Introduction — The Storefront That the Shopper May Never Visit
On September 29, 2026, Stripe published a deceptively simple observation that repays slow reading. Purchases made through its Link wallet by artificial-intelligence agents had increased thirty-eightfold over the preceding month, and the company named a cluster of personal agents — Meta’s newly launched Muse among them — as the vehicles through which consumers were beginning to spend in this new way.[1] The number was arresting, but not because it proved that agent-driven purchasing had suddenly conquered retail. Stripe did not disclose the original transaction count, nor an industrywide denominator against which the growth could be judged, and a thirty-eightfold rise from a very small base remains, in absolute terms, small. Its importance was architectural rather than quantitative. In the same announcement Stripe described agents that could navigate checkout obstacles, request incremental authorization when taxes or shipping changed the final total, draw on a consumer’s spending context to sharpen a recommendation, and offer purchase protection when something went wrong. Each of those capabilities presupposes a transaction in which the person expresses a preference, the assistant interprets it, structured product data makes goods legible to a machine, a protocol coordinates the exchange, an authorization mechanism governs the money, and a merchant fulfills the promise. The commercial relationship — once compressed into a single human visit to a single storefront — had begun to separate into parts, and each part had become a place where a different firm could compete for control.
“Over the past month, agentic purchases made with Link increased 38x.” — Dan Hill, Stripe [1]
Two weeks later, on October 15, Google began rolling out Universal Cart in the United States across Search, including AI Mode, and the Gemini app. Google’s own description was unglamorous and therefore revealing: a shopper could save products from multiple retailers into one persistent cart, receive price-drop and restock alerts, compare card benefits and promotions through Google Wallet, and complete purchases either with Google Pay or by handing off to the merchant’s own site, with the merchant remaining seller of record and responsible for confirmation and support.[2] The experience rested on the Universal Commerce Protocol that Google had developed with Shopify in January, just as OpenAI and Stripe’s Agentic Commerce Protocol had, a year earlier, established a way for merchants and conversational interfaces to exchange structured product and checkout information.[7][9] The paths were not identical. In March 2026, OpenAI had publicly conceded that its first Instant Checkout approach had not provided the flexibility it wanted and had shifted its near-term emphasis to product discovery and merchant-controlled checkout.[3] And on October 20, in a development that belongs to the same story even though it concerns businesses rather than shoppers, Visa, BILL, and Billtrust disclosed a live pilot in which an authenticated agent completed previously approved supplier-invoice payments across business portals — an agent recognized by the supplier’s platform in place of the CAPTCHA and bot-blocking defenses that had historically stood in its way.[4]
“The cart puts you in control.” — Suresh Ganapathy, Google [2]
Consider what these announcements, taken together, actually imply. The visible innovation is a more helpful shopping assistant. The deeper innovation is that the decisive commercial encounter — the moment at which a buyer’s needs are translated into a set of eligible sellers — can now occur outside the retailer’s own property, inside a system the retailer does not operate, under rules the retailer did not write, and, in some cases, without the buyer ever seeing which alternatives were considered and discarded. A brand can lose a sale without ever receiving a website visit. A small merchant can win a customer who would never have found it through conventional search. A payment network can become the trust layer between an instruction and a legally meaningful transaction. A technology platform can become a powerful commercial gatekeeper without ever taking title to the goods.
I choose the title Agentic Shelfspace because it captures a competition that the familiar vocabulary of e-commerce, search-engine optimization, and online marketplaces cannot fully describe. A physical retailer once competed for an eye-level display; an online merchant competed for a search result, a sponsored placement, a marketplace rank, a position in a social feed. In machine-mediated commerce, the scarce position is a place inside an assistant’s constructed set of recommendations — and even that set may be invisible to the buyer. The word shelfspace preserves the retail industry’s oldest economic insight: that distribution and selection determine what can sell, and that the party who controls the shelf extracts rent from everyone who needs to be on it. The modifier agentic identifies the new allocator: software capable of comparing options, pursuing delegated goals, invoking tools, and, with permission, transacting. The phrase is a claim about a changing location of commercial power. It is not a claim that stores, websites, or human judgment disappear.
I also choose the title because it converts a fashionable product story into a testable political-economic question for what I have elsewhere called the Five-Layer AI Economy. Who will decide which merchants an agent can see, which products it considers, which offers it presents, what it costs a seller to qualify, whose payment credential it uses, and who retains the customer after the sale? Those decisions connect applications and agents at Layer Five to models at Layer Four, datacenters at Layer Three, chips at Layer Two, and energy at Layer One. The upper layer generates transactions and revenue; the lower layers supply the inference and infrastructure that make such transactions practical at scale. The title therefore names an economic institution in formation: a digitally allocated, machine-interpreted commercial shelf whose governance may matter at least as much as its size.
The academic literature has begun to recognize the stakes in precisely these terms. In a chapter prepared for the National Bureau of Economic Research volume on the economics of transformative AI, Peyman Shahidi, Gili Rusak, Benjamin Manning, Andrey Fradkin, and John Horton of MIT, Harvard, and Boston University describe AI agents as market participants that can search, negotiate, and transact directly, and they argue that by lowering the costs of preference elicitation, contract enforcement, and identity verification, agents expand the feasible set of market designs while introducing frictions such as congestion and price obfuscation.[21] Their conclusion is appropriately modest, and it sets the epistemic tone for everything that follows in this paper.
“The net welfare effects remain an empirical question.” — Shahidi, Rusak, Manning, Fradkin, and Horton, NBER [21]
The thesis of this paper is not, then, that AI shopping eliminates competition. It is that AI shopping relocates several of competition’s most consequential decisions. An assistant may help a consumer uncover overlooked products, assess trade-offs, and avoid manipulative advertising. Equally, an assistant owned by a platform may privilege integrated merchants, default payment arrangements, sponsored offers, or products whose data is easier for its systems to process. Both outcomes are possible; neither follows automatically from the existence of a protocol or the announcement of a cart. The critical questions are therefore empirical and institutional. Does an agent optimize for the user’s specified interests? How can the user verify that claim? Can an independent merchant be discovered without contracting separately with each powerful interface? Does the consumer understand when the agent is authorized to explore, to recommend, to reserve, or to purchase? And can the market distinguish genuine improvements in decision quality from increases in conversion caused merely by narrowing what buyers are shown?
Six sections develop the answer. Section 1 explains how shopping migrates from the browsing of pages to the delegation of tasks, and why product data becomes a form of commercial legibility. Section 2 maps the competing commerce and payment protocols and the firms behind them, drawing on earnings disclosures through the second quarter of 2026. Section 3 examines recommendation power, advertising inside conversational systems, the custody of customer intent, and the conflicts of interest that research has already documented. Section 4 studies the business models of retailers, brands, payment firms, and startups, and connects transaction economics to inference economics across the five layers. Section 5 sets out four conditional futures through 2030 and a governance, legal, and measurement agenda. Section 6 distills seven pillars for executives, founders, payment networks, and public officials. The central proposition throughout is straightforward and, I think, durable: the next storefront is a decision process, and the new shelfspace may be allocated before the consumer knows a store has entered the contest.
Section 1 — From Human Browsing to Machine-Mediated Purchasing
1.1. The four commercial shelves: physical placement, search, marketplaces, and agents
For most of modern retail history, the shelf has functioned as a commercial allocation mechanism rather than as a piece of furniture. Its location affected which products consumers noticed, compared, and purchased. A manufacturer could invest in packaging, retailer relationships, point-of-sale promotion, and distribution, but success also depended on being stocked in the right store and presented in the right place; slotting fees, planograms, end-cap negotiations, and the quiet politics of adjacency were the institutional expression of that dependence. The shelf was access to the consumer’s limited attention, and the retailer who controlled it could charge for that access directly or indirectly. Selection, signage, adjacency, and out-of-stock performance jointly shaped demand, and a product that never reached the shelf could not sell no matter how good it was.
The internet did not remove this logic; it transformed and multiplied it. Search engines shifted commercial placement into indexed pages and keyword relevance, and the auction for a position on the results page became one of the largest advertising markets ever created. Retail media networks sold prominence near the point of purchase, converting the retailer’s own website into a second shelf whose space was auctioned to the brands already on the first. Marketplaces introduced their own ranking systems, review signals, fulfillment badges, and paid placements. Social commerce moved discovery into recommendation feeds governed by engagement models. In each system, the merchant sought consumer attention within a surface that another institution controlled. The new allocator could be a retailer, a search engine, a marketplace, or a social network, but the commercial problem remained recognizably about visibility and conversion, and the consumer remained, in each case, a human being looking at a screen.
Agentic shopping adds a fourth shelf that differs from the other three in a way that is easy to state and hard to overstate: the consumer may not be looking at anything. A person may say, “Find a durable carry-on bag under my budget, compatible with my airline’s size rules, available before Friday, and not made of leather.” The assistant interprets those constraints, searches product catalogs, rejects items that fail the requirements, compares warranties and review evidence, and perhaps asks a clarifying question about color or wheel type. Only a small fraction of the market appears in the final answer. Unlike a conventional page of search results, the entire assortment considered, the reasons for exclusion, and the influence of commercial arrangements on the ordering may not be visible at all. The user sees a recommendation, perhaps two alternatives, and a path to checkout. Columbia Business School’s Omar Besbes and Yash Kanoria, with co-authors, have audited this behavior experimentally and found that agents can exhibit what they call choice homogeneity, concentrating demand on a few “modal” products while ignoring others entirely, with preferences that reshuffle when the underlying model is updated.[20]
“Agentic markets are volatile and fundamentally different from human-centric commerce.” — Allouah, Besbes, Figueroa, Kanoria, and Kumar, Columbia Business School [20]
This is why an agentic shelf should not be equated simply with the first result in an AI-generated answer. It includes the upstream formation of the candidate set, the interpretation of the buyer’s preferences, the ranking and explanation of eligible products, and the downstream technical ability to transact. A merchant whose catalog is omitted has no opportunity to win an order; one whose product is retrieved but misunderstood may lose on false grounds; one that is accurately recommended but cannot support authorized checkout may lose to a more transaction-ready competitor. Visibility is therefore necessary but insufficient. The contest spans machine readability, recommendation quality, transaction readiness, and post-sale confidence, and a firm that wins on three of those four dimensions can still lose the order.
The structural analogy has limits worth stating plainly. A supermarket shelf has a known location and finite physical dimensions; an AI agent’s candidate set can change continuously and can be personalized to an individual. Its capacity is constrained by computation, retrieval design, latency, user attention, policy, and commercial objectives rather than by square footage. Nonetheless, the economic resemblance is substantial: a privately organized decision environment mediates access between sellers and buyers, and the organizer of that environment holds a form of power that the parties on either side of it cannot easily contest. What changes in the agentic case is that the organizer is increasingly able not merely to display but to interpret, to advise, and to act.
1.2. The shopping journey becomes a chain of delegated decisions
The phrase agentic commerce is often used as though a purchase were a single event, but a more accurate description separates at least six distinct stages, each of which can be performed by a different actor under different rules: intent formation, discovery, evaluation, selection, authorization, and fulfillment. At the first stage, an individual expresses a goal, sometimes incompletely or contradictorily. At discovery, the system locates possible suppliers. Evaluation verifies attributes, prices, shipping conditions, reviews, and trade-offs. Selection narrows the options and may recommend one. Authorization establishes whether, and under what limits, the agent may place an order. Fulfillment, returns, refunds, warranty administration, and customer support remain substantial commercial obligations that outlast the transaction by months or years.
These stages can be assembled in strikingly different configurations, and the 2026 product landscape illustrates nearly all of them. A consumer may ask ChatGPT to compare refrigerators and then purchase on a retailer’s website, which is the arrangement OpenAI explicitly embraced in March.[3] A Google user may save products from different merchants in Universal Cart and finalize with Google Pay or on the seller’s own site.[2] A Meta Muse user may ask a persistent assistant to manage a task and approve a payment through Link, with a one-time virtual card standing between the agent and the consumer’s real credentials.[5] An Amazon customer may use Alexa for Shopping, which the company formed in May by bringing together the former Rufus assistant and Alexa+, entirely within Amazon’s own ecosystem, including automated features such as price alerts and auto-buy.[6] A corporate accounts-payable function may delegate to an agent the completion of invoices that a human has already approved, as in the Visa–BILL–Billtrust pilot.[4] These are variations on a common architecture, not proof that all agents possess the same autonomy or that every demonstration is generally available.
The stage distinction resolves a persistent confusion in public debate. A company can lead in shopping discovery without owning payments. A payment network can provide agent authentication without ranking products. A commerce platform can make merchants accessible to many agents while leaving recommendation logic to those agents. The economic value and the regulatory responsibility of each participant vary according to the stage it controls, and to treat all of these firms as interchangeable “AI shopping platforms” is to miss the emerging division of labor — and, with it, the emerging division of power.
It also reveals where authority can be contested. If a consumer has approved only price comparisons, the assistant has no license to submit an order. If a consumer permits purchase below a monetary ceiling, that ceiling must be enforced against the final total, including taxes, shipping, and any subscription the merchant bundles into the price. If the merchant changes the price during checkout, the platform needs a rule for reauthorization. Stripe’s September 2026 discussion of incremental authorization illustrates how real checkout creates conditional decisions that are absent from a polished product demonstration.[1] The most valuable assistant is not necessarily the one that executes the most tasks; it may be the one that reliably knows which tasks require permission.
The buyer’s intent also evolves across a conversation, and the system must track that evolution faithfully. At the beginning, “best value” may mean the lowest purchase price. After a clarification, it may mean a five-year warranty and local repair. The final order should reflect the clarified objective, not the first broad phrase the agent encountered. This creates a new type of process integrity: preserving the history of the buyer’s preferences sufficiently to justify a recommendation without collecting or sharing more personal information than the task requires. A purchase record shows what happened; a decision record can explain why it happened, and in disputes the second will matter as much as the first.
1.3. The 2025–2026 inflection: progress without inevitability
September 2025 supplied a foundational milestone. OpenAI and Stripe introduced the Agentic Commerce Protocol alongside Instant Checkout in ChatGPT, with a design that sought to let buyers, agents, merchants, and payment providers communicate through structured messages while keeping the merchant responsible for fulfillment and for the commercial relationship.[7][8] The point was not that the chatbot became the merchant. It was that discovery and checkout could be coordinated inside a conversational interface through agreed technical conventions, published as an open specification, rather than through a bespoke integration negotiated one retailer at a time.
January 2026 brought a different architectural ambition. Google and Shopify introduced the Universal Commerce Protocol to represent parts of a transaction journey — merchant capability discovery, checkout, identity linking, orders, and extensions — and Google described a vendor-agnostic design with reference implementation guidance for eligible merchants in Search AI Mode and Gemini.[9][10] Shopify emphasized continuity of merchant business logic, including the rules governing discounts, subscriptions, and fulfillment, so that a merchant’s commerce engine would not need to be rebuilt for every assistant.[11] A protocol alone does not guarantee equal access. It can, however, reduce repetitive integration work when a merchant would otherwise need to connect separately to each interface, and in doing so it lowers the barrier to entering the contest even as it leaves the contest itself unresolved.
March 2026 then provided an especially useful corrective to technological determinism. OpenAI expanded ACP to carry richer product feeds, comparisons, and promotions, rolled out visual browsing and side-by-side comparison to its free and paid tiers, and — in a sentence that would be quoted across the industry — acknowledged that its original Instant Checkout design had fallen short.[3][58] Subsequent reporting indicated that only a few dozen Shopify merchants had actually gone live on the feature and that several participants regarded it as a channel that had not driven sales, which is consistent with a decision to let ChatGPT handle discovery while merchants own the rest of the transaction.[62]
“The initial version of Instant Checkout did not offer the level of flexibility that we aspire to provide.” — OpenAI [3]
This was not evidence that agentic commerce had failed. It was evidence that the position of the checkout boundary, the merchant relationship, and the consumer experience remained unsettled. A technology capable of suggesting products may be far easier to deploy at scale than a universal transaction system that must accommodate every merchant’s promotions, inventory changes, tax arrangements, identity needs, loyalty programs, returns policies, and exceptions. The retreat from in-chat checkout was, read correctly, an advance toward a more defensible division of labor.
By October 2026, the field had broadened considerably. Google’s Universal Cart was launching in the United States; Meta’s Muse was integrating merchant connectors and payment options; Stripe was reporting rapid growth in one subset of wallet-enabled agent purchases; Mastercard and Visa were expanding frameworks for trusted agent payments; and PayPal was helping existing merchant sites become agent-ready through browser-facing capabilities.[1][2][4][12][13][54] Yet deployment scope differed by company, geography, merchant, and product category. A launch announcement, a pilot, a technical specification, an adoption statistic, and a generally available checkout experience are five different kinds of evidence, and serious forecasting requires keeping them separate.
