The Macro Shift
Agentic Commerce in 2026: The Rails Are Built, the Trust Is Not
Every major payment network and AI platform has now shipped a way for agents to buy on a shopper's behalf, but the verified data shows trust collapses at the exact moment an agent reaches for a card.
Part of The Macro Shift in the Insights library.
Abstract
In under two years, OpenAI, Google, Visa, Mastercard, Coinbase, and Shopify have each shipped a competing protocol for AI agents to discover, compare, and pay for products, and the first live checkouts are already running inside ChatGPT, Perplexity, and Google's AI surfaces. The consumer research behind the announcements tells a more cautious story: shoppers already lean on AI heavily to research a purchase and say they trust an agent in the abstract, but only a single-digit share will hand over full purchasing authority, and every protocol is architected around signed human consent rather than blanket autonomy. How much commerce is actually moving through this layer today remains a question only the platforms themselves are currently answering, which is exactly the gap an independent measure of attention is built to close.
The shift happened faster than most retailers noticed. Anthropic introduced the Model Context Protocol in November 2024 as a plumbing standard for connecting AI models to outside tools, and within a year it had quietly become the shared substrate under nearly every competing commerce protocol that followed it. Mastercard and Visa shipped rival agent-payment protocols a single day apart in April 2025. Coinbase launched x402, reviving the dormant HTTP 402 'Payment Required' status code to settle machine-to-machine payments in USDC, the following month. Google published the Agent Payments Protocol in September 2025 with more than sixty launch partners including Mastercard, PayPal, Coinbase, American Express, and Salesforce, built around cryptographically signed Intent and Cart Mandates. OpenAI and Stripe co-developed the Agentic Commerce Protocol and took it live as Instant Checkout inside ChatGPT for US users on February 16, 2026, with Etsy live at launch and more than a million Shopify merchants, including Glossier, Vuori, Spanx, and SKIMS, rolling out behind it. Shopify and Google followed with the Universal Commerce Protocol, and Google's Universal Cart went live in early July 2026. Visa formalized its own version as the Trusted Agent Protocol with Cloudflare, using RFC 9421 HTTP signatures, and named Anthropic, OpenAI, Perplexity, Mistral, and Samsung as model partners. Perplexity took a narrower route, pairing a subscriber-only Buy with Pro feature backed by PayPal, covering more than five thousand merchants, with a separate, fee-free Perplexity Merchant program. Card networks are also treating stablecoins as core infrastructure rather than a side bet: Visa's stablecoin activity reached a seven billion dollar annualized run rate by April 2026, and Mastercard agreed in March 2026 to acquire the stablecoin platform BVNK for up to 1.8 billion dollars. None of this is confined to the United States. Google's AP2 and Visa's model-partner list span multiple regions by design, and the consumer research cited below was fielded across the US, Brazil, the UAE, and, in Accenture's case, sixteen countries, which suggests the protocol race and the trust question are both global, not a US-only phenomenon.
The verified survey data draws a sharper line than any protocol announcement does. PYMNTS Intelligence, working with Visa Acceptance Solutions across 5,241 consumers, 1,185 merchants, and 150 acquirers in the US, Brazil, and the UAE, found that 48% of online shoppers had already used AI to research their most recent purchase, which tells you discovery and comparison are already an AI-mediated behavior for close to half of shoppers. But the same study found 27% trust no organization to operate an AI shopping agent at all, and 24% say they will never delegate a purchase to AI under any circumstance. Accenture's larger sixteen-country Consumer Pulse survey, fielded across 25,590 consumers, found something that looks contradictory at first: 74% said they would trust a personal AI agent more than their own best friend to make a purchase for them. Read next to the rest of Accenture's own numbers, the contradiction resolves. Only 9% of those same respondents are open to fully autonomous purchasing. Thirty-two percent would let an agent choose within a set budget and brand parameters, and 56% insist on specifying up front which brands the agent may even consider. Put the two studies together and a consistent shape emerges: shoppers trust an agent to listen, research, and narrow choices, and they trust it in the abstract more than they expected to, but they want a hand on the wheel the moment money moves. That is not a marketing preference, it is why every protocol described above, AP2's Cart Mandates, ACP's Shared Payment Token, Visa's Trusted Agent Protocol, is built around a signed, revocable consent artifact rather than a standing authorization. The industry built its payment rails around the exact trust boundary the survey data describes, a rare case of infrastructure and behavior lining up before either side had proof the other existed.