The most persuasive sign of market evolution may be this willingness to revise product architecture in public. Early systems favored tightly integrated conversational purchase flows. Later deployments increasingly emphasized modular discovery, merchant-controlled conversion, reusable wallets, and explicit authorization. That is a characteristic pattern in emerging markets: a first generation proves a possibility; a second generation must solve distribution, incentives, trust, and exceptions. Agentic Shelfspace belongs to the second generation because its central problem is not whether a model can recommend a product, but how commercial access will be organized once many systems can do so.
1.4. Product data becomes a form of commercial legibility
An AI agent does not physically inspect the items it recommends. It depends on representations: product feeds, structured attributes, images, text, prices, inventory signals, merchant policies, reviews, and the responses returned by tools. A seller with excellent merchandise but missing or inconsistent data may become difficult for an agent to evaluate, and a seller with clearer machine-readable information may be easier to include even when its product is not objectively superior. I call this condition merchant legibility: the ability of a commercial offering to be correctly discovered, interpreted, compared, and acted upon by automated systems. Legibility is the agentic era’s equivalent of being stocked.
The technical details are economically important in a way that non-specialists tend to underestimate. A product listing should distinguish a manufacturer’s model from a merchant-specific stock-keeping unit, a color variant from a separate product, a one-time price from a subscription price, and a promotional price from the standard price. It should carry accurate shipping restrictions, warranty terms, seller identity, stock status, and return rules. OpenAI’s merchant guidance for ACP emphasizes structured product feeds and attribute quality; Google’s UCP documentation requires eligible merchant data and checkout integration; Shopify argues explicitly that the quality of product information determines participation in AI commerce.[10][14][15] Shopify’s own second-quarter disclosures gave that argument a number: searches powered by its structured Catalog converted at roughly twice the rate of searches built on scraped data, and about 80 percent better than traditional organic search.[27] These are not formatting niceties. They determine whether automated comparison can preserve meaning, and whether a merchant is in the contest at all.
Legibility also invites manipulation, and the research record already documents how. The Columbia team showed that a seller-side agent making simple, query-conditional tweaks to product descriptions could drive significant gains in market share against AI buyers — a finding that anticipates an “AI-SEO” contest in which listings are tuned to the known biases of buyer agents rather than to the needs of human readers.[20] A vendor might stuff a feed with attractive but unverifiable claims, omit an unfavorable condition, or describe incompatible goods using fashionable keywords. An assistant may overweight repeated marketing phrases unless it cross-checks official specifications and independent evidence. Conversely, strict data standards can burden small merchants that lack integration budgets. Market design therefore faces two simultaneous objectives: improving the truthfulness and comparability of inputs while keeping participation feasible for legitimate sellers of different sizes.
The difference between data being available and data being reliably actionable deserves emphasis. An agent may know that a camera is advertised at a certain price but not whether that exact variant is in stock. It may identify a discount but fail to apply the eligibility rule. It may recommend a product on a warranty promise that applies only in another jurisdiction. Checkout protocols can test live conditions closer to purchase, but the recommendation may already have shaped demand by then. Merchants and agent platforms should monitor not only feed ingestion but discrepancies between what the agent said and what the buyer ultimately received — a metric I formalize in Section 3 as offer integrity.
For a brand, this suggests an investment agenda broader than what is now called generative-engine optimization. It includes trustworthy structured data, first-party specification pages, service and repair information, stable product identifiers, accessible customer-support policies, and attribution systems that can be reconciled against agent-reported behavior. Adobe’s analysts have repeatedly found that substantial portions of retail websites remain unreadable to machines, which is a quiet form of exclusion from a channel that, on Adobe’s July 2026 data, converted 60 percent better than non-AI traffic.[16][17] For regulators, the concern is whether machine-ready distribution creates a new form of exclusion. For startups, it creates an opportunity to build catalog normalization, verification, and real-time commerce infrastructure. The agent’s shelf is built from data, but the quality of the shelf depends on truth, not on volume alone.
1.5. Consumers want reduced effort, not unrestricted autonomy
The appeal of delegated shopping is understandable and, in many categories, substantial. A consumer buying a familiar household item may not want to visit six websites, translate technical specifications, compare shipping fees, and repeatedly enter payment details. The value of an assistant is not simply faster typing; it is lower search cost and better matching between needs and available products. That benefit is potentially greatest when product attributes are complex, the consumer has many constraints, or the financial stakes justify careful comparison. It is less obvious when shopping is recreational, experiential, social, or expressive, and the consumer enjoys the browsing itself. Mark Bartholomew of the University at Buffalo School of Law has warned that Americans spend more time shopping than on education, volunteering, or talking on the phone, and that automating the activity could quietly erode economic, psychological, and social benefits people derive from shopping on their own terms.[24]
“AI agents aren’t just designed to assist; they’re designed to influence.” — Mark Bartholomew, University at Buffalo School of Law [24]
Available evidence supports growing use, but it must be read narrowly. Adobe reported that in July 2026 referrals from AI sources to U.S. retail websites were 62 percent higher year over year, converted 60 percent better than non-AI traffic, kept visitors on site 59 percent longer, and were 33 percent less likely to bounce — the eleventh consecutive month in which AI-referred traffic outperformed other channels on conversion.[16][17] Those are measured referred visits, not all AI-mediated transactions, and referral selection can favor buyers who are already near purchase. Salesforce’s data for the 2025 holiday season attributed 20 percent of global online retail sales, or about $262 billion, to the influence of AI and agents, a figure that blends first-party and third-party data with market assumptions and should be read as an influence estimate rather than a count of agent-placed orders.[36] Academic shopping-agent benchmarks, meanwhile, continue to find persistent difficulty with personalized curation and with hidden user requirements distributed across a conversation.[18][19] More shopping through AI does not mean every current agent makes reliable purchasing decisions.
Consumer trust is not a single variable. A person may trust an assistant to produce a shortlist but not to charge a credit card; trust it to buy a low-cost replenishment item but not medical equipment; or authorize a transaction while insisting that the merchant, not the assistant, manage returns. Trust can also be delegated differently across individuals within a household. The system must distinguish advice, commitment, and transfer of funds. These are not merely user-interface states; they are changes in authority with potentially different legal and financial consequences, and Meta’s design choice to require explicit approval before Muse sends an email or completes a transaction reflects an industry learning that lesson early.[5]
The strongest consumer proposition may be bounded autonomy: automate repetitive low-risk work while preserving meaningful checkpoints. An assistant could monitor a price and present an option when a threshold is met, but require explicit acceptance before purchase. It could reorder a known product below a spending limit with a clear record. It could prepare a cart and explain two trade-offs. Stanford economist Neale Mahoney, director of the Stanford Institute for Economic Policy Research, has pointed to a less obvious benefit of this kind of delegation: his research with Liran Einav and Ben Klopack found that sellers of subscriptions roughly double revenue through consumer inertia and cancellation friction, and that forcing an active decision changes behavior dramatically.[25]
“They are about four times more likely to cancel.” — Neale Mahoney, Stanford Institute for Economic Policy Research [25]
An agent that surfaces forgotten recurring charges and asks whether to keep them is, in effect, attacking one of the economy’s most profitable weak spots on the consumer’s behalf. A platform that interprets convenience as permission to purchase anything, by contrast, may discover that one mistake destroys more trust than hundreds of successful suggestions create. The test of economic value should therefore be task-specific. How much time was saved? Did the selected product meet the constraints? Was the total cost accurate? Were better alternatives accessible? Did the customer understand the recommendation and the authorization? Was the order fulfilled without avoidable dispute? Measuring only gross merchandise value or conversion rate risks rewarding overconfident agents. The consumer’s welfare, not simply the merchant’s sales, is the defensible long-term basis of this market.
1.6. Mapping the agentic transaction onto the Five-Layer AI Economy
The Five-Layer AI Economy provides a way to connect apparently lightweight shopping interactions to physical infrastructure, and to resist two opposite errors: the error of treating a chatbot’s product suggestion as computationally free, and the error of treating every incremental order as a measurable draw on the electrical grid. At Layer Five, an application receives the task, provides the interface, calls merchant tools, and records consent. At Layer Four, models interpret natural-language requirements, reason across alternatives, assess uncertainty, and generate responses or plans. At Layer Three, datacenters host inference, retrieval, orchestration, databases, and network services. At Layer Two, accelerators and associated hardware determine performance per inference and the economics of memory-intensive, tool-using workloads. At Layer One, electricity and associated infrastructure support the computational system. Payment networks and merchant backends cut across this stack as institutional rails rather than fitting neatly into one technical layer.
The effect is not a simple equation in which every additional purchase requires a large increase in energy. A merchant feed can be cached; routine eligibility checks can run on conventional computing; smaller specialized models can handle narrow tasks; and agents can use event-driven scheduling rather than permanent model inference. The marginal compute required for an incremental order depends on architecture, task complexity, retries, model choice, and traffic patterns. Nevertheless, if commerce shifts from occasional human searches to continuous background monitoring across hundreds of millions of consumers — the kind of always-on price tracking that both Universal Cart and Alexa for Shopping now advertise — aggregate inference, storage, and networking demand can become strategically relevant.[2][6] Amazon’s own disclosure that it expects roughly $220 billion of capital expenditure in 2026, overwhelmingly for AI and cloud infrastructure, is a reminder that the firms building shopping agents are also the firms building the datacenters those agents run on.[28]
Layer Five also sends information downward and money upward. Transaction fees, advertising revenue, subscriptions, merchant tools, and payment services may help finance model development and datacenter expansion. Better model reliability can improve selection and reduce support costs; more efficient chips can lower the cost of evaluating richer catalogs; resilient energy and datacenters can sustain availability during high-volume commerce periods. The stack is economically circular rather than only vertically consumed, and the health of the loop depends on whether the commerce at the top produces genuine value or merely more activity.
This is the first structural insight of Agentic Shelfspace: the commercial shelf is no longer a page owned entirely by a retailer. It is a distributed decision and execution chain, dependent on technical layers and governed by contracts, defaults, data policies, and consumer authorizations. A firm may control one important link without controlling the others. The question of who captures value, and who can be excluded, must be investigated across the whole chain — which is the task of the next section.
Section 2 — The Protocol Contest: OpenAI, Google, Shopify, Stripe, Amazon, Meta, and the Payment Networks
2.1. A protocol is a language; a platform controls a distribution channel
Every agentic transaction requires a way for separately operated systems to express facts and instructions to one another. The agent must know that a product exists, how much it costs, whether it can be delivered, what information is needed to purchase it, and whether the consumer has approved the relevant commitment. A merchant must understand the request, calculate a current total, apply its policies, and return an order state. A wallet or payment provider must decide whether the authorization is valid and which credential may be used. A protocol defines how some of these entities communicate. It does not, by itself, decide which products are recommended, how they are ranked, or who owns the relationship with the buyer — and the distance between those two things is where most of the economic value in this market will be captured or lost.
This distinction is indispensable because the industry’s own vocabulary blurs it. A technically open specification can reduce the expense of integration without neutralizing the market power of the interface where customers begin their searches. Conversely, a closed, vertically integrated marketplace can provide highly reliable shopping precisely because it tightly coordinates catalog data, payments, warehousing, and returns. Openness and quality are related but not interchangeable. The relevant competition involves at least three dimensions: the technical compatibility of merchants and agents; the distribution power of consumer-facing platforms; and the control of identity, payment, and post-purchase records.
Imagine a small maker of specialized cycling equipment. It could implement a neutral commerce interface and thereby avoid writing separate checkout code for several assistants. Yet if a major agent never retrieves the maker’s products, the standardized interface creates no sale. If an assistant selects the maker’s product but routes the shopper through a wallet that does not serve the consumer’s region, conversion may fail. If the buyer completes a transaction but the platform withholds attribution or restricts permitted follow-up, the brand’s customer lifetime value may be impaired. The engineering problem and the distribution problem must be solved together, and a merchant that solves only the first has bought a ticket to a lottery whose rules it cannot read.
There is also a governance issue within protocol design itself. An open standard can have optional capabilities, proprietary extensions, certification requirements, and reference implementations that become de facto defaults. The institution controlling the most widely deployed extension may exercise power even when the core format is public. Buyers of technology should ask not simply whether a standard is “open” but whether they can change agent providers, wallets, catalog intermediaries, and checkout surfaces without forfeiting product availability, historical records, consumer consents, or seller policies. Portability of the relationship, not merely of the message format, is the test that matters.
2.2. ACP: from conversational checkout to product-discovery infrastructure
The Agentic Commerce Protocol developed by OpenAI and Stripe was introduced publicly in September 2025 with a deliberately practical ambition: connect an AI shopping interface to merchants’ existing commercial systems. Stripe described a specification that could work with different backends and payment providers, preserve the merchant’s status as merchant of record, and support evolving transaction types; it was published as an open-source specification under the Apache 2.0 license.[8] OpenAI’s launch emphasized that customers could discover products through conversation and, for supported transactions, confirm a purchase within ChatGPT itself.[7]
An important design boundary is that the agent does not become entitled to spend merely because it can assemble a cart. The consumer must select or approve a transaction; merchants still manage availability, price, taxes, fulfillment, and customer service. OpenAI’s checkout specification makes the merchant responsible for payment processing, tax calculation, risk analysis, and order acceptance, and states that OpenAI does not act as merchant of record.[61] Delegated-payment arrangements can help avoid exposing the buyer’s complete payment credentials to the agent, but the surrounding consent and dispute processes remain essential. This is why agentic checkout should not be treated as a synonym for unattended purchasing. Structured checkout can be deeply assisted while final authorization remains human.
OpenAI’s March 24, 2026 announcement changed the near-term emphasis in ways that reverberated through the sector. The company expanded ACP to provide richer product feeds, comparisons, and promotions; introduced visual browsing, side-by-side comparison, conversational filtering, and image-based lookalike search; and allowed merchants to use their own checkout experiences, including through dedicated apps inside ChatGPT. It named retailers such as Target, Sephora, and Nordstrom as supporting the new discovery experience, described a Walmart in-ChatGPT app with account linking, loyalty, and Walmart’s own payments, and announced that Shopify’s Catalog would feed eligible merchants’ products into ChatGPT by default.[3][58] This public adjustment has unusually high analytical value. It demonstrates that a firm may choose to own the beginning of a shopping journey — the interpretation of intent and the assembly of candidates — without insisting on owning every subsequent interaction.
Such a strategy could be commercially powerful precisely because it is less demanding operationally. Product discovery can scale across many categories while leaving the difficult work of payment, returns, tax, local delivery, and loyalty with existing merchants. But discovery power is not trivial. A recommendation interface that influences which products enter consideration may shape where purchases occur even when the checkout happens elsewhere. The March pivot therefore strengthens, rather than weakens, the Agentic Shelfspace thesis: the control point can sit upstream of payment, and the party that relinquishes the buy button may be the party that most firmly holds the shelf.
The ACP documentation also reveals the practical disciplines required for broad merchant participation. Sellers must provide consistent feed attributes and keep price and availability current; integrations may be mediated through providers; and approved partner access and specific checkout experiences carry eligibility restrictions.[14][61] The existence of an open technical specification should not be confused with universal admission to every commercial surface. A robust industry benchmark would separately measure specification accessibility, merchant onboarding success, live product coverage, actual recommendation inclusion, and completed purchases.
The long-term question is whether ACP evolves into widely reusable commercial infrastructure or remains closely associated with ChatGPT’s distribution advantages. The first would lower multihoming costs for merchants across many agents. The second could still lower integration friction while concentrating demand through a particular interface — one that, by OpenAI’s own account in October, reaches more than a billion people each week.[37] These outcomes are not mutually exclusive. An ecosystem can have portable data formats and concentrated consumer attention at the same time, and that combination is precisely the one that deserves the most careful watching.
2.3. UCP: Google, Shopify, and commerce across several surfaces
Google and Shopify approached the challenge with the Universal Commerce Protocol. Announced in January 2026, UCP was presented as a standard covering the shopping journey from merchant capability discovery and checkout to related identity and order functions.[9][10] Google emphasized vendor neutrality and a reference implementation for eligible businesses appearing in Search AI Mode and Gemini. Shopify highlighted preservation of merchant rules and the possibility of exposing inventory and promotions without reconstructing the store’s commerce engine for every assistant.[11] By the time of its second-quarter earnings call in August, Shopify was telling investors that dozens of retailers and platforms had adopted UCP and that major technology partners — OpenAI, Google, Meta, and Microsoft among them — were building on its open ecosystem.[27]
Google’s October 15 Universal Cart release made the architectural stakes tangible. It described one consumer cart containing products encountered across Google surfaces and from different retailers, with AI-enabled monitoring of price and stock, awareness of card benefits through Google Wallet, automatically surfaced promotion codes, and checkout either through Google Pay or through merchant-controlled paths; launch partners reportedly included Nike, Sephora, Target, Walmart, Ulta Beauty, Wayfair, and many Shopify merchants.[2][57] The participating seller remains seller of record and provides confirmation and customer support. This is a meaningful hybrid between the old browser-based journey and a fully autonomous buyer agent. The consumer still controls the purchase; the cross-merchant organization and some transaction preparation move to the platform.