For a business owner, the practical question is not which protocol wins, it is where buyer attention actually sits right now, and the answer is that nobody outside the platforms themselves can say with certainty. Shopify reports AI-driven traffic to its stores up eight times year over year and orders from AI-powered searches up nearly thirteen times, and OpenAI cites more than 700 million weekly ChatGPT users as the surface Instant Checkout sits on top of, with PayPal noting more than 400 million of its own users already shop through its platform. Those figures describe real scale, but every one of them is self-reported by a company with a direct commercial stake in the growth story, and none has been corroborated by an outside measurement. That is precisely the blind spot a demand-side attention map is built to close: not whether a platform says AI traffic is growing, but where a given market's attention actually sits across search, social, and the AI-answer layer today, and how fast it is migrating. It also reframes what 'agent-buyable' means for a retailer. The evidence gathered here has no figures at all on what it costs a merchant to become discoverable and transactable by an agent, structured product feeds, schema.org markup, mandate support, clear merchant-of-record terms, versus what that investment returns, which means the operational core of this shift is still an open question rather than a solved playbook. Until an independent read of where attention sits exists, a business's safest move is to build for structural readiness (clean, machine-readable product data, honest disclosure of any sponsored placement) while treating every vendor growth number in this space as directional, not proof of where its own customers already are.
The data, in one read
Six Rails, Eighteen Months
Between April 2025 and July 2026, six competing protocols for agentic payment shipped in quick succession, and none of them come from the same company. Mastercard's Agent Pay and Visa's Intelligent Commerce launched a single day apart in April 2025, an unusually tight coincidence for two networks that normally announce on their own schedule. Coinbase's x402 arrived the following month, reviving the long-dormant HTTP 402 'Payment Required' status code to let machines settle payments to each other directly in USDC. Google published the Agent Payments Protocol in September 2025, an open, payment-agnostic standard built around cryptographically signed Intent and Cart Mandates, with more than sixty launch partners on day one including Mastercard, PayPal, Coinbase, American Express, Salesforce, Etsy, Adyen, and Worldpay. OpenAI and Stripe co-developed the Agentic Commerce Protocol around the same window. Shopify and Google then built the Universal Commerce Protocol on top of that emerging set of standards, aiming to let a single merchant catalog transact across ChatGPT, Copilot, Google AI Mode, Perplexity, and Gemini's Universal Cart without separate integration work for each. Visa, not content with its April protocol, formalized a second, more technical layer in October 2025: the Trusted Agent Protocol, built with Cloudflare on RFC 9421 HTTP signatures, naming Anthropic, OpenAI, Perplexity, Mistral, and Samsung as model partners.
Two card networks announced competing agent-payment protocols a single day apart, an unusually tight coincidence that says more about the pressure in the room than either press release does.
From Announcement to Checkout
What separates this moment from a normal standards race is that several of these systems are already live, not just documented. OpenAI's Instant Checkout went live inside ChatGPT for US Plus, Pro, and Free users on February 16, 2026, with Etsy transacting from day one and more than a million Shopify merchants, including Glossier, Vuori, Spanx, and SKIMS, rolling out behind it. The mechanism uses a Shared Payment Token so ChatGPT itself never holds a buyer's card details, a design choice that matters given how much of the trust research below turns on the payment step specifically. Perplexity took a narrower path: a subscriber-only Buy with Pro feature underwritten by PayPal, covering more than five thousand merchants, paired with a separate, free Perplexity Merchant program that charges zero commissions or listing fees and has Perplexity itself absorbing shipping costs. Google's Universal Cart, the consumer-facing expression of the Shopify-built Universal Commerce Protocol, went live in early July 2026, with more than five million Shopify merchants eligible to appear in it. None of these are pilots anymore. They are live checkout surfaces with real merchants and real transactions running through them today, which is a meaningfully different claim than 'protocol announced.'
The Protocol Everyone Quietly Built On
Underneath the branded competition sits a detail worth pulling forward on its own: nearly every one of these systems, including Google's own AP2, is built on top of the Model Context Protocol, an open standard Anthropic introduced in November 2024 to let AI models call outside tools and data sources in a consistent way. By the time AP2 launched in September 2025, MCP had already been adopted by OpenAI and Google DeepMind as a de facto cross-vendor substrate, meaning the companies racing each other publicly on payment protocols are, at the infrastructure layer beneath the branding, building on the same open standard. That is a quieter story than the headline protocol war, but it is the more durable one: whichever payment protocol wins the next round of consolidation, the tool-calling layer underneath it is, for now, still Anthropic's.
The Stablecoin Undercurrent
A second infrastructure race is running alongside the payment-protocol one, and it is about settlement rails rather than checkout experience. Coinbase's x402 protocol has processed 169 million payments in its first year across 590,000 buyers and 100,000 sellers, settling in roughly 200 milliseconds on the Base network, with an x402 Foundation established in September 2025 to steward the standard. Visa is not standing outside that trend: its own stablecoin activity reached a seven billion dollar annualized run rate by April 2026, up 50% quarter over quarter, spanning more than 130 stablecoin-linked card programs across more than fifty countries, and Visa is actively working with Coinbase on x402 interoperability rather than competing head-on. Mastercard's move was even more direct: it agreed in March 2026 to acquire the stablecoin platform BVNK for up to 1.8 billion dollars, just under a year after launching Agent Pay. Read together, these three data points say the card networks do not see stablecoins as a fringe technology to watch from a distance. They are treating stablecoin settlement as core infrastructure for the machine-to-machine payment volume that agentic commerce is expected to generate.
Where Trust Actually Breaks
This is the finding that matters most for a business trying to plan around agentic commerce rather than just read about it. PYMNTS Intelligence, in a study with Visa Acceptance Solutions covering 5,241 consumers, 1,185 merchants, and 150 acquirers across the US, Brazil, and the UAE, found that 48% of online shoppers had already used AI to research their most recent purchase, essentially normal behavior for close to half of shoppers already. But the same study found 27% of shoppers trust no organization at all to operate an AI shopping agent on their behalf, and 24% say they will never delegate a purchase to AI under any circumstance. Accenture's much larger sixteen-country Consumer Pulse survey, 25,590 consumers fielded in January 2026, appears at first to contradict that caution: 74% said they would trust a personal AI agent more than their own best friend to buy for them. But look at what the rest of Accenture's numbers actually describe, and the picture resolves rather than conflicts. Only 9% of respondents are open to fully autonomous purchasing. Thirty-two percent would let an agent choose within a set budget and named brands. Fifty-six percent insist on specifying up front which brands the agent is even allowed to consider. Trust, in other words, is real and growing at the discovery and comparison stage of a purchase, and it collapses specifically at the point an agent would need authority to spend money without asking first.
Shoppers trust an agent to listen, research, and narrow the field, more than they expected to. They want a hand on the wheel the moment money actually moves.
Regulators Are Watching the Same Failure Mode
The trust research above explains why regulators are moving now rather than waiting for the market to mature. On July 1, 2026, the Federal Trade Commission published a proposed policy statement extending the FTC Act's ban on unfair or deceptive conduct to AI systems configured to steer purchase recommendations toward sponsored or affiliated options in a way a reasonable consumer would not expect, with a public comment period running through July 31, 2026. Separately, Senator Mark Warner released a discussion draft on June 29, 2026, the Artificial Intelligence Access, Gatekeeper Exchange, and Nondiscriminatory Transfer Act, that would require large online platforms, those with fifty million or more US users, to permit authorized third-party AI agents to act on a user's behalf while imposing fiduciary-style duties on the agent providers themselves, with the FTC administering registration and NIST setting technical standards. It has not yet been formally introduced as legislation. The specific behavior both proposals target has some early evidence behind it: a study cited by Forbes and attributed to Princeton researchers, titled 'Ads In Chatbots?,' reportedly found that almost every model tested favored a sponsored product over a cheaper unsponsored alternative once advertising entered the conversation, with some instances of concealed pricing or undisclosed paid placement. We were not able to locate the underlying paper directly in our research, so that finding should be read as reported by a business-press secondary source rather than confirmed against the primary study, but it names the exact mechanism, sponsored steering disguised as neutral recommendation, that both the FTC and Senator Warner's draft are now writing rules against.
What 'Agent-Buyable' Actually Requires, and What Nobody Has Measured Yet
Every protocol described above assumes a merchant on the other end with clean, structured, machine-readable product data: accurate feeds, consistent schema markup, clear merchant-of-record terms, and support for the mandate or verifiable-credential handshake each protocol uses to confirm a purchase. What is conspicuously absent from the public record is any figure for what that actually costs a retailer to build, or what it returns once built. We found no data on the merchant-side effort or ROI of becoming agent-buyable, no benchmark for how a structured, agent-ready catalog compares to an unstructured one in referral traffic or conversion, and no data yet on fraud, chargeback, or liability-dispute rates for agent-initiated purchases now that live checkout systems have real transaction history behind them. That is not a small gap. It is the operational center of the entire question a mid-size retailer actually has: not whether to care about agentic commerce, but what specifically to build first.