Universal Cart is not literally a single universal order that collapses distinct merchants into one legal seller. Different retailers retain separate inventory, terms, and fulfillment obligations. The value lies in shared organization and continuity of intent: the shopper’s items persist across Search and Gemini, and the agentic layer can track relevant changes without requiring a new search each time. That continuity may create switching costs based on saved context and preferences, even where the underlying commerce standard is portable. It is also worth recording what the launch did not include. The YouTube and Gmail surfaces previewed at Google I/O in May were absent from the October rollout, which was confined to the United States, and the announcement said nothing about merchant fees, ranking criteria, or data-sharing terms.[57]
UCP’s composability is attractive because retail is irregular. Promotions may depend on loyalty status, orders may contain multiple variants, pickup requires location and time, and post-purchase changes need stateful records. A protocol that merely sends a product identifier and a card token is unlikely to satisfy complex merchants. Yet richer commerce semantics create more governance questions. Who decides which optional functions become supported? Can a small seller implement only the core? Does the agent surface a merchant that cannot support every optional capability? Does an assistant privilege sellers with fast, deeply integrated checkout over those whose prices or quality are better? The answers can alter market access as significantly as any formal ranking criterion.
Google also occupies a different economic position from a standalone agent developer. Search, advertising, the Shopping Graph with its tens of billions of product listings, Merchant Center, Gemini, Wallet, and the commerce protocol can reinforce one another.[2] Such breadth can improve relevance and freshness. It also means policy analysts must examine how a consumer-facing assistant, a merchant acquisition channel, and an advertising business coexist inside one firm — a question that, as Section 5 discusses, European regulators had already answered unfavorably for Google’s conventional search results in July 2026.[46] A protocol’s published openness does not answer whether its distribution channel applies neutral eligibility, disclosure, and ranking rules.
2.4. Stripe, Visa, Mastercard, and PayPal: payments become an agent trust architecture
AI agents expose a transaction problem that conventional checkout was never designed to handle well. A human browser can be challenged with a password, a one-time code, or a CAPTCHA, and can visually inspect a final price. A machine agent may need to operate through an API or an existing website while proving that it is legitimate, identified, and bounded by the user’s authority. It must carry evidence of intent without revealing unnecessary credentials. Payment firms are therefore building an identity-and-control layer around the act of spending, and their earnings calls through mid-2026 show that they regard this layer as a growth market rather than a defensive necessity.
Stripe has worked on both merchant enablement and agent-side payment capabilities. Its Agentic Commerce Suite, announced in December 2025, aims to bring catalogs and checkout connections into new AI channels.[53] In September 2026, Stripe described Link’s wallet for agents handling additional authorization when checkout totals change, providing spending context to improve recommendations, and extending purchase protection to eligible agent purchases.[1] Where a merchant accepts Link, the agent checks out with the consumer’s saved method; everywhere else, Link issues a single-use virtual card capped at the approved purchase, so that the agent never sees the real card number — an arrangement that, according to reporting in late October, had begun to concern the card networks about where fees and data would flow, even though no public evidence showed volume actually moving off their rails.[56] Stripe’s thirty-eightfold figure is a company-specific growth observation, not a marketwide spending share, and the fact that Muse launched in the same month it was measured means a large part of the increase may reflect one product switching on.
Visa’s Trusted Agent Protocol addresses the problem of recognizing a legitimate agent and communicating verified identity and intent to merchants.[49] Visa’s October 20 pilot with BILL and Billtrust applied that idea to approved business invoices across a supplier portal: the BILL payment agent, registered through Visa’s protocol and its Agentic Directory, used tokenized credentials under buyer-defined controls, and Billtrust authenticated the agent in place of the CAPTCHA and bot-blocking mechanisms that had previously forced finance teams to log in and key in payment details by hand.[4] The pilot is narrower than the claim that machines can autonomously purchase anything on any website — the agent executes only payments a human has already approved — but its importance is the demonstration that existing portal defenses and payment rules can be adapted to verified machine actors rather than simply disabled.
“Businesses should not have to choose between automation and control when making payments.” — Mark Nelsen, Visa [4]
On its fiscal third-quarter call in July, Visa reported net revenue of $11.6 billion, up 14 percent, with payments volume exceeding $4 trillion in a quarter for the first time and value-added services revenue up 34 percent; it announced partnerships with OpenAI and Meta to carry Visa credentials into agentic experiences, and it named Agent Score, Agent Directory, and a Token Assurance Framework as the instruments by which it intends to make AI-initiated transactions more transparent.[29][30] Chief executive Ryan McInerney framed the opportunity as market expansion rather than disintermediation.
“Agentic commerce is a when, not an if.” — Ryan McInerney, Visa [30]
Mastercard’s Agent Pay initiatives likewise center authenticated participants, user intent, tokens, and traceability. Its Verifiable Intent work with Google addresses the creation of a tamper-resistant representation of what a user authorized, with existing chargeback processes resolving disputes; its Agent Pay for Machines service, launched in June with more than thirty industry partners, targets high-frequency, low-value interactions among software and service providers.[31][50][51] On September 30, 2026, the company announced new trust and intelligence services forming what it called the intelligence layer of an Agent Pay Trust Framework — identity, intent, controls, execution, and intelligence — and projected that one in ten consumers would routinely use agents to make purchases by 2030.[12] Chief executive Michael Miebach’s second-quarter message to investors was competitive and unambiguous.
“We expect cards will prevail in an agentic world.” — Michael Miebach, Mastercard [32]
PayPal and Braintree have emphasized making existing merchants accessible to agentic commerce, including ACP- and UCP-related integration paths, and PayPal’s September 2026 WebMCP discussion highlighted the possibility of agents interacting with merchant websites through structured browser-facing capabilities while reusing storefronts that already exist.[13][60] This model matters for merchants unwilling to build a complete new API channel. PayPal’s own second-quarter results — revenue of $8.68 billion, up 5 percent, on total payment volume of $486 billion, with branded checkout growing only 2 percent on a currency-neutral basis — also illustrate why an incumbent wallet might regard agentic payments as a future source of growth rather than a present one; chief executive Enrique Lores placed the payoff in 2028 and beyond.[33]
“Agentic payments and digital identity, to become increasingly meaningful contributors to our business.” — Enrique Lores, PayPal [33]
The contest among payment networks is therefore more than a competition to become the default buy button. It is a competition over the format of consent, identity, tokens, fraud evidence, liability, dispute handling, and merchant acceptance. These functions could become essential infrastructure for agents even if model providers change. A consumer may rotate among several assistants but still prefer a familiar wallet, bank, network, or payment credential. Conversely, a tightly integrated agent could make a particular payment choice feel inevitable. The distribution of bargaining power will depend on portability and user choice as much as on cryptography.
2.5. Amazon and Meta show why there will not be only one storefront model
Amazon enters agentic commerce with a structurally different asset base from any protocol consortium. It already integrates a marketplace, retail assortment, seller services, payments, reviews, delivery capabilities, and a $19.8 billion-per-quarter advertising business.[28] In May 2026 it combined the experiences previously known as Rufus and Alexa+ into Alexa for Shopping, and on its second-quarter call it reported that more than 350 million customers had used the assistant over the trailing twelve months, that active users had nearly doubled year over year with interactions up more than fivefold, that U.S. customers who use Alexa for Shopping spend an average of more than 40 percent more per order, and that customers who tried Alexa+ were signing up for Prime at a nearly 25 percent higher rate.[6][28][59] Its advantage is not merely a language model: it has extensive transaction and fulfillment data and a functioning distribution network. Its strategic risk is that a sufficiently capable external assistant could reduce the number of consumers beginning their journey on Amazon’s own surfaces.
“Customers love Alexa for Shopping, our agentic AI shopping assistant.” — Andy Jassy, Amazon [28]
Amazon’s response to that risk has been as instructive as its product. In November 2025 it sued Perplexity, alleging that the Comet browser’s assistant accessed customer accounts and placed orders in violation of its terms and of anti-hacking law; it won a preliminary injunction in March 2026; and on August 4, 2026, the Ninth Circuit vacated that injunction, holding that on the record before it the user, not Perplexity, accessed Amazon’s systems, while expressly leaving contract and terms-of-service claims open.[47] Six weeks later, on September 20, Amazon blocked Meta’s Muse from shopping on Amazon.com, telling customers that continued access by an unauthorized agent violated the Conditions of Use they had accepted; Amazon said Meta had not notified it that Muse would reach its store, that the agent did not identify itself when browsing, and that it appeared to capture and store credentials, claims Meta disputed.[48] Whatever the merits, the sequence shows a marketplace fortress defending its shelf with the one instrument the courts had left in its hands — and running its own agents to shop elsewhere while doing so.
Meta illustrates another entry point: persistent, socially situated assistance. Its September 8 launch of Muse described a personal agent, available in the United States on iOS, Android, the web, and WhatsApp, that runs in an isolated virtual machine, opens its own browser, fills forms, negotiates, and checks out through Link with a one-time card, with sensitive actions requiring approval and a separate oversight agent reviewing what the assistant sends to the internet; Meta said Muse was the first agent covered by Link’s purchase protections.[5] Two weeks later, at Connect, the company announced merchant connectors with Walmart, Best Buy, and Sephora among others, and payment options including Shop Pay and PayPal.[54] A user might begin shopping from a message, a social context, or a wearable device rather than a dedicated retail search. Such interfaces add relevance when a purchase is part of a larger task. They also raise sensitive questions about whether information drawn from communications, contacts, or persistent memory is appropriate to use for recommendations, questions Meta addressed at launch by stating that Muse’s conversations and virtual-machine data are not shared with its advertising systems.
The ecosystem is thus not a neat duel between ACP and UCP. At least five models can coexist. An open-web assistant can recommend goods across many merchants. A marketplace agent can optimize within an established ecosystem. A personal agent can initiate shopping as one of many household tasks. A merchant-owned agent can give expert guidance within one brand’s or retailer’s inventory. A business-to-business purchasing agent can apply approved vendor lists, invoices, procurement policies, and enterprise payment systems. Different product categories and customer preferences will likely support different combinations, and the same consumer will move among them in a single week.
This heterogeneity is a competitive safeguard if users and merchants can move among providers. It becomes a new fragmentation risk if each assistant requires separate feeds, credentials, authorization rules, promotions, and attribution arrangements. The practical goal for merchants should not be a bet on the one winning protocol. It should be a modular capability stack: reliable catalog data, standardized inventory and order APIs, payment flexibility, portable product identity, security controls, and measurable access to multiple discovery channels. Platform diversity is valuable only when participation is economically feasible.
2.6. Protocol neutrality, extensions, and the risk of the new commercial tollbooth
Open protocols can shift bargaining power in two opposing directions at once. They can reduce technical lock-in, because merchants need fewer custom connectors and agents can access more products. But they can also expand the addressable commerce flowing through a small number of popular assistants. If a standardized interface makes it easy for every store to reach a dominant agent, the agent’s control over recommendation becomes more consequential, not less. The protocol lowers the cost of entering the market while the distribution surface potentially becomes more concentrated.
I call this tension the interoperability paradox. Greater compatibility does not guarantee greater competition at the consumer interface. A common payment network can support many banks, yet particular wallets may dominate where users initiate transactions. A common merchant feed can support several agents, yet a few interfaces may receive most user intent. A specification can be neutral while the ranking, identity, data access, or default checkout rules built around it are not. This is a reason to evaluate protocols as institutional arrangements, not merely as developer tools, and it is why the MIT Initiative on the Digital Economy’s observation that outcomes hinge on whether agents operate within or across platforms is the right frame.[21][22]
“AI agents fit the same niche for the same reasons.” — Peyman Shahidi, MIT Sloan [22]
Merchants should examine the total cost of access rather than the headline fee. A nominally free or open integration may still involve certification, intermediary fees, higher support costs, paid promotion, discounts required for visibility, restrictive product attributes, or dependence on platform-controlled identity. A payment arrangement may be inexpensive per transaction but transfer valuable customer data or constrain the methods available at checkout. A sponsored-offer program may increase sales while reducing incremental profit. The meaningful denominator is not technical setup cost; it is contribution margin and customer relationship value after the entire agent-mediated journey.
The policy questions follow naturally. Should an assistant disclose material commercial incentives that affect which offers it includes? Should platforms provide merchants with meaningful information about eligibility and exclusions? Can customers export their saved shopping preferences and consent records? How should liability be assigned when a protocol message is valid but the product recommendation is materially misleading? Does an agent provider use one ranking treatment for its own retail operations and another for rivals? These are competition and consumer-protection questions that arise after interoperability succeeds, not only when it fails.
By 2030, the decisive protocol may not be the one with the most elegant specification. It may be the one that aligns the incentives of consumers, merchants, agents, networks, and regulators while remaining technically adaptable. That alignment is difficult because each actor wants something different: the buyer wants accuracy and control; the merchant wants profitable access and a customer relationship; the assistant wants relevance, engagement, and revenue; the payment provider wants secure volume; the public wants contestable markets and enforceable rights. A commercial standard survives when these interests can coexist without requiring one actor to surrender all bargaining power.
The second structural insight follows: whoever writes the transaction language does not necessarily own the transaction, but whoever controls the interface, the ranking, the credentials, or the authorization defaults may shape the market even under an open protocol. Agentic Shelfspace is the arena in which those layers of authority meet.
Table 1. The principal agentic-commerce architectures as of October 2026
| Architecture | Principal sponsors | What it standardizes or controls | Checkout locus | Representative evidence |
| Agentic Commerce Protocol (ACP) | OpenAI, Stripe | Product feeds, discovery, checkout sessions, delegated payment; Apache 2.0 | Shifted in March 2026 to merchant-controlled checkout and in-ChatGPT apps | [3][7][8][61] |
| Universal Commerce Protocol (UCP) | Google, Shopify | Capability discovery, cart, identity linking, checkout, orders, extensions | Google Pay or merchant site; merchant remains seller of record | [2][9][10][11] |
| Link wallet for agents | Stripe (used by Muse, others) | Agent-side payment credential, single-use virtual cards, incremental authorization, purchase protection | Any merchant; 38x monthly growth reported Sept 2026 | [1][5][56] |
| Trusted Agent Protocol / Visa Intelligent Commerce | Visa (with Cloudflare) | Agent recognition, identity, intent signaling, tokenized credentials, Agent Directory | Merchant or supplier portal; B2B pilot live Oct 2026 | [4][29][49] |
| Agent Pay / Verifiable Intent / Agent Pay for Machines | Mastercard (with Google) | Authenticated participants, tamper-resistant intent, chargeback-compatible disputes, machine micropayments | Card rails; Trust Framework expanded Sept 2026 | [12][31][50][51] |
| Alexa for Shopping | Amazon | Closed-ecosystem agent with price alerts and auto-buy; external agents blocked under Conditions of Use | Amazon checkout only | [6][28][47][48] |
| Muse | Meta | Persistent personal agent in isolated VM; merchant connectors; Link, Shop Pay, PayPal | Merchant sites via agent browser | [5][54] |
| WebMCP / Agent Ready | PayPal, Braintree | Browser-facing structured capabilities on existing storefronts | Existing merchant checkout | [13][60] |
Section 3 — Who Controls Product Selection, Customer Intent, and Commercial Authority?
3.1. Recommendation is the new allocation decision
The market importance of search rankings is familiar, but an agent changes the user’s relationship to the ranking in a way that two decades of search-advertising economics did not anticipate. Conventional search offers visible links, advertisements, filters, and pages that a person can inspect, and the person remains the decision-maker who chooses to click. An agent may instead produce a conclusion: “These are the three best choices for your needs.” The answer compresses research, ranking, and explanation into a small set of statements. If it proceeds to checkout, the same chain can convert a recommendation into an action. The economic unit of visibility changes from an impression or a click to consideration by a delegated decision process, and consideration is a far scarcer good than an impression ever was.
An agent’s selection can be modeled as several successive gates, each of which can be governed by a different actor and each of which can quietly remove a merchant from the effective shelf. First is coverage: which sellers and products exist in the system’s accessible data? Second is retrieval: which become candidates for a particular task? Third is eligibility: which survive hard constraints such as price, location, availability, and authorized category? Fourth is ranking: in what order are eligible products evaluated and presented? Fifth is actionability: can the customer inspect, reserve, or purchase them through a supported path? Sixth is retention: who keeps the customer’s relationship after the transaction? Failure at any gate ends the merchant’s participation in that purchase, and the merchant may never learn which gate it failed.
These gates are frequently distributed across companies. A commerce platform supplies catalog data. An AI model interprets user intent. A retrieval system assembles candidates. Platform policy filters some content. A ranking model prioritizes offers. A wallet resolves payment. The retailer fulfills the order. Technical transparency at one stage does not reveal what happened at another. For example, a merchant may know its feed was successfully ingested and still not know why it failed to appear for a relevant consumer query. A buyer may see a well-written explanation but not the alternative products the system never retrieved. The experimental evidence that agents exhibit strong position biases — biases that vary across providers and model versions and persist even in text-only interfaces — means that there may be no universal notion of a “top” rank to appeal to, only a set of model-specific preferences that sellers will attempt to learn and exploit.[20]
This motivates a new economic variable: qualified recommendation share. For a well-defined set of shopping tasks where a merchant’s product genuinely meets the stated criteria, what fraction of those tasks results in the merchant being considered and presented? That measure differs from raw brand mentions. A brand can be mentioned often but only for low-value questions; it can appear as an alternative rather than the recommended choice; and it can be eligible for a purchase but absent because of incomplete inventory signals. Qualified recommendation share requires a carefully designed benchmark of consumer tasks, verified product attributes, and observed outcomes — and it is the single number a chief executive should ask for before approving another dollar of agentic marketing spend.