The Numbers Nobody Has Reconciled
Market-size forecasts for agentic commerce should be read with real skepticism right now, and not only because they come from parties with an interest in a large number. Juniper Research's own materials are internally inconsistent: the same report cites an $8 billion transaction-value figure for 2026 rising to $3.5 trillion by 2031 on its data page, while separate press messaging for the identical report quotes '1.5 trillion dollars globally by 2030,' a roughly twofold disagreement inside one publisher's own numbers. Secondary coverage then frequently blends that with Gartner's differently scoped figures, a business-to-business-specific $15 trillion by 2028 and a claim that 20% of digital commerce transactions will run through AI platforms by 2030, neither of which we were able to trace to a primary Gartner publication. Shopify's own growth claims sit in a similar place: 8x year-over-year AI-driven traffic and nearly 13x growth in orders from AI-powered search are real numbers from a real company, but they describe a vendor's own new product with no outside corroboration, which is why we hold them as contested rather than established. Set against that, OpenAI's reported reach, more than 700 million weekly ChatGPT users, and PayPal's more than 400 million users who already shop through its platform, describes the size of the surface agentic checkout is being layered onto. It is a real, measured audience number. It is not, on its own, a transaction census, and nothing in the public record currently is.
Where This Leaves a Mid-Size Business Today
None of the above adds up to a single number a business owner can plan a budget around. That reflects the state of the evidence: no reconciled figure yet exists. What does hold up: the underlying technical substrate is converging even while the branded protocols compete, real checkout is live today inside three major AI surfaces, the payment step is where every protocol and every consumer survey agrees trust runs out, and regulators are already writing rules against the one failure mode, sponsored steering disguised as neutral advice, that the research has actually documented. The practical response is not to bet on a single winning protocol. It is to make sure your own product data, pricing, and policies are clean and structured enough to be read correctly by whichever agent a customer happens to be using, to be transparent about anything that could look like sponsored placement before a regulator asks, and to measure where your own market's attention is actually moving rather than borrow a platform's growth chart as if it described your customers too.
The evidence, in numbers
Key findings, dated and sourced
-
OpenAI launched Instant Checkout, an in-chat purchase flow inside ChatGPT for US users, powered by the Agentic Commerce Protocol co-developed with Stripe, using a Shared Payment Token so ChatGPT never holds the buyer's card credentials. Protocol announced Sept 29, 2025; live Feb 16, 2026 for US Plus/Pro/Free users; Etsy live at launch, more than a million Shopify merchants (including Glossier, Vuori, Spanx, SKIMS) rolling out.
established OpenAI / Stripe, Stripe Newsroom, 'Stripe powers Instant Checkout in ChatGPT'; OpenAI, 'Buy it in ChatGPT' (2025-09-29 / 2026-02-16)
-
Google published the Agent Payments Protocol (AP2), an open, payment-agnostic standard using cryptographically signed Intent and Cart Mandates (W3C Verifiable Credentials), extending Anthropic's Model Context Protocol and Google's Agent2Agent protocol. More than 60 launch partners including Mastercard, PayPal, Coinbase, American Express, Salesforce, Etsy, Adyen, Worldpay.
established Google Cloud Blog, 'Announcing Agent Payments Protocol (AP2)' (2025-09-17)
-
Mastercard (Agent Pay, extending its Digital Enablement Service) and Visa (Intelligent Commerce, later formalized as the Trusted Agent Protocol with Cloudflare using RFC 9421 HTTP signatures) shipped competing agent-payment protocols about one day apart in April 2025. Mastercard Agent Pay launched April 2025 (LatAm/Caribbean expansion Dec 2025, partners Microsoft, IBM watsonx, Braintree, Checkout.com); Visa Intelligent Commerce launched the day after Mastercard, Trusted Agent Protocol with Cloudflare Oct 2025, model partners named as Anthropic, OpenAI, Perplexity, Mistral, Samsung.