Research should also distinguish relevance from merchant preference. If a running shoe is unsuitable for a consumer’s foot shape or budget, exclusion is good service rather than discrimination. If two sellers offer identical goods at comparable delivered prices and one is systematically preferred because it uses a platform-affiliated checkout path, the issue is different in kind. Evaluating agentic shelfspace therefore cannot assume that equal visibility for every seller is desirable. It should ask whether the differences in treatment are explainable by the buyer’s legitimate requirements and by clearly disclosed commercial arrangements, and nothing else.
3.2. Advertising travels from the screen into the decision environment
The arrival of advertising in conversational systems complicates recommendation governance considerably. OpenAI began broader testing of ads in ChatGPT for free and lower-tier users in February 2026, and on October 5 announced a new visual ad format to be tested during image generation, expanded measurement integrations with third-party attribution and incrementality partners, and brand-suitability pilots with DoubleVerify and Integral Ad Science; throughout, it has maintained that ads are labeled, kept separate from generated content, and do not influence ChatGPT’s answers.[37] These are important corporate design commitments. They are not an independent measurement of every recommendation outcome, and they should be evaluated through auditing and disclosure rather than accepted or rejected on assertion alone. Google has pursued promotional commerce formats as well, including a Direct Offers initiative under which advertisers present special offers to buyers expressing purchase intent in AI Mode.[9]
A discount can be useful to consumers and may improve price competition. It can also change the set of products that capture attention. A promotion inserted clearly alongside a neutral recommendation is analytically different from a paid consideration invisibly changing the agent’s product shortlist. The design question is not whether commerce should be allowed to fund AI services — it plainly will — but whether the system preserves a meaningful boundary between advice to the buyer and inducements from the seller.
The distinction becomes harder because a consumer may delegate the interpretation of advertising itself. A person reading a webpage can notice a “Sponsored” label, however imperfectly. An agent summarizing product choices may incorporate a sponsored listing into its reasoning and then deliver a persuasive narrative without exposing the original label. The experimental literature of 2026 is sobering on this point. A preregistered study of 2,012 participants found that conversational LLM agents given a persuasive objective nearly tripled the rate at which users selected a sponsored product relative to traditional search placement, that explicit sponsorship labeling only marginally reduced the effect, and that most users did not notice the bias at all.[39] Researchers at Princeton and the University of Washington, testing twenty-three models under a soft sponsorship cue in the system prompt, found that most recommended a sponsored flight roughly twice as expensive as the cheapest option, with only a minority consistently presenting the lowest-cost choice.[40] And in September 2026, Davood Wadi and Yu Ma reported that assigning an agent a platform role, rather than a traveler role, reduced its skepticism toward sponsored listings and weakened the effect of disclosure in their tested settings.[38] None of this is evidence that any named commercial platform secretly manipulates rankings. It does identify a credible mechanism: the agent’s perceived principal, and the objectives it is given, change how it treats paid recommendations.
“Consumer trust, once compromised, is hard to repair.” — Bob Hedges, MIT Initiative on the Digital Economy [23]
The emerging policy principle should be disclosure continuity. If a paid influence exists at retrieval, ranking, or presentation, its material status should not disappear when information is transformed into an agent’s final recommendation. This need not require every technical feature of a ranking model to be public. It requires an auditable separation between user-serving criteria and financial incentives, and accessible disclosure when the latter materially affect what the buyer sees. In sensitive categories, additional safeguards may be justified. A financial incentive buried inside an automated choice process is not cured simply because a disclosure appeared in a machine-readable source the user never encountered.
The old marketing distinction among organic search, paid search, and retail media will not vanish; it will become more complex. Platforms may sell impressions within AI interfaces, paid inclusion in shopping experiences, sponsored comparisons, attribution services, or merchant tools. Different arrangements should carry different labels and rules. Paid advertising should not be casually conflated with ordinary merchant data delivery or processing fees; nor should a platform be allowed to evade scrutiny by calling an economically preferential placement a “technical optimization.” The contract, the selection mechanism, and the consumer effect matter more than the marketing name.
3.3. Customer intent is an economic asset — and a responsibility
Search advertising historically extracted value from a relatively thin representation of intent: keywords, location, device, and behavioral history. Conversational commerce can yield far richer information. A buyer might explain household composition, aesthetic preferences, allergies, travel dates, budget flexibility, accessibility needs, previous disappointments, and the reasons an item matters. An assistant with memory could combine these constraints across interactions. Shopify’s observation that three-quarters of AI-attributed purchases in its second quarter fell outside the top one hundred product categories — a car seat that fits three across a sedan, a standing desk for a small apartment — is a glimpse of how much long-tail intent conversational interfaces can elicit that keyword search never captured.[27] This context may produce better recommendations. It can also become valuable commercial intelligence about a person’s plans, sensitivities, and willingness to pay.
I propose the term intent custody for the obligation and power attached to holding and using this context. Custody is more precise than ownership: the individual may have supplied the information for a limited purpose, and several systems may process parts of it under different permissions. The assistant may need enough data to interpret the task; the merchant may need shipping and product information; the payment provider may need authorization evidence; the advertising system may seek a commercially useful signal. These are not automatically the same entitlement, and a system that treats them as one is making a policy choice whether or not it says so.
Several institutional models are possible. A platform could retain most context and give merchants only the information needed to fulfill an order. That may protect privacy but increase merchant dependency, because the brand loses access to preference and attribution information. A merchant could obtain extensive customer context through a direct integration, improving service but expanding exposure. A consumer-controlled agent could retain a portable preference profile and disclose selected attributes separately to each seller. Such a system would require robust permission management, secure identity, and practical ways to correct or delete stale preferences. None of these models is inevitable, and the choice among them will be made largely by default settings that most consumers will never change.
The commercial stakes are considerable. If an agent knows that a consumer will buy a particular category repeatedly, it can decide when to prompt replenishment, which sellers to compare, and whether to suggest substitutes. A merchant might have secured the first sale through a strong product, only to find that subsequent purchases are initiated by a general-purpose assistant that optimizes across competing brands. Conversely, an excellent retailer could gain access to consumers who would never have discovered it through advertising. Intent custody can redistribute power in both directions.
The boundary between personalization and exploitation must be made explicit rather than left to emerge. A system that knows a consumer is price-sensitive can search for discounts; the same knowledge could be used to estimate an individual’s maximum willingness to pay. A platform that knows a person urgently needs a replacement appliance may prioritize delivery speed; it could also surface higher-margin products. The Federal Trade Commission’s January 2025 surveillance-pricing study documented the breadth of behavioral and personal signals that intermediaries already use to tailor prices.[41] That investigation did not establish that agentic assistants were performing such pricing, but it describes a preexisting mechanism that becomes more salient as agents gather richer intent. For California residents and the businesses that serve them, the state’s privacy regulations that took effect in January 2026 — addressing risk assessments, cybersecurity audits, and certain automated decision-making uses, with compliance timelines extending into 2027 — add a further governance backdrop, even though not every product recommendation will qualify as a regulated significant decision.[42] The broader lesson is that consent for shopping assistance is not a universal waiver of privacy rights.
3.4. From conversion optimization to consumer-outcome optimization
A commercial platform can optimize for clicks, checkout completion, revenue, gross margin, repeat purchasing, or customer satisfaction. Those objectives are correlated but not identical. A recommendation that induces the highest immediate conversion may be less suitable than one that leads to fewer returns and greater long-term trust. A cheaper product can be poor value if it fails quickly; a highly rated product can be wrong for a buyer’s specific constraints; a recommendation can be plausible while omitting an essential compatibility detail. Agentic commerce forces the industry to confront the objective function behind product selection, because for the first time the selection is being made by a system whose objective function can, in principle, be inspected.
Consider a laptop purchase. The consumer requests a quiet, repairable machine suitable for a particular engineering application, within a fixed budget. A superficial agent might prioritize models with the strongest general reviews. A better agent verifies software compatibility, memory requirements, serviceability, warranty coverage, and current delivered prices. A responsible agent identifies gaps and asks one or two high-value clarifying questions before offering a shortlist. If the agent is rewarded solely for producing an order, it may skip those clarifications; if rewarded solely for conversational engagement, it may never reach a decision. The right performance measure must balance outcome quality, time, cost, and uncertainty, and it must be chosen by someone who is accountable for the consumer’s result rather than only for the platform’s revenue.
Academic evidence warns against assuming present capability is sufficient. The 2026 AgenticShop benchmark investigated personalized product curation in realistic web-shopping situations and reported substantial shortcomings in the evaluated systems; EComAgentBench modeled hidden consumer requirements distributed across an initial query, personal context, and clarifying interactions, and found long-horizon tasks difficult.[18][19] The purpose of these studies is to examine systems under controlled tasks, and their scores should not be extrapolated mechanically to every production agent or merchant category. They nevertheless show why purchase accuracy requires more than conversational fluency, and why Amazon’s reported statistic that assistant users spend 40 percent more per order should be read as a measure of engagement and basket size rather than as evidence that those users were better served.[6]
Consumer-outcome optimization needs observable post-purchase data, but access creates new dependencies. An assistant cannot measure whether its recommendation worked if it never learns about delivery, return, defect, or customer satisfaction. A merchant may not want to surrender all after-sale information to the assistant. Privacy-preserving or aggregated feedback mechanisms could help evaluate outcomes without transferring unnecessary personal data. Independent evaluation may also be needed, because the organization that earns revenue from a recommendation has an incentive to define success in its own terms.
I propose four core outcome dimensions. Constraint fidelity asks whether the selected product met every hard user requirement. Decision calibration asks whether the agent expressed uncertainty appropriately and escalated unclear choices. Total-cost accuracy tests whether tax, delivery, subscriptions, and other charges were faithfully represented before authorization. Post-purchase fitness measures whether the product performed as expected and the consumer could resolve problems. These dimensions should be evaluated by product category, because a minor accessory and a safety-critical household device do not present equivalent risks.
3.5. The merchant’s old customer relationship becomes a contested handoff
Retailers have spent decades building direct customer relationships through membership programs, email, apps, service, returns, and loyalty benefits. An assistant can potentially weaken the importance of direct traffic by attracting a buyer at the moment of need. That does not mean the merchant loses ownership of every subsequent interaction. OpenAI, Stripe, and Google have all described arrangements in which the merchant remains seller or merchant of record and handles fulfillment and support.[2][8][61] Legal seller status is an important boundary. It is not identical to controlling the consumer’s future attention.
An agent can mediate a relationship even when the merchant fulfills the order. The consumer may remember that “my assistant found it” rather than remembering the store. The assistant may own the reminder that triggers the next purchase. It may also control default substitution rules when an item is unavailable. From the retailer’s perspective, the strategic question becomes whether a sale obtained through an agent is a one-time acquisition or an enduring relationship. The answer depends on consent-based contact opportunities, service quality, brand differentiation, loyalty portability, and what information the agent shares. Shopify’s claim that new-buyer orders from AI channels arrive at nearly twice the rate of other channels is encouraging for merchants on the acquisition side; whether those buyers return directly or only through the agent is the question the next several quarters will answer.[27]
This is particularly important for independent brands. Large platforms can subsidize discovery, invest in protocol integrations, and negotiate payment terms. Smaller producers may benefit from the agent’s ability to compare niche attributes beyond brand fame. But they could also become more exposed to wholesale-style competition on price and specification if the agent compresses their stories into a narrow comparison table. A brand’s long-term defense is not to obscure product data. It is to make distinctive value verifiable: craftsmanship, fit, service, repair, provenance, community, reliability, and outcomes that matter to the user.
Loyalty arrangements create another test. An assistant that does not understand member pricing, accrued points, or product warranties may recommend a nominally cheaper rival even when the customer’s total economic advantage lies with an incumbent seller. UCP’s emphasis on merchant business logic and Universal Cart’s attention to card benefits show how the ecosystem is responding to such complexity, as does Walmart’s insistence on account linking and loyalty inside its ChatGPT app.[2][3][11] The more accurately agents can incorporate these benefits, the more a retailer’s relationship capital can survive outside its own website. But benefits must be represented fairly; a platform’s preferred loyalty partners should not be silently treated as the only ones that count.
3.6. An Agentic Shelfspace Audit and its metrics
To move the argument from metaphor to measurement, I propose an Agentic Shelfspace Audit conducted across standardized shopping tasks and real merchant availability. Each task would specify a consumer objective, mandatory constraints, acceptable trade-offs, geography, date, and approval limits. Evaluators would then observe which products an assistant retrieved, considered, recommended, and could transact. Tests would be repeated across agent providers and, where permitted, across diverse consumer profiles. The audit must respect platform terms, privacy, and applicable law; it should not assume access to proprietary reasoning traces, and it should adopt the randomized, causal design that the Columbia ACES framework has shown to be feasible.[20]
Table 2. Proposed Agentic Shelfspace Audit metrics
| Metric | Definition | Denominator | Primary user |
| Qualified coverage rate (QCR) | Share of verified eligible products the assistant can access in a defined test set | Products meeting prevalidated conditions, not the whole internet | Merchants, regulators |
| Qualified recommendation share (QRS) | Share of relevant tasks in which a given eligible merchant appears in the presented shortlist | Tasks for which that merchant genuinely offers a suitable product | Merchants, brands |
| Constraint fidelity (CF) | Share of final recommendations or completed purchases satisfying all specified hard constraints, severity-weighted | All completed tasks | Consumers, platforms |
| Offer integrity (OI) | Share of displayed prices, stock statements, shipping promises, and conditions matching independently checked merchant data at the relevant time | All displayed offers | Consumers, merchants |
| Authorization integrity (AI) | Share of actions staying within user-approved scope and value limits; any unauthorized payment is a critical event, never averaged away | All agent actions | Payment networks, regulators |
| Post-purchase resolution rate (PRR) | Share of disputes, returns, and order problems resolved through the promised service path within defined time limits | All reported problems | Consumers, merchants |
| Merchant multihoming cost (MMC) | Fully loaded cost of maintaining access, feeds, compliance, payments, and measurement across several agent channels | Per merchant, per channel | Merchants, startups |
| Commercial influence transparency (CIT) | Share of materially paid or affiliated treatments correctly labeled to users and available for audit | All paid or affiliated treatments | Regulators, consumers |
No composite index should conceal safety or consent failures. A system with excellent recommendation relevance but one unauthorized purchase can be unacceptable for a particular deployment. Likewise, a merchant’s low ranking need not constitute unfair treatment if its offer is objectively less appropriate. Audits should report disaggregated metrics and confidence intervals, document the task sample, and distinguish missing data from adverse results. Independent testing can be supplemented by supervised merchant and platform records, but it should not depend exclusively on the platform’s chosen success stories.
The audit also requires counterfactual thinking. Would the consumer have found a better option through ordinary search? Did the assistant improve the decision or simply shorten the path to a familiar seller? Did a sponsored offer lower the customer’s total cost or merely move attention from an equivalent non-sponsor? A/B tests can measure behavior, while structured product-quality evaluation can assess whether that behavior served the buyer. Both are necessary. Without outcome assessment, the industry may mistake greater automation for greater welfare.
The third structural insight is thus a governance proposition: Agentic Shelfspace should be evaluated as a decision-allocation system, not simply as a traffic source. Ranking, intent custody, commercial influence, and authorization jointly determine who benefits from the new storefront. Better consumer outcomes and competitive merchant access are goals to be demonstrated, not assumed.
Section 4 — The New Business Model of Retailers, Brands, Startups, and Technology Platforms
4.1. A sale is not profitable merely because an agent completes it
The excitement around conversational conversion can obscure basic retail economics, and the most important corrective is to insist on the distinction between volume and contribution. A merchant earns revenue when a transaction completes, but its contribution depends on product cost, shipping, payment fees, returns, customer service, advertising, discounts, fraud, and the cost of maintaining the distribution channel. An agent might raise conversion and simultaneously reduce margin by increasing promotional dependence or generating orders with high return rates. An accurate economic model must separate the volume of transactions from the profitability and durability of the customer relationship — a separation that the strongest 2026 disclosures, to their credit, began to make. Shopify’s leadership was explicit that agentic transactions carry the same economics as any online-store transaction for Shopify itself, that the volume remained small relative to a $116 billion quarter of gross merchandise volume, and that the dollars appeared incremental rather than cannibalizing search.[26][27]
“Agentic transactions carry the exact same economics as an online store transaction.” — Harley Finkelstein, Shopify [27]
Consider a purely illustrative merchant selling an item for $100. Suppose product cost is $55, fulfillment and expected returns cost $12, payments and fraud prevention cost $4, and ordinary marketing costs $9. The $20 contribution remaining before fixed costs looks attractive. Now suppose an agent channel replaces some marketing expense with an $8 combination of referral services and incremental discounting while introducing $3 of integration and support overhead. If sales are genuinely incremental and the original $9 acquisition expense disappears, the merchant may improve its economics. If the agent merely redirects loyal buyers who would have purchased anyway, the merchant could pay new distribution costs without receiving new demand. Neither outcome can be inferred from gross sales growth alone. The assumptions are illustrative, not estimates of any company’s actual fees.
Table 3. Illustrative unit economics of an agent-mediated order (hypothetical, per $100 sale)
| Line item | Direct channel | Agent channel, incremental order | Agent channel, redirected loyal buyer |
| Revenue | $100 | $100 | $100 |
| Product cost | ($55) | ($55) | ($55) |
| Fulfillment and expected returns | ($12) | ($12) | ($12) |
| Payments and fraud prevention | ($4) | ($4) | ($4) |
| Ordinary marketing | ($9) | $0 | ($9) still incurred elsewhere |
| Agent referral services and discounting | — | ($8) | ($8) |
| Integration and support overhead | — | ($3) | ($3) |
| Contribution before fixed costs | $20 | $18 on a sale that would not otherwise exist | $9 on a sale that would have happened anyway |
The correct accounting therefore begins with incremental contribution margin per agent-mediated order, not gross merchandise value. The merchant should measure whether the order would have occurred without the agent channel, which costs were avoided, what new fees appeared, and whether the buyer returns directly or through the intermediary. In practice, causal attribution is difficult because an individual may see an advertisement, ask an agent, read independent reviews, and then buy on a merchant site. Treating the final click as the sole cause of the purchase will systematically misprice the contribution of earlier interactions — and will systematically flatter whichever intermediary stands closest to the checkout.
Agentic distribution may be especially valuable for products requiring comparative advice: complex electronics, specialized equipment, home improvement, business software, or purchases with detailed compatibility requirements. It may be less transformative for impulse goods selected through images or for branded products whose demand is primarily cultural. Yet even an established brand can benefit from an agent correctly answering product-fit questions, reducing pre-sale support, and matching accessories. The economic opportunity is not confined to receiving a new stream of customers; it includes lower decision friction and fewer avoidable wrong purchases.
Retail executives should build a channel-level profit-and-loss statement that includes revenue, gross margin, promotions, payment charges, fulfillment costs, returns, support, platform services, attributable compute, and repeat-customer behavior. They should track the marginal effect on direct, search, marketplace, and social channels rather than declare every agent order incremental. Boards should demand a clear distinction among adoption, transaction throughput, and profit improvement. Fast-growing new channels often look exceptional before cannibalization and service costs are measured.
4.2. Brand differentiation migrates into verifiable product claims
Traditional brands use packaging, stores, storytelling, influencers, advertising, and customer experience to create preference. In agentic commerce, these capabilities still matter, but they must survive translation into machine-readable evidence. An assistant choosing a winter jacket may evaluate insulation, water resistance, size, repair policy, labor standards, shipping time, and price. A premium brand must make the reasons for paying more understandable in that context. It cannot rely solely on the shopper encountering an emotionally persuasive landing page, because the shopper may never encounter the landing page at all.
This is not the end of branding. It may reward brands with stronger genuine differentiation. An assistant can explain why a specific manufacturer’s repair network or long warranty matters to a particular person, even when the brand has a smaller advertising budget. It can surface a niche artisan to a buyer who values provenance, or identify a small manufacturer whose product uniquely satisfies accessibility requirements. However, this benefit depends on unbiased retrieval and trustworthy evidence. If an agent merely repeats the most abundant or best-optimized text, it may favor well-resourced content producers over better goods, and the research on seller-side manipulation suggests that the best-optimized text will not always belong to the best product.[20]
Brands should construct a product-evidence layer: authoritative specifications; variant-level identifiers; current stock and price; quantified durability or performance where verifiable; warranty and return rules; original imagery; compatibility documentation; safety and compliance certifications; and channels for correcting errors. Independent reviews and test results can supplement merchant claims. The point is not to flood AI systems with promotional language. It is to reduce ambiguity between a buyer’s stated need and the offering’s actual properties.
The risk of commoditization is real when agents normalize products to a few comparison fields. Sellers could respond by creating ever-more artificial attributes to look unique, repeating the old search-optimization contest in a new vocabulary. Better practice would make comparisons more complete rather than more deceptive. A retailer that can demonstrate superior service or reliability should want those dimensions included. An assistant that reduces everything to price could harm both consumers and high-quality suppliers. The design of the evaluation rubric is therefore commercially consequential, and brands have a legitimate interest in shaping it.
Product information also changes over time. Inventory may be depleted, a model may be recalled, an accessory may be replaced, or a discount may end. A brand cannot treat a static AI training corpus as a reliable catalog. Dynamic feeds, provenance, update timestamps, and verifiable source attribution become essential. The software infrastructure responsible for those tasks may be less glamorous than a frontier model, but it could become a durable and valuable part of the commerce stack, and it is already where Shopify, Google’s Merchant Center, and OpenAI’s feed APIs are competing for merchant attention.
4.3. A new startup map: agent-native intermediaries and merchant infrastructure
The rise of agentic shopping creates startup opportunities at several points between consumer intent and completed order. Some are familiar categories with new interfaces; others involve genuinely new coordination problems. The strongest opportunities solve an expensive failure point rather than merely decorating a conventional storefront with a chatbot.
One category is merchant-data infrastructure: tools that validate catalog attributes, reconcile product identifiers, map variants, compare merchant policies, and publish accurate information to multiple agents. A second is recommendation verification: independent services that test whether a proposed product actually meets hard constraints, whether quoted prices are current, and whether important alternatives were excluded — the commercial form of the audit described in Section 3. A third is transaction trust: authorization ledgers, agent identity verification, fraud analytics, mandate management, and recovery from incomplete checkout, a category in which the card networks are both potential customers and potential competitors. A fourth is post-purchase orchestration: systems that track shipments, coordinate returns, handle warranties, and help customers resolve exceptions across several sellers.
A fifth category is agent-native merchant services. A retailer or small brand may need a specialized assistant that understands its products and can interact with consumer-controlled agents. Such software might answer technical compatibility questions, negotiate allowable configuration options, offer local pickup, or explain service agreements in standardized form. Salesforce’s holiday data suggesting that retailers deploying their own branded agents grew markedly faster than those that did not — whatever the attribution caveats — indicates that merchants perceive value in having a representative of their own in the conversation.[36] A sixth is measurement and attribution: services that estimate qualified recommendation share, merchant inclusion, outcome quality, customer acquisition incrementality, and channel profitability. These are not guaranteed markets. Each must overcome data access, trust, and distribution barriers, and some functions may be absorbed by large commerce platforms or payment networks.
One promising distinction is between agents that sell and tools that help agents buy well. A seller-facing agent is loyal to the merchant within lawful and transparent bounds. A consumer-facing assistant is expected to serve buyer preferences. Confusing these roles creates conflicts — exactly the conflicts the role-assignment research has begun to quantify.[38] A startup can gain credibility by making its principal explicit: it is a merchant representative, an independent comparison tool, an authorized buyer agent, or a neutral transaction service. Attempts to portray a paid sales agent as an impartial buyer adviser could erode the very trust the category needs.
Startups should also resist dependence on a single protocol or hyperscaler. A vendor built only around one proprietary shopping interface could face abrupt changes in ranking rules, onboarding policy, fees, or API access; the merchants who built for Instant Checkout in late 2025 learned this within six months.[62] Multihoming raises engineering expense but preserves bargaining power. Open standards can reduce that cost, though they do not eliminate business-specific extensions. Investors should test whether the startup owns scarce capability — verified data, merchant relationships, a difficult compliance workflow, unique consumer consent, or measurable outcomes — or merely resells access that a platform can later internalize.
4.4. Independent merchants, local commerce, and the changing meaning of proximity
Agentic Shelfspace could either improve or worsen the prospects of local commerce, and which it does will depend on choices still being made. A customer searching conventionally may never discover a neighborhood shop that stocks a hard-to-find product. A capable assistant with accurate local inventory could identify that shop, verify pickup availability, and route the order to it. For a merchant with limited advertising resources, this could be an important distribution improvement. It could also reduce environmental and delivery costs when a nearby product substitutes for a long-distance shipment — though such benefits must be measured rather than assumed. The ICSC–McKinsey report of April 2026 drew the complementary conclusion for physical retail: that stores will need to choose between a convenience-optimized identity and a discovery-led one, because the store caught in between is the one an agent can most easily bypass.[34]
The obstacles are practical. Many independent stores lack real-time inventory feeds, consistent product identifiers, online reservation systems, or technical resources for multiple protocol integrations. They may also have complex local policies for returns, special orders, deposits, or services bundled with products. A large merchant’s sophisticated information systems are an advantage. If an agent requires perfect structured data to consider an offer, the resulting market may be biased toward the digitally mature, even when a local seller has the best real-world option.
Commerce platforms such as Shopify can reduce this entry burden by connecting merchant catalogs to larger systems, and the default inclusion of eligible Shopify merchants in ChatGPT’s discovery layer and in Universal Cart is the clearest 2026 example of a platform lowering the barrier on behalf of the small.[3][11] Payment services can make agent checkout practical without bespoke integration. Local chambers of commerce, economic-development agencies, and small-business support programs could help merchants improve product data, identity verification, and secure digital payments. Such assistance would be more useful than generic exhortations to “adopt AI.” The task is specific: make legitimate local offerings reliably discoverable and transaction-ready without requiring a small business to hire an AI engineering department.
The local dimension also complicates recommendation criteria. A consumer may value same-day pickup, repair support, avoiding shipping, face-to-face product advice, or a local return counter. These preferences should be represented as first-class attributes, not afterthoughts to national price comparison. An assistant that only optimizes the displayed sticker price may choose an inferior outcome. The economic importance of place remains even if discovery becomes conversational and geographically dispersed.
There are risks of fraud and impersonation. An agent may encounter false storefronts, counterfeit listings, fabricated reviews, or sellers using misleading location claims. Strong merchant identity and marketplace verification remain necessary. Because a machine can complete more steps more quickly than a human, automation can scale legitimate commerce and fraudulent activity simultaneously. A local merchant’s trust advantage may depend on being verifiably local and responsible for real post-sale support, not merely claiming those attributes in a feed.
4.5. Loyalty, returns, and service determine who retains the buyer
The sale is only the middle of the customer relationship. Returns, defects, exchanges, support, subscription changes, warranties, recalls, and loyalty rewards often determine whether a buyer will return. A shopping assistant that recommends a product but cannot help with a failed delivery may offer less total value than an ordinary retailer with excellent service. The strongest agentic experiences will therefore require durable order references, clear seller identity, authorized service actions, and transparent handoffs — which is one reason Link’s extension of purchase protection to agent transactions, and Meta’s emphasis on it at Muse’s launch, mattered more than their modest scale suggested.[1][5]
This opens a potentially important distinction between transaction authority and service authority. A shopper might authorize an assistant to purchase a product but not automatically return it, accept a store credit, share medical or identity documents, or agree to a subscription renewal. Permissions should be task-specific and revocable. A service agent may need to communicate with a merchant’s support system without knowing the buyer’s complete payment credentials. Durable transaction records can help establish what the customer approved and what the merchant promised.
Retailers and brands should decide which after-sale interactions they want to expose to external assistants. A repair-status lookup might be low risk. Changing an address on a valuable shipment might require stronger authentication. Authorizing a replacement item may require explicit confirmation. Failure to distinguish these actions could create security and liability problems — the very problems Amazon cited when it blocked Muse from reaching account pages and order history.[48] At the same time, refusing all automated service could make a retailer inconvenient compared with competitors whose customer-service APIs work with legitimate agents.
Loyalty programs face a deeper strategic choice. If an assistant evaluates all offers across retailers, loyalty can be a user benefit or a form of commercial lock-in. A membership discount, accumulated points, or free repair can rationally change a recommendation. However, a program that makes benefits difficult to understand or redeem across agents may lose relevance. The winners will be those able to express loyalty value clearly and honor it reliably, without hiding key restrictions.
These relationships explain why OpenAI’s March 2026 decision to support merchant-controlled checkout is strategically significant beyond the immediate product. Retailers need flexibility over account linking, loyalty, payment methods, and returns, while consumers may wish to preserve familiar service channels.[3] A platform can create value by bringing qualified buyers without displacing every merchant interaction. The long-term boundary will likely vary by category and customer preference. There is no economic law requiring a single entity to own discovery, recommendation, payment, and post-sale service.
4.6. From checkout fees to inference economics across all five layers
The revenue architecture of agentic commerce may combine subscription fees, commerce-enablement services, payments revenue, merchant tooling, advertising, and perhaps transaction-related compensation. These streams should not be conflated. A platform may monetize its assistant through subscriptions while promising that organic recommendations remain independent of ads. A payment provider earns from successfully processed authorized transactions, which is why Visa and Mastercard can describe agentic commerce as addressable-market expansion with a straight face.[30][31] A merchant platform charges for software and services. An advertising system sells attention or promotion. A consumer-controlled agent could charge the buyer directly for unbiased procurement assistance. Each model creates a different incentive for how shelfspace is allocated.
At the application layer, per-task costs include model tokens, retrieval, product-feed queries, verification, tool calls, session state, wallet interfaces, and customer support. A complex purchase may involve many model calls before a decision. A recurring replenishment task may require little active inference if most checks are event-driven. The economics therefore depend heavily on system design, not merely on the frontier model’s advertised cost per million tokens. The relevant denominator is the cost per successful, authorized, satisfactory shopping outcome.
For illustration, suppose an agent spends $0.12 in model and tool costs on a complex shopping journey that yields a genuinely incremental $20 merchant contribution before agent-related fees. The computational cost is small relative to the commercial value. But if it must conduct many failed searches, use costly model reasoning for low-value items, or provide intensive post-purchase support, the margin can deteriorate. If agents run background price checks for hundreds of millions of users — Amazon’s 350 million assistant users and Google’s Universal Cart alerts give a sense of the potential population — unnecessary polling can increase infrastructure demand without proportional sales.[2][6] Figures here are hypothetical; they demonstrate unit-economic sensitivity, not observed industry costs.
At Layer Four, model providers compete on instruction fidelity, product reasoning, tool reliability, uncertainty management, and cost. At Layers Two and Three, chips and datacenters support increasingly high-volume inference, and the capital commitments disclosed in 2026 earnings — Amazon’s roughly $220 billion annual plan foremost among them — are in part a bet that agentic workloads of this kind will materialize.[28] At Layer One, energy prices, reliability, and grid access influence the cost and geography of that capacity. The connection does not imply that one consumer order can be translated mechanically into megawatts. It implies that sustained growth in continuous agentic workflows could become one among several drivers of inference infrastructure investment. The industry should resist both exaggerating and dismissing that connection.
The strongest economic feedback loop would be a virtuous one: lower inference costs allow richer comparison; richer comparison produces better consumer outcomes; improved trust supports more use; transaction and service revenue finances reliable infrastructure; and scale encourages technical efficiency. A weaker loop would generate more unsolicited agent activity, more advertising pressure, more low-quality recommendations, and higher operating costs without improving welfare. Both are plausible. The Five-Layer AI Economy helps reveal that the upper-layer commercial design determines whether lower-layer capital produces durable value.
The fourth structural insight is that Agentic Shelfspace creates a new distribution market, but the economic winner is not necessarily the company collecting the most checkouts. Sustainable advantage depends on incremental value, trust, cost discipline, qualified discovery, customer retention, and the capacity to work across several technical and commercial systems.