emerging Mastercard / Visa, Forbes (Digital Assets), 'Visa, Mastercard And Coinbase Are Fighting Over How AI Agents Pay' (2025-04 through 2026-06)
-
Coinbase's x402 protocol, which revives HTTP 402 'Payment Required' to settle agent-to-agent payments in USDC, has processed real machine-to-machine transaction volume in its first year; Visa is working with Coinbase on x402 interoperability. 169 million payments processed in first year across 590,000 buyers and 100,000 sellers; roughly 200ms settlement on Base; x402 Foundation established Sept 2025; Visa stablecoin activity reached a $7 billion annualized run rate by April 2026 (up 50% quarter over quarter; 130+ stablecoin-linked card programs, 50+ countries).
emerging Coinbase / Visa, Forbes (Digital Assets), 'Visa, Mastercard And Coinbase Are Fighting Over How AI Agents Pay' (2025-05 launch; figures as of approx. 2026-06)
-
Shopify and Google co-developed the Universal Commerce Protocol (UCP) so AI agents (ChatGPT, Copilot, Google AI Mode, Perplexity, Gemini's Universal Cart) can transact against Shopify's merchant base; Shopify self-reports early AI-referred traffic and order growth. More than 5 million Shopify merchants eligible; Google Universal Cart went live early July 2026; Shopify's own Q1 2026 reporting states AI-driven traffic to stores up 8x year-over-year, orders from AI-powered searches up nearly 13x year-over-year.
contested Shopify / Google, Shopify, 'AI commerce at scale' / 'How agentic commerce works on Shopify', a vendor blog reporting growth in its own new product, uncorroborated by independent measurement (2026-Q1 metrics; UCP/Universal Cart live 2026-07)
-
Perplexity launched a subscriber-only agentic shopping feature, Buy with Pro, with PayPal as merchant-of-record, plus a separate free Perplexity Merchant program charging zero commissions or listing fees. PayPal integration covers more than 5,000 merchants; Perplexity funds free shipping itself; the Merchant Program has zero fees or commissions.
established Perplexity / PayPal, CNBC, 'Perplexity announces free product to streamline online shopping' (Nov 19, 2025); Perplexity Blog, 'Shop like a Pro' (program build-out through 2025-11 into 2026)
-
A large multi-country survey found nearly half of online shoppers already use AI to research purchases, but trust collapses at the payment-authorization step; a meaningful minority say they will never delegate purchasing to AI at all. 48% of online shoppers used AI to research their most recent purchase; 27% trust no organization to operate an AI shopping agent; 24% say they will never delegate purchases to AI.
established PYMNTS Intelligence / Visa Acceptance Solutions, 'The Agentic Commerce Deep Dive: Payment Infrastructure and the Path to Trust for Merchants' (5,241 consumers, 1,185 merchants, 150 acquirers; US, Brazil, UAE) (published 2026-07-16 (fieldwork: consumers/merchants Mar 2026, acquirers Jan 2026))
-
Accenture's large 16-country Consumer Pulse survey found most consumers say they trust a personal AI agent more than their best friend to make a purchase for them, but actual willingness to grant full purchasing autonomy is low and most want to keep control over which brands are considered. 74% would trust a personal AI agent more than their best friend to purchase on their behalf; only 9% are open to fully autonomous purchasing; 32% would let an agent choose within set budget/brand parameters; 56% would specify which brands the agent may consider.
established Accenture Consumer Pulse Research, 'Talk to My AI Agent: The New Rules of Brand Value,' 25,590 consumers across 16 countries, fielded Jan 2026 (2026-06)
-
On July 1, 2026 the FTC published a proposed policy statement extending the FTC Act's ban on unfair or deceptive conduct to AI systems configured to steer purchase recommendations toward sponsored or affiliated options in a way a reasonable consumer would not expect. Comment period ran through July 31, 2026; frames AI-agent purchase-steering-for-hidden-commercial-gain as Section 5 deceptive advertising.
established Federal Trade Commission, FTC proposed policy statement on AI accuracy / deceptive steering (via Forbes, Reed Smith, Consumer Finance Monitor coverage; the primary FTC.gov page was not directly accessible in this research pass, so treat as corroborated by secondary-source consensus rather than primary-confirmed) (2026-07-01)
-
Senator Mark Warner released a discussion draft of a federal AI Agent Act that would require large online platforms to permit authorized third-party AI agents to act on a user's behalf, while imposing fiduciary-style duties and technical governance requirements on agent providers. Discussion draft (the Artificial Intelligence Access, Gatekeeper Exchange, and Nondiscriminatory Transfer Act of 2026) released June 29, 2026; not yet formally introduced as legislation; applies to platforms with 50 million or more US users, with the FTC administering registration and NIST setting technical standards.