Table 4. Agentic commerce in the disclosures of six firms, through the second quarter of 2026
| Firm | Reporting period | Headline results | What was said about agentic commerce | Source |
| Shopify | Q2 2026 (Aug 5) | GMV $115.6B (+32%); revenue $3.58B (+34%); FCF margin 18%; payments penetration 68% | AI-driven traffic and orders tripled YoY; new-buyer orders ~2x other channels; 75% of AI-attributed purchases outside top 100 categories; agentic GMV small but incremental | [26][27] |
| Amazon | Q2 2026 (Jul 30) | Net sales $200.6B (+20%); operating income $27.5B (+43%); ads $19.8B (+26%); Q2 capex $53.1B | Alexa for Shopping used by 350M+ customers in 12 months; active users ~2x; assistant users spend 40%+ more per order; external agents blocked | [6][28][59] |
| Visa | Fiscal Q3 2026 (Jul 28) | Net revenue $11.6B (+14%); payments volume >$4T; VAS +34% | Partnerships with OpenAI and Meta; Agent Score, Agent Directory, Token Assurance Framework; “when, not if” | [29][30] |
| Mastercard | Q2 2026 (Jul 30) | Net revenue ~$9.3B (+14% reported); adjusted EPS $5.04 | Agent Pay live transactions; Agent Pay for Machines with 30+ partners; Verifiable Intent with Google; “cards will prevail” | [31][32] |
| PayPal | Q2 2026 (Jul 28) | Revenue $8.68B (+5%); TPV $486B (+10%); branded checkout +2% cc | Agentic payments and digital identity expected to become meaningful contributors from 2028 | [33] |
| Stripe (private) | Sept 2026 disclosure | Link agentic purchases +38x month over month; Agentic Commerce Suite | Incremental authorization, spending-context recommendations, purchase protection for agents | [1][53] |
Section 5 — The Structure of Machine-Mediated Commerce in 2030
5.1. Four plausible futures, not a single inevitable winner
Forecasting agentic commerce requires more than projecting a growth rate from early transactions. The trajectory depends on model reliability, trust architecture, merchant participation, consumer willingness to delegate, regulation, and the economics of inference. These variables can interact in several different ways, and the most-cited projections in the field are explicitly scenario-based rather than point forecasts: McKinsey’s estimate that agents could orchestrate $3 trillion to $5 trillion of global consumer commerce by 2030, with up to $1 trillion of U.S. business-to-consumer retail revenue, assumes agent mediation of roughly a fifth of U.S. consumer spend under a moderate adoption path and is framed by its authors as a curve of delegation levels rather than a single destination.[34][35] Mastercard’s own September 2026 projection — one in ten consumers routinely using agents to purchase by 2030 — points in the same direction while implying a slower pace.[12] To avoid treating the most dramatic scenario as destiny, I propose four conditional market structures for 2030.
Scenario A: The interoperable bazaar. Common commerce and payment interfaces become sufficiently robust that merchants participate through a modest number of integrations. Consumers choose among several personal and general-purpose agents, carrying preferences and payment permissions where feasible. Agents compete on recommendation quality, privacy, service, and price. Merchant visibility remains contested but is open to meaningful independent testing. This scenario favors modular standards, portable data, strong agent identity, and relatively low switching costs. It is a plausible aspiration, not a guaranteed consequence of ACP, UCP, or any other standard.
Scenario B: The assistant oligopoly. A small number of general-purpose interfaces accumulate most high-intent consumer interactions. Protocols remain open and many merchants connect, yet the consumer’s choice set is primarily assembled by a few agents with large distribution, memory, and advertising businesses. Retailers compete fiercely for qualified recommendation share and may become dependent on those interfaces for new customers. This market can deliver excellent convenience, but it requires scrutiny of commercial incentives, self-preferencing, portability, and access. In this scenario, openness at the technical layer coexists with concentration at the attention layer — the interoperability paradox realized.
Scenario C: The marketplace fortress. Large retailers and ecosystems develop highly capable internal assistants that exploit transaction history, fulfillment reliability, returns infrastructure, and exclusive benefits, while using contractual terms to exclude external agents. Consumers use these agents for routine purchasing and external agents mainly for cross-market comparison or unusual categories. Independent sellers often reach buyers through established marketplaces rather than direct agent connections. This scenario can produce strong service integration, but the diversity of products considered is bounded by each fortress’s ecosystem. Amazon’s integration of shopping, fulfillment, advertising, and customer service, combined with its decision to block Perplexity’s Comet and Meta’s Muse, makes this structure especially relevant as an analytical possibility.[6][47][48]
Scenario D: The sovereign consumer agent. Personal assistants operate more explicitly on behalf of individuals, with user-controlled preferences, portable consent, independent product verification, and permission-bound payments. Merchants publish offers through interoperable interfaces; buyers can inspect why the agent selected one offer over another. This model would require strong security, sustainable consumer-side economics, and practical solutions to personal data custody. It could coexist with all the other scenarios for particular high-value categories rather than replace them. A consumer might use a marketplace assistant for groceries and an independent agent for a major appliance or business procurement.
Table 5. Four conditional futures for machine-mediated commerce in 2030
| Scenario | Who assembles the consideration set | Protocol openness | Attention concentration | Principal risk | Leading indicators to watch |
| A. Interoperable bazaar | Many competing agents | High | Low to moderate | Fragmented trust and liability | Falling merchant multihoming cost; portable consent records; audited QRS parity |
| B. Assistant oligopoly | A few general-purpose interfaces | High | High | Self-preferencing, sponsored distortion, dependency | Ad revenue share of agent platforms; share of retail referrals from top three agents; disclosure litigation |
| C. Marketplace fortress | Vertically integrated retailers | Low | High within ecosystems | Bounded choice; exclusion of external agents | Terms-of-service enforcement against agents; growth of in-ecosystem assistants; antitrust complaints |
| D. Sovereign consumer agent | User-controlled agents | High | Low | Security burden; consumer-side economics | Paid consumer-agent subscriptions; independent verification services; portable preference standards |
These scenarios are not mutually exclusive across regions or sectors. A 2030 consumer may encounter all four in one week. The variables that determine their balance include the portability of consent, the cost of agent multihoming, the reliability of cross-platform product data, the degree of advertising influence, the quality of third-party audits, and the availability of dispute remedies. The paper makes no claim to know their future market shares. Its purpose is to expose strategic alternatives and the institutions that could steer the outcome.
5.2. Consumer protection moves upstream from checkout to recommendation
Conventional commerce protections concentrate heavily on payment authorization, truthful advertising, returns, and seller accountability. Agentic commerce preserves those needs and adds earlier decision points. A buyer can be harmed before entering payment credentials if an assistant invents product attributes, excludes suitable alternatives for hidden commercial reasons, or misrepresents a seller’s warranty. A transaction can be technically authorized and still be based on materially misleading advice. Security and consumer welfare are related but distinct, and a regulatory regime that protects only the second half of the journey will miss where the new harms concentrate.
At minimum, agent-mediated shopping should preserve four boundaries. The information boundary distinguishes verified product facts from uncertain or seller-supplied claims. The commercial boundary distinguishes organic assistance from paid promotion and platform affiliation. The authorization boundary distinguishes research, recommendation, cart preparation, reservation, and irreversible purchase. The remedy boundary identifies which party must resolve non-delivery, defects, unauthorized charges, or incorrect information. These boundaries should remain intelligible to ordinary consumers, not merely present in technical documentation.
U.S. consumer-protection law already applies to deceptive advertising and unfair practices. The FTC’s endorsement guidance covers material advertising relationships in relevant contexts, and its surveillance-pricing work highlights risks created by intensive personal-data use.[41][44] On July 1, 2026, the FTC voted to seek public comment on a proposed policy statement concerning what it called the suppression of accuracy in AI systems, taking the position that AI companies have represented their systems as designed to produce accurate and faithful output and that distorting outputs toward undisclosed objectives could constitute deception under Section 5 of the FTC Act; the statement, which grew out of a December executive order and also advanced a theory that certain state AI laws are preempted, drew comments through July 31.[43] That proposal should not be described as a finalized agentic-commerce regulation, and its stated concern was ideological rather than commercial steering. Its significance for this paper is nonetheless real: the conduct of AI systems, not only conventional promotional copy, has become an explicit subject of federal consumer-protection debate, and the logic that an undisclosed objective can deceive a user who expected faithfulness applies with at least equal force to an undisclosed commercial objective. Industry participants themselves acknowledge that the rules are trailing the products.
“Regulation always follows innovation.” — Courtney Robinson, Akoya, at Fortune Brainstorm Tech 2026 [55]
The practical standard should be stronger than a lengthy disclaimer. Before a consequential purchase, the system should present the item, seller, total price, key recurring obligations, delivery conditions, and approval scope in a form the user can understand. It should explain when it cannot verify an important claim. Consumers need accessible histories of authorized actions and simple ways to revoke permissions. Transaction service providers should be able to determine which agent acted, under whose mandate, and within which controls. These safeguards can be designed to minimize unnecessary data sharing.
Not every mistaken recommendation is fraud or a violation of law. An agent can make a good-faith comparison using incomplete information; a seller can change inventory between retrieval and checkout; a consumer can prefer an unusual trade-off. Governance should distinguish error, negligence, deception, malicious manipulation, and unauthorized action. Overly broad liability could suppress useful innovation; nonexistent accountability could make delegated commerce untrustworthy. The challenge is proportionate responsibility tied to the stage and the actor involved.
5.3. Competition law confronts a potentially new gateway
The concentration question is not hypothetical in the underlying digital economy. U.S. and European authorities have already scrutinized control over search distribution, ranking, and platform access. In September 2025, the U.S. Department of Justice described remedies in its Google search monopolization case, including restrictions on certain distribution and revenue-sharing arrangements, and the appeals that followed in early 2026 ensure the matter will shape the legal environment for years.[45] The analytical relevance is clear whatever the eventual outcome: control over the default path by which users reach information can affect competition even when alternative services technically exist.
The European Union’s Digital Markets Act offers a second and more pointed reference. On July 23, 2026, the European Commission adopted two non-compliance decisions against Google, fining it €460 million for giving its own shopping, hotel, transport, and sports services preferential placement, visuals, and filters in Search over rival services, and €430 million for restricting app developers from steering users to alternative purchase channels in Google Play, with sixty days to comply or face periodic penalty payments.[46] Those decisions concerned designated services and specific conduct, not a determination that AI shopping assistants are gatekeepers. Still, the first DMA ruling on ranking self-preferencing arrived in the same month that the same company was preparing to launch a cross-retailer cart inside its AI surfaces, and the juxtaposition will not be lost on complainants.
“Determination to apply the Digital Markets Act to safeguard business and innovation.” — Henna Virkkunen, European Commission [46]
Agentic Shelfspace could create a gateway with different characteristics from classic search. A search engine presents a ranked list; an agent may interpret the user’s objectives, choose a shortlist, and sometimes execute the result. That combination can strengthen the influence of defaults. If a user trusts the assistant’s answer, the cost of inspecting alternatives may feel high even when links to competitors are available. A nominal choice to click “show more” is not equivalent to a competitive market if the system’s first recommendation systematically channels demand.
Competition analysis should therefore examine several behaviors: exclusive distribution or wallet arrangements; discriminatory merchant eligibility; platform self-preferencing; tying a commerce tool to unrelated services; restrictions on external checkout or steering; anticompetitive use of merchant transaction data; refusal to deal with legitimate third-party agents where the refusing party operates competing agents of its own; and acquisitions of essential comparison, catalog, or attribution infrastructure. None is automatically unlawful. Relevant market definition, competitive effects, business justifications, and applicable jurisdiction remain necessary. The public-interest objective is not equal sales for every merchant; it is meaningful contestability and consumer choice.
Open standards and auditability can assist enforcement without requiring regulators to design a universal recommendation algorithm. A platform could document eligibility criteria, disclose material payment or affiliate preferences, and provide structured explanations for major merchant exclusion categories. Independent auditors could test whether similarly situated merchants receive systematically different treatment, controlling for product fit and price. Public authorities could seek evidence of switching costs and exclusive arrangements. These measures would help distinguish legitimate quality differentiation from the abuse of gatekeeping power.
5.4. Identity, authorization, and liability form the constitutional rules of a purchase
An AI agent may be capable of clicking a button, but capacity is not authority. In law and business, the actor entitled to commit a person to a transaction depends on consent, contract, agency principles, and relevant payment rules. The technology industry’s use of the term agent does not automatically confer legal agency in every jurisdiction. The central design problem is to translate a human’s bounded instructions into verifiable technical permissions and to preserve evidence when a dispute arises. The Ninth Circuit’s August 2026 holding that it was the user, not the agent’s developer, who accessed Amazon’s systems when directing Comet to shop — a comparison the court drew to a Safari user visiting a site from her own machine — is an early judicial recognition that the human remains the legally operative actor, even as it left the contractual terms governing agents entirely intact.[47]
Google’s Agent Payments Protocol work, Mastercard’s Verifiable Intent, Visa’s Trusted Agent Protocol, and Stripe’s incremental-authorization logic are complementary pieces of a wider trust architecture rather than interchangeable standards.[1][49][50][52] A robust transaction may require evidence of which user approved the action, which agent executed it, which merchant accepted it, what item and total were authorized, when the approval occurred, and whether any material conditions changed. Mastercard’s framing of its Trust Framework as five elements — identity, intent, controls, execution, and intelligence — is as good a summary of the required evidence as the industry has produced.[12]
“Trust is the foundation that will determine how far agentic commerce goes.” — Stephanie Cohen, Cloudflare [12]
Different actions need different permissions. A buyer might grant a daily ceiling for consumable office supplies on an approved vendor list, but demand manual sign-off for a new supplier. A consumer might authorize a price-alert service but not automatic purchase. A business might authorize invoice payment only after a separate approval workflow has matched an invoice to a purchase order. The Visa–BILL–Billtrust pilot illustrates a payment agent completing an already approved obligation, which is a narrower and more defensible proposition than unrestricted autonomous spending, and the pilot’s design — tokenized credentials instead of shared portal logins, and no agent decision-making about whether to pay — shows what proportionate authority looks like in practice.[4]
Disputes expose the importance of records. If an agent buys the wrong product, was the user instruction ambiguous, the agent’s reasoning defective, the merchant’s data incorrect, or the checkout confirmation misleading? If a credential is stolen, did the wallet properly authenticate the agent? If delivery fails, did the seller fulfill its obligation? A well-designed system should make it possible to identify the relevant stage and responsible party. Shared responsibility may be appropriate in some situations; automatic blame of the merchant or model provider in every case is not. Link’s decision to offer purchase protection on eligible agent purchases, and Mastercard’s reliance on existing chargeback processes to resolve challenges under Verifiable Intent, are two different answers to the same question of who absorbs the loss when the machine errs.[1][50]
The industry’s trust designs should also minimize the amount of sensitive information an agent must retain. A system can prove authorization through purpose-bound tokens or mandate records without copying complete card details into conversation history. Payment credentials should be revocable, transaction limits enforceable, and access appropriately scoped. Such controls may initially add friction, but they are essential to the repeat use on which a durable agentic economy depends.
5.5. Regional fragmentation and the politics of cross-border commerce
Agentic shopping may appear borderless when a user speaks to a global assistant, but commerce remains local in important respects. Sales tax, value-added tax, customs, sanctions, consumer protection, returns, product safety, data protection, payment acceptance, and delivery rules differ. A product available in one region may be prohibited, impractical, or subject to very different obligations in another. A model that reasons well about products but ignores jurisdiction cannot reliably serve a cross-border buyer. The fact that every major 2026 consumer launch — Universal Cart, Muse, the ChatGPT discovery revamp — was confined to the United States at introduction is a reminder that the legal and payments perimeter still binds even the most ambitious agent.[2][3][5]
The United States is itself a patchwork of consumer and privacy rules. California’s privacy requirements illustrate how data handling and certain automated decisions may create compliance obligations; other states differ, and the FTC’s July 2026 proposal to treat certain state AI statutes as impliedly preempted signals that the boundary between federal and state authority over AI conduct is itself contested.[42][43] Federal rules and enforcement provide additional layers, and payment networks impose contractual standards. In the European Union, the DMA and the data-protection framework shape the conduct of certain platforms and processors, though specific obligations vary by designation and service.[46] A multinational merchant cannot assume that a protocol’s technical compatibility equals legal permission to sell every product everywhere.
Cross-border agentic commerce may benefit smaller exporters by making language, catalog normalization, and price comparison easier. An assistant could explain duties, translate a warranty, and identify local alternatives. It can also introduce risks if it misstates import conditions or encourages a purchase the seller cannot legally fulfill. Government agencies, customs authorities, and payment firms may eventually need more structured information channels that agents can use for jurisdiction-specific checks. These would have to be carefully governed and updated; no single commercial protocol can resolve public law by itself.
There is also a geopolitical dimension. Countries with strong domestic payment systems, data-localization requirements, or national AI strategies may favor different implementations, and China’s platforms have been building agent-payment services on their own rails in parallel with the Western protocol contest. The global AI commerce system could fragment into regional combinations of identity, wallets, merchant data standards, and model providers. This would raise compliance and multihoming costs. Alternatively, open technical standards with jurisdiction-aware extensions could preserve cross-border reach. Corporate strategy should therefore include localization at the policy and data levels, not merely translation of the user interface.
The relevant public policy is not to promote one national assistant at any cost. It is to ensure that domestic merchants and consumers can participate safely in evolving commerce networks while maintaining competitive choice, privacy, and lawful trade. Such policies may involve small-business digital infrastructure, cyber-resilience, standards participation, payment-fraud prevention, and enforcement capacity. They need not be framed as a subsidy race for the most computationally intensive model.
5.6. A 2027–2030 scorecard for companies and policymakers
By 2027, leading merchants should be able to answer basic questions that many cannot yet answer confidently: Which agents can discover our products? What portion of our catalog is represented correctly? Which qualified tasks include us? Which transactions are incremental? What are the true fees and service costs? Who owns permission to contact the customer after the sale? How do we contest inaccurate information? Without those answers, an agent strategy is likely to become another advertising budget without a measurement discipline.