emerging US Senate (Sen. Mark Warner), Davis Wright Tremaine, 'The Federal AI AGENT Act: Consumer Protection in AI Clothing?' (2026-06-29 (discussion draft))
-
A study titled 'Ads In Chatbots? An Analysis of How Large Language Models Navigate Conflicts of Interest,' attributed to Princeton researchers, reportedly found LLMs almost universally favored a sponsored product over a cheaper unsponsored alternative once advertising entered the conversation, with higher-income simulated users shown sponsored content more often and some concealed pricing or undisclosed paid placement. 'Almost all' tested models recommended the sponsored option over a cheaper non-sponsored one. The underlying paper could not be located directly via academic search in this research pass; the finding is reported here only via a secondary business-press writeup and should be re-verified against the primary paper before being treated as confirmed.
emerging Princeton University (via Forbes secondary report), Forbes (Pamela N. Danziger), 'Washington Watchdogs Take Notice As AI Becomes Retail's Front Door' (cited in coverage dated 2026-07-19)
-
Secondary reporting cites a University of Virginia Darden School figure that most shoppers have already used AI in a shopping context, plus eMarketer forecasts for AI-mediated ad spend; neither figure was verified against its primary source. Nearly 60% of shoppers have used AI to shop (Darden School, via Forbes); 46% of consumers trust AI more than friends for outfit advice (Darden School, via Forbes); AI ad spending projected at $32 billion in 2026, rising to $68 billion by 2030 (eMarketer, via Forbes).
emerging University of Virginia Darden School / eMarketer (via Forbes), Forbes (Pamela N. Danziger), 'Washington Watchdogs Take Notice As AI Becomes Retail's Front Door'; primary Darden and eMarketer sources not independently located (cited 2026-07-19)
-
Mastercard made a large stablecoin-infrastructure acquisition shortly after launching its agent-payments protocol, signaling card networks treat stablecoin rails as core to the agentic-commerce payment stack. Agreed in March 2026 to acquire stablecoin platform BVNK for up to $1.8 billion.
established Forbes (Digital Assets), 'Visa, Mastercard And Coinbase Are Fighting Over How AI Agents Pay' (2026-03)
-
The technical substrate underlying most 2025-2026 agentic-commerce protocols (AP2, and agent-to-merchant tool calls generally) is the Model Context Protocol, an open standard Anthropic introduced in November 2024, subsequently adopted by OpenAI and Google DeepMind as a de facto cross-vendor substrate. Introduced Nov 2024; adopted by OpenAI and Google DeepMind by the time AP2 (Sept 2025) explicitly built on top of it.
established Anthropic, Wikipedia, 'Model Context Protocol' (2024-11 (origin); ongoing adoption through 2025-2026)
-
Market-size forecasts for agentic commerce are not reconciled across or even within research houses: Juniper Research's own materials state both a $1.5 trillion globally by 2030 headline figure and a separate $8 billion (2026) to $3.5 trillion (2031) transaction-value trajectory on the report page; secondary sources conflate this with Gartner's differently scoped B2B forecast.
contested Juniper Research / Gartner, Juniper Research, 'Agentic Commerce Market Report: Size, Trends 2026-31' (Juniper report published 2026-04-07; Gartner figures secondary-sourced, undated at primary)
-
OpenAI's reported weekly reach for ChatGPT, cited in coverage of its commerce rollout, establishes the scale of the surface Instant Checkout is being layered onto, alongside PayPal's user base as a payments partner. 'More than 700 million people turn to ChatGPT each week'; more than 400 million PayPal users shop via the platform.
established OpenAI / PayPal (via Digital Commerce 360), Digital Commerce 360, 'OpenAI expands agentic commerce push' (Feb 16, 2026) (reported as of the Instant Checkout / ACP rollout period, 2025-09 through 2026-02)
Learning outcomes
What this study teaches
- Don't wait for a single protocol to win. Every system in this race, ACP, AP2, UCP, TAP, Agent Pay, x402, assumes clean, structured, machine-readable product data underneath it, so building that now pays off regardless of which brand of checkout a customer's agent happens to use.
- Design for an agent that asks permission, not one with a blank check. The trust data is consistent across two separate large surveys: shoppers will let an agent research and narrow choices, but only a single-digit share will grant it full purchasing authority, so any agent-facing flow you build should assume a human confirms the final step.
- Treat platform growth numbers as a claim, not a measurement of your own market. Shopify's traffic and order multipliers and OpenAI's reach figures describe those companies' own results; they are not evidence of what is happening in your specific category or customer base until you look at your own numbers.
- Review your own recommendation and placement practices now. The FTC's proposed steering policy and the pattern the Princeton-attributed study describes, sponsored results dressed as neutral advice, are the exact behavior regulators are moving to police, and it is worth checking your own disclosures before that scrutiny arrives.
- Expect the payment step to stay human-gated for a while. Every major protocol is architected around signed, revocable consent rather than standing authorization, which matches what consumers say they want, so plan checkout experiences around confirmation, not autonomy, for the foreseeable future.
Honest limits
What this does not yet settle
- No independent, non-vendor-reported census yet exists for how much shopping attention or transaction volume actually flows through AI agents versus classic search or direct-to-site traffic. Every growth figure gathered here, Shopify's 8x AI traffic and nearly 13x AI orders, OpenAI's checkout reach, is self-reported by a platform with a commercial stake in the narrative. This is the exact gap an independent measure of attention is built to close, and it should not be treated as settled until measured outside the platforms themselves.
- Market-size forecasts for agentic commerce are internally inconsistent even within a single research house (Juniper Research's own report page and press messaging disagree by roughly twofold on the 2030/2031 endpoint) and are frequently conflated across sources with differently scoped figures (Gartner's B2B-specific $15 trillion-by-2028 claim gets cited elsewhere as if it were a total agentic-commerce number). No reconciled, methodology-transparent market-size figure currently exists.
- No peer-reviewed academic literature specific to agentic-commerce trust, adoption, or behavior was located in this research pass. The evidence base here is trade press, vendor communications, and paid consultancy or survey research (PYMNTS, Accenture, Juniper), not academic literature. This is a field still ahead of scholarly measurement.
- The Princeton 'Ads in Chatbots' study on sponsored-recommendation steering could not be located and read directly; it is included here only via a secondary business-press writeup, and its exact figures, sample, and methodology should be re-verified against the primary paper before it is cited as settled fact.
- No data yet exists on real-world fraud, chargeback, or liability-dispute rates for agent-initiated purchases now that Instant Checkout and card-network agent tokens are live in the US. This is a foreseeable next data point once these systems have run a full return-and-dispute cycle.
- No figures were found for the merchant-side cost, effort, or return of becoming agent-buyable, implementing ACP, AP2, or UCP, structured product feeds, schema.org markup, versus the traffic or revenue lift it produces. This is the operational core of what agentic commerce actually asks of a merchant, and it remains open.
- Gartner's specific $15 trillion-by-2028 and 20%-by-2030 figures were found only through secondary aggregator coverage, not a primary Gartner publication, and should be treated as directional until sourced to Gartner directly.
This is a synthesis of dated, attributed evidence, not a census. The AI-answer layer in particular has no independent, Nielsen-grade measurement yet, so readings of it are directional and named as a frontier, never presented as settled.
Straight answers
Frequently asked questions
Is AI agent checkout actually live, or is this still just protocol announcements?
It is live, not just announced. OpenAI's Instant Checkout went live inside ChatGPT for US Plus, Pro, and Free users on February 16, 2026, with Etsy transacting from day one and more than a million Shopify merchants rolling out behind it. Perplexity's Buy with Pro (backed by PayPal, covering more than five thousand merchants) and Google's Universal Cart, live in early July 2026, are also running real checkouts today, not pilots.
Do shoppers actually trust AI agents to make purchases for them?
It depends on which part of the purchase you mean. Accenture's 16-country survey of 25,590 consumers found 74% would trust a personal AI agent more than their own best friend to buy for them, but only 9% are open to letting it purchase fully autonomously, and 56% insist on specifying up front which brands the agent may even consider. Separately, PYMNTS Intelligence and Visa Acceptance Solutions found 27% of shoppers trust no organization at all to operate an AI shopping agent, and 24% say they will never delegate a purchase to AI under any circumstance. Trust is real at the research and comparison stage and collapses specifically at the moment an agent would need authority to spend money without asking.
Why are Visa, Mastercard, and Coinbase all pushing stablecoins into agent payments?