By 2028, agent providers should be expected to mature in authorization evidence, merchant eligibility documentation, post-purchase handoffs, and independently testable recommendation quality. Payment networks should measure authentication success and fraud, but also false rejections of legitimate agents. Standards organizations should publish compatibility and extension guidance. Consumer advocates should investigate whether people understand the difference between a personalized recommendation, a commercial placement, and an authorized purchasing action. The chronology is an agenda, not a claim that any firm has committed to these dates.
By 2029, researchers could evaluate longitudinal data on merchant concentration, small-business access, product diversity, repeat-purchase dynamics, and consumer welfare. The field needs to know whether AI expands the effective choice set or merely makes an already concentrated one easier to transact with. It needs causal evidence of price savings, time savings, returns, product fit, and fraud. It should compare categories and countries rather than present a single global adoption number. A thirty-eightfold wallet growth claim is newsworthy, but a durable political economy requires denominators, baselines, and outcomes.[1]
By 2030, the decisive test will be whether agentic commerce is simultaneously useful, secure, and contestable. The market could be technically interoperable yet economically concentrated. It could also be fragmented into incompatible private networks that burden small merchants. Neither future should be accepted as the default. Public officials can promote transparency, competition, and enforcement without dictating every model architecture; firms can invest in portable infrastructure and independent evaluation without surrendering legitimate intellectual property.
Table 6. A scorecard by actor, 2027–2030
| Actor | What to track | Warning signs |
| Retailers and brands | Qualified coverage, incremental contribution margin per agent order, direct repeat rate after agent acquisition, catalog accuracy, dispute resolution time | Rising agent GMV with flat or falling contribution; buyers returning only through the intermediary |
| AI platforms and agents | Constraint fidelity, merchant diversity among qualified alternatives, consent integrity, sponsored-treatment transparency, post-purchase satisfaction | Conversion rising while constraint fidelity falls; sponsored share of recommendations rising without disclosure |
| Payment networks and wallets | Agent verification rate, unauthorized-transaction incidence, dispute resolution outcomes, accessibility of multiple wallets per agent | False rejections of legitimate agents; single-wallet defaults inside dominant agents |
| Startups | Multihoming resilience, differentiation, measurable cost reduction for merchants, dependency on any single gatekeeper | Revenue concentrated in one platform’s revocable program |
| Regulators | Barriers to entry, consumer harm, discriminatory access, effectiveness of remedies, terms-of-service exclusion of third-party agents | Open protocol adoption coinciding with rising attention concentration |
This section’s structural conclusion is that the governance of Agentic Shelfspace is not a future problem to be considered after adoption. Early technical defaults, contracts, consent designs, attribution rules, and ranking incentives can become hard to reverse once embedded in large consumer and merchant networks. The choices made between 2027 and 2030 may determine whether the new storefront is a competitive, public-facing marketplace of opportunities or a set of privately controlled recommendation corridors.
Section 6 — What Have We Learned? Seven Pillars of Agentic Shelfspace
Pillar 1 — Merchant Legibility: The Right to Be Considered Begins with Reliable Data
The first pillar is that commerce cannot be meaningfully competitive in an agentic environment if legitimate products are not machine-legible. The old shelf demanded physical distribution; the search era demanded indexing and relevance; the agentic shelf demands trusted, structured, current, and comparable representations of offers. A merchant’s eligibility for a consumer task depends partly on information systems it controls, and partly on the catalog gateways and retrieval systems it does not. Shopify’s evidence that structured-catalog searches convert at twice the rate of scraped-data searches is the commercial expression of this pillar, and Adobe’s finding that large portions of retail sites remain unreadable to machines is its warning.[16][27]
This insight should change investment priorities. A retailer can spend heavily on advertising yet remain invisible for a task if the agent cannot verify dimensions, compatibility, availability, or shipping. The correct response is not an arms race in synthetic marketing copy. It is a durable product-evidence capability: normalized variants, authoritative specifications, provenance, timestamped availability, meaningful warranty terms, and merchant-identification controls. Merchant platforms and startups can supply much of this infrastructure for smaller businesses.
But legibility must not become a pretext for exclusion. Standards should make essential functions accessible without unnecessary certification or proprietary extensions. Platforms should distinguish an offer that fails a genuine quality requirement from one that is absent because onboarding or commercial access is restricted. Independent audits should examine the qualified coverage rate across large and small merchants, product categories, and regions.
For corporate leaders, the question is, Can an authorized agent accurately discover and understand what we sell? For policymakers, it is, Can a legitimate merchant enter the effective choice set without purchasing preferential treatment? The first is an operational challenge; the second is a competition challenge. Agentic Shelfspace requires both to be answered.
Pillar 2 — Decision Duty: A Shopping Assistant Must Be Evaluated by Whom It Serves
The second pillar concerns the principal of the decision. An assistant may be deployed by a company that sells ads, collects transaction fees, owns a marketplace, or operates a preferred payment network. Those business models can finance useful services, but they create circumstances in which the platform’s interests and the consumer’s interests diverge. The existence of a sponsored label or an open protocol does not by itself prove that recommendations are unbiased. The relevant test is whether the system faithfully represents the user’s objectives and discloses material commercial influence.
The economic standard should be outcome-based. Recommendation quality is not the same as conversion. A trustworthy assistant asks clarifying questions when necessary, verifies important requirements, distinguishes source evidence from inference, and respects a buyer’s preferences even when a different product would earn the platform more revenue. It should be possible to compare its decisions against a prevalidated set of qualified alternatives.
Companies should publish meaningful principles about sponsored treatment, affinity arrangements, and how they isolate commercial incentives from organic advice. Independent evaluations can test those commitments. The 2026 experimental record — conversational persuasion nearly tripling sponsored selection, most frontier models choosing a pricier sponsored option under a soft cue, and role assignment changing an agent’s skepticism toward sponsorship — provides every reason to take such testing seriously, even though none of it establishes misconduct by specific firms.[38][39][40] The goal is not an impossible demand that every recommendation satisfy every consumer. It is a system that can be scrutinized for systematic distortions and corrected when they are found.
This pillar converts the title’s metaphor into an ethical and competitive obligation. If the shelf is assembled by an agent, the industry must ask on whose behalf it is assembled. A merchant may reasonably pay to advertise; a consumer must still be able to tell the difference between an advertisement and advice.
Pillar 3 — Consent Architecture: Permission Must Travel with the Action
The third pillar is the separation of capability from authority. A model can navigate checkout, invoke APIs, or prepare payment instructions. That technical ability does not prove that an individual or business authorized the action. The human must remain able to define scope: the product category, amount, seller, timing, recurrence, and circumstances requiring confirmation. Evidence of that authorization must survive handoffs between the assistant, the commerce protocol, the wallet, the network, and the merchant.
This is where otherwise separate initiatives converge. Visa’s authenticated-agent approach, Mastercard’s verifiable intent, Google’s payment mandates, and Stripe’s authorization services reflect a shared problem: useful automation needs proofs that can be interpreted across systems.[1][49][50][52] The standards differ, and the industry has not established one final architecture. The principle is stable, however. A trusted transaction should be attributable to a recognized actor acting within a specific, valid permission, with records adequate for dispute resolution.
Consumer experience should make authorization understandable. “Research options,” “add to cart,” “reserve,” “buy once,” and “repeat under a limit” are materially different permissions. A final total that exceeds a threshold must trigger the appropriate reauthorization. Revocation and error correction should be straightforward. B2B workflows may place approval in a procurement system before the agent makes the payment, as the October 2026 Visa–BILL–Billtrust pilot illustrates.[4]
For executives, consent architecture is a condition of scalable adoption, not a compliance afterthought. For payment networks, it is a chance to provide enduring trust infrastructure. For regulators, it is the boundary between legitimate automation and unauthorized economic action. In the new storefront, the right to recommend and the right to spend must never be mistaken for the same right.
Pillar 4 — Protocol Plurality: Open Connections Must Preserve Real Commercial Choice
The fourth pillar is that no one protocol should be mistaken for the entire market. ACP, UCP, payment mandates, agent-verification systems, browser-facing integrations, and proprietary marketplace tools address different parts of a transaction. Merchants should be able to reach multiple suitable assistants without recreating their businesses each time. Consumers should be able to choose agents and payment methods where feasible. Technical compatibility is valuable because it reduces switching and integration costs; it is insufficient when distribution or ranking becomes concentrated.
The policy concern is the interoperability paradox: an open standard can make a dominant interface more attractive by connecting it to more merchants. The response is not to oppose interoperability. It is to pair openness with contestability: transparent eligibility criteria, reasonable access conditions, avoidance of coercive tying, meaningful steering options, and portable records where technically and legally appropriate. Competition authorities should examine actual exclusionary effects rather than infer fairness from the word “open” in a specification.
The corporate implication is to build modularly. Product data, inventory, checkout, payment, identity, and service functions should be governed as separable capabilities. A company should understand its dependencies on each platform and negotiate accordingly. A merchant that cannot change its wallet or agent channel without losing its customers is economically less independent than one that can. A startup whose entire business rests on a single revocable integration should recognize the risk in its valuation.
By 2030, the quality of agentic commerce will depend not only on model intelligence but on whether competition remains possible among the institutions using that intelligence. Protocol plurality is not fragmentation for its own sake. It is the preservation of alternative paths for consumers and merchants when one intermediary becomes inconvenient, expensive, or untrustworthy.
Pillar 5 — Access Parity: A Marketplace’s Right to Exclude Must Be Weighed Against a Consumer’s Right to Delegate
The fifth pillar emerged from the litigation and blocking episodes of 2026 rather than from any protocol. When Amazon sued Perplexity, won and then lost an injunction, and subsequently turned to its Conditions of Use to exclude Meta’s Muse while operating its own agents that shop elsewhere, it posed a question the market had not previously needed to answer: when a consumer directs a trusted agent to act on an account the consumer owns, whose permission governs?[47][48] The Ninth Circuit’s answer under anti-hacking law was that the consumer is the one accessing the site; its silence on contract law left marketplaces free to write agent terms, and Amazon’s terms already required agents to identify themselves and to stop when asked.
Both sides of this dispute have legitimate interests. A marketplace that invests in fraud prevention, customer service, and account security has reason to fear undisclosed automation scraping account data or placing erroneous orders, and the right to demand that agents identify themselves is a reasonable condition of access. A consumer who has chosen an agent has reason to expect that the agent can act where the consumer could act, and a dominant marketplace that blocks all external agents while running its own is engaged in conduct that competition authorities will recognize from earlier platform disputes. The resolution that best serves the market is access parity under identification: agents that disclose themselves, honor rate and scope limits, and accept liability for their errors should be able to transact on a consumer’s behalf, and a platform’s decision to exclude them should be justified by demonstrable security or integrity concerns rather than by the protection of advertising revenue or internal agents.
The practical instrument for this pillar is already visible in the payment networks’ agent directories and trusted-agent registries. A registered, authenticated, and scored agent is a different proposition from an anonymous browser masquerading as a human, and the Visa–BILL–Billtrust pilot showed a portal operator accepting precisely such an agent in place of its bot defenses.[4] If marketplaces can be persuaded, or required, to apply the same logic to consumer agents that meet an industry standard of identification and accountability, the fortress scenario loses much of its force without any platform being compelled to accept genuinely abusive automation.
For executives at marketplaces, the question is whether blocking will remain defensible once registered-agent standards mature. For agent builders, it is whether they will accept the identification obligations that make access defensible. For regulators, it is whether the refusal to deal with identified agents, by a platform that fields its own, is a practice that merits scrutiny. The right to exclude and the right to delegate will have to be reconciled, and 2026 showed that courts, contracts, and protocols will each have a role in the reconciliation.
Pillar 6 — Auditability: Claims About Neutrality Must Be Testable by Someone Other Than the Claimant
The sixth pillar follows from the second and makes it operational. Every major agent platform in 2026 made some public commitment about the independence of its recommendations from its commercial interests — OpenAI’s statement that ads do not influence answers, Google’s and Shopify’s emphasis on merchants remaining seller of record, Meta’s assurance that Muse’s data does not feed its ad systems.[2][5][37] These commitments are valuable precisely to the degree that they can be checked, and the research community has already demonstrated the method: randomized, provider-agnostic audits that vary position, sponsorship, price, and endorsement and measure the causal response of agent choices.[20]
Auditability does not require platforms to publish their models or ranking weights. It requires three more modest things. First, stable interfaces through which independent evaluators can submit standardized tasks and observe the resulting consideration sets, under terms that protect privacy and prevent abuse. Second, documentation of material commercial arrangements — paid inclusion, affiliate relationships, preferred-partner programs — sufficient to let an auditor distinguish a commercial effect from a quality effect. Third, retention of decision records adequate to reconstruct, in a dispute, which candidates were considered and why the chosen one prevailed. The Agentic Shelfspace Audit proposed in Section 3 is designed around these three requirements.
The analogy to financial audit is instructive. Firms are not required to let the public inspect their ledgers, but they are required to let an independent professional do so and to publish the result. Agent platforms that aspire to custody over consumer intent, and to a decisive role in allocating commercial access, should expect the same discipline — and the platforms that embrace it first will find it a competitive asset rather than a burden, because trust in a recommendation that has been tested is worth more than trust in one that has merely been asserted.
For executives, auditability is a design requirement to be built in now, before the records that make it possible are discarded. For regulators, it is the most proportionate alternative to algorithmic mandates. For consumers, it is the difference between a shelf they can inspect and one they must simply believe in.
Pillar 7 — Five-Layer Value Discipline: The Real Denominator Is a Satisfactory, Authorized Outcome
The seventh pillar reconnects the commercial story to the Five-Layer AI Economy. A shopping assistant is the visible endpoint of a deeper industrial system: applications and agents, models, datacenters, chips, and energy. Its commercial success could stimulate recurring inference demand, payment services, merchant software, and infrastructure investment, and the capital-expenditure plans disclosed in 2026 suggest that the largest participants are already counting on it.[28] But the size of the industry should not be measured solely by the number of agent messages, agent runtime, protocol connections, or announced transaction capabilities.
The economically meaningful denominator is a satisfactory, authorized outcome. At Layer Five, did the assistant help the user obtain the right product with appropriate permission and a workable remedy? At Layer Four, did the model reason and calibrate uncertainty reliably? At Layer Three, was the system available, secure, and efficient? At Layer Two, did the computing architecture offer sustainable performance per useful task? At Layer One, did energy and grid resources support that activity at a defensible cost and reliability level? These questions translate an abstract AI boom into an accountable chain of value creation.
Financial analysts should distinguish gross merchandise value from incremental contribution margin, and company-reported transaction growth from market penetration. Executives should track failed tool calls, unnecessary background polling, dispute rates, returns, and the marginal compute needed for a completed task. Policymakers should examine whether infrastructure investment increases consumer welfare and business productivity, not merely platform engagement. The goal is not to suppress consumption of compute; it is to understand what that compute produces.
This final pillar reveals why Agentic Shelfspace belongs in the Five-Layer AI Economy rather than solely in a marketing or retail-technology discussion. The new commercial interface can create a feedback loop from user intent and transaction revenue to models and physical infrastructure. The quality of that loop depends on institutions — trust, fair access, data integrity, authorization, auditability, and competitive incentives — as much as on model capability. Sustainable growth requires technical and economic disciplines to reinforce each other.
Conclusion — The New Shelf Is a Decision, Not a Display
Retail history has repeatedly changed where consumers encounter goods. The street market gave way to the department store; the department store to the shopping mall and the big-box retailer; the web browser to search, marketplaces, mobile apps, and social feeds. Each transition altered which institutions controlled discovery, what sellers had to do to be visible, and how commercial value was divided. Artificial intelligence is producing another transition, but its most important feature is not a new screen. It is the possibility that software interprets a buyer’s goal, chooses a set of offers, negotiates the practical steps of purchase, and sometimes completes authorized actions. The storefront can be present even when the consumer never opens the seller’s website.
The events of 2026 make that change concrete, and they also counsel restraint. OpenAI and Stripe’s ACP demonstrated an architecture for AI-mediated commerce; OpenAI’s later emphasis on discovery showed that flexible checkout and merchant relationships remain difficult to standardize. Google and Shopify’s UCP built another route for commerce interoperability, and Google’s October Universal Cart launch organized cross-merchant shopping around persistent context and user control. Stripe’s thirty-eightfold Link growth suggested momentum within its own network, while Visa’s and Mastercard’s work illustrated how identity, consent, and payment verification must mature before delegation can scale. Amazon’s litigation and blocking showed that a marketplace will defend its shelf with every instrument available, and the Ninth Circuit showed that the law is only beginning to decide which instruments those are. The second-quarter earnings of Shopify, Amazon, Visa, Mastercard, and PayPal confirmed that every one of these firms regards agentic commerce as strategically central and, in the same breath, that none of them yet reports it as a material line of revenue.[1][2][3][4][26][28][29][31][33][47]
The central commercial question is now sharper than it was at the beginning of the paper: Will merchants primarily compete for the direct attention of human buyers, or for inclusion and preference within the decision processes of the agents acting for those buyers? The realistic answer is both. Some purchases will remain tactile, emotional, social, or deeply human. Others will be increasingly assisted, filtered, and executed by software. The mix will vary by category, risk, geography, trust, and individual preference. Yet even partial migration can alter the economics of customer acquisition, advertising, loyalty, payments, and product differentiation, and Shopify’s observation that three-quarters of AI-attributed purchases fall outside the top one hundred categories hints that the migration may begin in exactly the long tail where human search worked least well.[27]
This is precisely why I chose Agentic Shelfspace as the title. It is memorable because it joins an established commercial scarcity — access to the shelf — with a novel allocator, an agent capable of interpreting and acting. It is analytically useful because it identifies what should be measured: coverage, qualified recommendation share, product fit, commercial influence, authorization, merchant access, and the true cost of conversion. It is politically relevant because the distribution of shelfspace can become an issue of competition, consumer protection, and economic opportunity. And it is durable because the underlying struggle for visibility and authority will persist even as today’s specific protocols, model names, and interfaces evolve.