Because they are treating stablecoin settlement as core infrastructure for machine-to-machine payment volume, not a side experiment. Coinbase's x402 protocol processed 169 million payments in its first year across 590,000 buyers and 100,000 sellers, Visa's own stablecoin activity reached a seven billion dollar annualized run rate by April 2026 (up 50% quarter over quarter), and Mastercard agreed in March 2026 to acquire the stablecoin platform BVNK for up to 1.8 billion dollars. Visa is also working with Coinbase on x402 interoperability rather than competing head-on.
How much of my customers' shopping is actually happening through AI agents right now?
Nobody outside the platforms can say with certainty yet; that is the current state of the evidence. Shopify reports AI-driven traffic to its stores up eight times year over year and orders from AI-powered search up nearly thirteen times, and OpenAI cites more than 700 million weekly ChatGPT users as the surface Instant Checkout sits on top of, but every one of those figures is self-reported by a company with a direct commercial stake in the growth story and none has been corroborated by an outside measurement. The study treats these numbers as directional, not proof of where a given business's own customers already are.
What should a mid-size business actually do about agentic commerce right now?
Build for structural readiness rather than betting on a single winning protocol: every system in this race (ACP, AP2, UCP, TAP, Agent Pay, x402) assumes clean, structured, machine-readable product data underneath it, so that investment pays off no matter which checkout brand a customer's agent uses. Also design any agent-facing purchase flow to require a human confirmation step, since every protocol and both major consumer surveys agree trust runs out right at payment, and review your own recommendation and placement disclosures now, since the FTC's July 2026 proposed policy targets sponsored steering disguised as neutral advice.
Provenance
References
- Stripe Newsroom, 'Stripe powers Instant Checkout in ChatGPT' https://stripe.com/newsroom/news/stripe-openai-instant-checkout
- Google Cloud Blog, 'Announcing Agent Payments Protocol (AP2)' https://cloud.google.com/blog/products/ai-machine-learning/announcing-agents-to-payments-ap2-protocol
- Forbes (Digital Assets), 'Visa, Mastercard And Coinbase Are Fighting Over How AI Agents Pay,' Jun 7, 2026 https://www.forbes.com/sites/digital-assets/2026/06/07/visa-mastercard-and-coinbase-are-fighting-over-how-ai-agents-pay/
- Shopify, 'AI commerce at scale' https://www.shopify.com/news/ai-commerce-at-scale
- CNBC, 'Perplexity announces free product to streamline online shopping,' Nov 19, 2025 https://www.cnbc.com/2025/11/19/perplexity-ai-online-shopping-paypal.html
- PYMNTS Intelligence / Visa Acceptance Solutions, 'The Agentic Commerce Deep Dive: Payment Infrastructure and the Path to Trust for Merchants' https://www.pymnts.com/news/artificial-intelligence/2026/48percent-of-online-shoppers-now-use-ai-before-buying/
- Accenture Consumer Pulse Research, 'Talk to My AI Agent: The New Rules of Brand Value' https://www.accenture.com/us-en/insights/consulting/talk-my-ai-agent
- Federal Trade Commission, proposed policy statement on AI accuracy and deceptive steering, Jul 1, 2026 https://www.ftc.gov/news-events/news/press-releases/2026/07/ftc-seeks-public-comment-policy-statement-addressing-ai-accuracy
- Davis Wright Tremaine, 'The Federal AI AGENT Act: Consumer Protection in AI Clothing?' https://www.dwt.com/blogs/artificial-intelligence-law-advisor/2026/07/ai-agent-act-consumer-ai-regulation
- Forbes (Pamela N. Danziger), 'Washington Watchdogs Take Notice As AI Becomes Retail's Front Door,' Jul 19, 2026 https://www.forbes.com/sites/pamdanziger/2026/07/19/washington-watchdogs-take-notice-as-ai-becomes-retails-front-door/
- Wikipedia, 'Model Context Protocol' https://en.wikipedia.org/wiki/Model_Context_Protocol
- Juniper Research, 'Agentic Commerce Market Report: Size, Trends 2026-31' https://www.juniperresearch.com/research/fintech-payments/ecommerce/agentic-commerce-research-report/
- Digital Commerce 360, 'OpenAI expands agentic commerce push,' Feb 16, 2026 https://www.digitalcommerce360.com/2026/02/16/openai-expands-agentic-commerce-push/
Every measured figure is dated to its capture and tagged with an evidence tier. Every cited work is real and locatable. Where an engine could not be captured this round, it is named as uncaptured, not estimated. Small-sample readings are labelled as directional.