The Five-Layer AI Economy reveals the full significance of this shift. Intelligence does not simply generate more persuasive shopping text. It can become a revenue-producing coordination layer that mobilizes models, datacenters, chips, and energy to perform real economic work. The question for 2027–2030 is whether that work will be organized to expand informed choice, trustworthy automation, independent merchant participation, and measurable productivity — or to concentrate commercial influence in a few opaque decision systems. Protocols and payment tokens are necessary engineering ingredients. They are not sufficient governance.
The next retail frontier may therefore be measured less by the number of stores a consumer visits than by the number of genuinely qualified alternatives an assistant considers, the fidelity with which it serves a buyer’s goals, and the integrity of the permission under which it acts. A brand could lose a sale without ever receiving a website visit. A small merchant could gain a customer who would never have found it through conventional search. A payment network could become the trusted bridge between an instruction and a legally meaningful transaction. A technology platform could become a powerful commercial gatekeeper without ever taking title to the goods.
The central task for business leaders and public officials is to ensure that this new shelf does not become an invisible monopoly over consumer intent. Its legitimacy must come from outcomes that can be tested, choices that can be contested, permissions that can be verified, access that can be defended on its merits, and commercial relationships that remain accountable. Agentic Shelfspace names the emerging marketplace of machine-mediated consideration. Its future will be determined not only by what AI agents can buy, but by whose interests they faithfully represent when they decide what is worth buying.
Endnotes and Source References:
[1] Dan Hill, Stripe, “Helping personal agents shop more intelligently and reliably with Link,” September 29, 2026. https://stripe.com/blog/helping-personal-agents-shop-more-intelligently-and-reliably-with-link
[2] Suresh Ganapathy, Google, “Google’s new Universal Cart launches today in the U.S.,” The Keyword, October 15, 2026. https://blog.google/products-and-platforms/products/shopping/universal-cart-launch/
[3] OpenAI, “Powering Product Discovery in ChatGPT,” March 24, 2026. https://openai.com/index/powering-product-discovery-in-chatgpt/
[4] Visa, BILL, and Billtrust, “Visa, BILL and Billtrust Launch First Live Pilot for Trusted AI Agent Payments Between Buyer and Supplier Platforms,” press release, October 20, 2026. https://investor.visa.com/news/news-details/2026/Visa-BILL-and-Billtrust-Launch-First-Live-Pilot-for-Trusted-AI-Agent-Payments-Between-Buyer-and-Supplier-Platforms/default.aspx
[5] Meta, “Introducing Muse: The World’s First Personal AI Agent Built for Everyone,” September 8, 2026. https://about.fb.com/news/2026/09/introducing-muse-personal-ai-agent/
[6] Amazon.com, Inc., “Amazon.com Announces Second Quarter Results,” earnings release, July 30, 2026. https://www.aboutamazon.com/news/company-news/amazon-earnings-q2-2026-report
[7] OpenAI, “Buy it in ChatGPT: Instant Checkout and the Agentic Commerce Protocol,” September 29, 2025. https://openai.com/index/buy-it-in-chatgpt/
[8] Stripe, “Developing an open standard for agentic commerce,” September 29, 2025. https://stripe.com/blog/developing-an-open-standard-for-agentic-commerce
[9] Google, “New tech and tools for retailers to succeed in an agentic shopping era,” January 11, 2026. https://blog.google/products/ads-commerce/agentic-commerce-ai-tools-protocol-retailers-platforms/
[10] Google Developers, “Under the Hood: Universal Commerce Protocol (UCP),” January 11, 2026. https://developers.googleblog.com/under-the-hood-universal-commerce-protocol-ucp/
[11] Shopify, “Shopify connects any merchant to every AI conversation,” 2026. https://www.shopify.com/news/ai-commerce-at-scale
[12] Mastercard, “Mastercard advances agentic commerce with new trust and intelligence services,” press release, September 30, 2026. https://www.mastercard.com/us/en/news-and-trends/press/2026/september/new-trust-and-intelligence-services.html
[13] PayPal Developer, “WebMCP is here: PayPal is helping every merchant become agent-ready,” September 16, 2026. https://developer.paypal.com/community/blog/WebMCP_PayPal_Agent_Ready/
[14] OpenAI Developers, “Get Started — Agentic Commerce,” technical guidance, accessed October 2026. https://developers.openai.com/commerce/guides/get-started
[15] Shopify, “Agentic-Ready Product Data: How to Get It & the Cost of Inaction,” 2026. https://www.shopify.com/enterprise/blog/agentic-ready-product-data
[16] Adobe Digital Insights, “How AI is reshaping travel planning,” including July 2026 U.S. retail AI-referral analysis, August 2026. https://business.adobe.com/blog/us-consumers-are-embracing-llms-to-make-travel-plans
[17] Abbas Haleem, Digital Commerce 360, “Adobe: AI-referral traffic spending, converting more than counterparts,” August 19, 2026. https://www.digitalcommerce360.com/2026/08/19/adobe-ai-referral-traffic-data-july-2026/
[18] Sunghwan Kim, Ryang Heo, Yongsik Seo, Jinyoung Yeo, and Dongha Lee, “AgenticShop: Benchmarking Agentic Product Curation for Personalized Web Shopping,” arXiv preprint, February 2026. https://arxiv.org/abs/2602.12315
[19] “EComAgentBench: Benchmarking Shopping Agents on Long-Horizon Tasks with Distributed Hidden Intent,” arXiv preprint, June 2026. https://arxiv.org/abs/2606.17698
[20] Amine Allouah, Omar Besbes, Josué D. Figueroa, Yash Kanoria, and Akshit Kumar (Columbia Business School), “What Is Your AI Agent Buying? Evaluation, Biases, Model Dependence, & Emerging Implications for Agentic E-Commerce,” arXiv 2508.02630; Proceedings of the ACM Web Conference 2026. https://arxiv.org/abs/2508.02630
[21] Peyman Shahidi, Gili Rusak, Benjamin S. Manning, Andrey Fradkin, and John J. Horton, “The Coasean Singularity? Demand, Supply, and Market Design with AI Agents,” NBER Working Paper 34468, 2025; forthcoming in The Economics of Transformative AI (University of Chicago Press). https://www.nber.org/papers/w34468
[22] MIT Initiative on the Digital Economy, “How AI Agents for Commerce Could Shape Future Markets,” 2026. https://ide.mit.edu/insights/how-ai-agents-for-commerce-could-shape-future-markets/
[23] MIT Initiative on the Digital Economy, “AI Agents Want to Shop for You: The Future of Agentic Commerce” (remarks of Bob Hedges, MIT IDE Digital Fellow), 2026. https://ide.mit.edu/insights/ai-agents-want-to-shop-for-you-the-future-of-agentic-commerce/
[24] Mark Bartholomew, University at Buffalo School of Law, “What we lose when AI does our shopping,” UBNow / The Conversation, May 1, 2026. https://www.buffalo.edu/ubnow/stories/2026/05/bartholomew-conversation-ai-shopping.html
[25] CNBC, “Meta’s Muse agent is attacking one of the economy’s most profitable weak spots” (remarks of Neale Mahoney, Stanford Institute for Economic Policy Research), September 27, 2026. https://www.cnbc.com/2026/09/27/meta-muse-ai-personal-agent.html
[26] Shopify Inc., “Shopify Delivers Big: 30%+ Growth Across GMV, Revenue, Gross Profit, and Free Cash Flow,” second-quarter 2026 results, Form 8-K Exhibit 99.1, August 5, 2026. https://www.sec.gov/Archives/edgar/data/0001594805/000159480526000046/exhibit991pressreleaseq220.htm
[27] Shopify Inc., second-quarter 2026 earnings call transcript (Harley Finkelstein, Jeff Hoffmeister), August 5, 2026. https://www.theglobeandmail.com/investing/markets/markets-news/motley/3809161/shopify-shop-q2-2026-earnings-call-transcript/
[28] Amazon.com, Inc., second-quarter 2026 earnings call transcript (Andy Jassy, Brian Olsavsky), July 30, 2026. https://www.fool.com/earnings/call-transcripts/2026/08/07/amazon-amzn-q2-2026-earnings-call-transcript/
[29] Visa Inc., “Visa Reports Fiscal Third Quarter 2026 Results,” Form 8-K, July 28, 2026. https://www.sec.gov/Archives/edgar/data/0001403161/000140316126000103/q32026earningsrelease.htm
[30] Visa Inc., fiscal third-quarter 2026 earnings call transcript (Ryan McInerney), July 28, 2026, as reported by Investing.com. https://www.investing.com/news/transcripts/earnings-call-transcript-visa-tops-q3-2026-estimates-raises-outlook-93CH-4818303
[31] Sachin Mehra, Mastercard, “Earnings review: Delivering value across every stage of commerce,” second-quarter 2026, July 30, 2026. https://www.mastercard.com/us/en/news-and-trends/stories/2026/earnings-review-q2-2026.html
[32] PYMNTS, “Cards Make Their Case in an Agentic World” (remarks of Michael Miebach, Mastercard second-quarter 2026 earnings call), August 2026. https://www.pymnts.com/credit-cards/2026/cards-make-their-case-in-an-agentic-world/
[33] PayPal Holdings, Inc., second-quarter 2026 earnings call transcript (Enrique Lores), July 28, 2026. https://s205.q4cdn.com/875401827/files/doc_financials/2026/q2/PYPL-2Q-26-Earnings-Transcript.pdf
[34] ICSC and McKinsey & Company, “Shopping in the Age of AI: Redefining Stores for a New Era,” April 27, 2026. https://www.icsc.com/news-and-views/icsc-exchange/icsc-mckinsey-company-release-shopping-in-the-age-of-ai-redefining-stores-for-a-new-era-report
[35] McKinsey & Company (QuantumBlack), “The automation curve in agentic commerce,” January 2026. https://www.mckinsey.com/~/media/mckinsey/business%20functions/quantumblack/our%20insights/the%20automation%20curve%20in%20agentic%20commerce/the-automation-curve-in-agentic-commerce.pdf
[36] Salesforce, “AI and agents account for $262 billion of 2025 holiday spend” (remarks of Caila Schwartz), January 2026. https://www.salesforce.com/news/stories/2025-holiday-shopping-data/
[37] OpenAI, “Building advertising for the way people use AI,” October 5, 2026. https://openai.com/index/new-chatgpt-ads-format-and-measurement/
[38] Davood Wadi and Yu Ma, “Whom Do AI Agents Work For? Role Assignment Induces Sponsorship Bias in LLM Recommenders,” arXiv preprint, September 2026. https://arxiv.org/abs/2609.17989
[39] “Commercial Persuasion in AI-Mediated Conversations,” arXiv 2604.04263, April 2026 (preregistered experiment, N = 2,012). https://arxiv.org/abs/2604.04263
[40] Addison J. Wu, Ryan Liu, Shuyue Stella Li, Yulia Tsvetkov, and Thomas L. Griffiths (Princeton University and University of Washington), “Ads in AI Chatbots? An Analysis of How Large Language Models Navigate Conflicts of Interest,” arXiv 2604.08525, 2026. https://arxiv.org/abs/2604.08525
[41] Federal Trade Commission, “FTC Surveillance Pricing Study Indicates Wide Range of Personal Data Used to Set Individualized Consumer Prices,” January 2025. https://www.ftc.gov/news-events/news/press-releases/2025/01/ftc-surveillance-pricing-study-indicates-wide-range-personal-data-used-set-individualized-consumer
[42] California Privacy Protection Agency, “CCPA Updates, Cybersecurity Audits, Risk Assessments, Automated Decisionmaking Technology (ADMT), and Insurance Regulations,” regulations effective January 1, 2026. https://cppa.ca.gov/regulations/ccpa_updates
[43] Federal Trade Commission, “FTC Seeks Public Comment on Policy Statement Addressing AI Accuracy,” July 1, 2026 (proposed policy statement; comments through July 31, 2026). https://www.ftc.gov/news-events/news/press-releases/2026/07/ftc-seeks-public-comment-policy-statement-addressing-ai-accuracy
[44] Federal Trade Commission, “Endorsements, Influencers, and Reviews,” business guidance. https://www.ftc.gov/business-guidance/advertising-marketing/endorsements-influencers-reviews
[45] U.S. Department of Justice, “Department of Justice Wins Significant Remedies Against Google,” September 2, 2025. https://www.justice.gov/opa/pr/department-justice-wins-significant-remedies-against-google
[46] European Commission, “Commission fines Google €890 million for breaches of the Digital Markets Act” (remarks of Executive Vice-President Henna Virkkunen), July 23, 2026. https://digital-strategy.ec.europa.eu/en/news/commission-fines-google-eu890-million-breaches-digital-markets-act
[47] Wilson Sonsini Goodrich & Rosati, “Ninth Circuit Addresses CFAA and Agentic AI Tools in Groundbreaking Decision” (Amazon.com Services, LLC v. Perplexity AI, Inc., 9th Cir., August 4, 2026), August 2026. https://www.wsgrdataadvisor.com/2026/08/ninth-circuit-addresses-cfaa-and-agentic-ai-tools-in-groundbreaking-decision/
[48] Todd Bishop, GeekWire, “Amazon blocks Meta’s Muse AI assistant in new standoff over agentic shopping,” September 21, 2026. https://www.geekwire.com/2026/amazon-blocks-metas-muse-ai-assistant-in-new-standoff-over-agentic-shopping/
[49] Visa Developer, “Trusted Agent Protocol,” technical documentation. https://developer.visa.com/capabilities/trusted-agent-protocol
[50] Mastercard, “How Verifiable Intent builds trust in agentic AI commerce,” March 5, 2026. https://www.mastercard.com/us/en/news-and-trends/stories/2026/verifiable-intent.html
[51] Mastercard, “Mastercard launches Agent Pay for Machines to unlock super-fast, always-on payments,” June 10, 2026. https://www.mastercard.com/us/en/news-and-trends/press/2026/june/mastercard-launches-agent-pay-for-machines.html
[52] Google Developers, “Developer’s Guide to AI Agent Protocols,” including Agent Payments Protocol (AP2) mandates, March 18, 2026. https://developers.googleblog.com/developers-guide-to-ai-agent-protocols/
[53] Stripe, “Introducing the Agentic Commerce Suite: A complete solution to make your business agent-ready,” December 11, 2025. https://stripe.com/blog/agentic-commerce-suite
[54] Meta, “The Biggest News From Connect 2026,” September 2026. https://about.fb.com/news/2026/09/the-biggest-news-from-connect-2026/
[55] Jeremy Kahn, Fortune, “AI shopping agents are coming. No one is ready for them” (Fortune Brainstorm Tech 2026 panel), June 12, 2026. https://fortune.com/2026/06/12/ai-shopping-agents-are-coming-no-one-is-ready-for-them/
[56] Dealroom, “Stripe’s virtual cards for AI shopping agents put Visa and Mastercard on edge,” summarizing The Information’s October 26, 2026 reporting, October 27, 2026. https://app.dealroom.co/news/note/stripe-s-virtual-cards-for-ai-shopping-agents-put-visa-and-mastercard-on-edge
[57] Digital Commerce 360, “Google launches Universal Cart for purchases across retailers,” October 16, 2026. https://www.digitalcommerce360.com/2026/10/16/google-universal-cart-launch-features/
[58] CNBC, “OpenAI revamps shopping experience in ChatGPT after struggling with Instant Checkout offering,” March 24, 2026. https://www.cnbc.com/2026/03/24/openai-revamps-shopping-experience-in-chatgpt-after-instant-checkout.html
[59] Retail Dive, “Amazon customers are embracing Alexa for Shopping,” August 2026. https://www.retaildive.com/news/amazon-alexa-for-shopping-agentic-ai-adoption/826806/
[60] PayPal Developer, “Agent Ready overview,” updated June 10, 2026. https://developer.paypal.com/agent-ready/overview/
[61] OpenAI Developers, “Agentic Checkout Spec — Agentic Commerce,” technical specification. https://developers.openai.com/commerce/specs/checkout[62] Modern Retail, “What went wrong with ChatGPT’s Instant Checkout,” March 2026. https://www.modernretail.co/technology/what-went-wrong-with-chatgpts-instant-checkout/
