Most of It Went Through the Mesh. They Counted What Stayed.

Twenty-three per cent can see it. The rest are guessing#

PYMNTS Intelligence, in research commissioned by Visa Acceptance Solutions and published on 2 September 2026, reports that only 23 per cent of merchants can clearly identify both AI-driven traffic and AI-driven purchases. A further 21 per cent can see agentic traffic arriving but cannot connect it to a completed sale. Size helps only slightly: 27 per cent of large merchants can identify agent-driven purchases against 19 per cent of small and medium businesses.

Merchants are not, however, sanguine about where this goes, and the scent of the thing is already on the trail. In the same study 61 per cent agreed AI-generated results will influence purchases more than traditional search, 58 per cent expect agents to choose payment methods based on fees and rewards, and 56 per cent expect agents to complete transactions themselves. Thirty-eight per cent expect agent purchases to exceed 15 per cent of sales within two years.

The survey predates the standard that answers it#

One detail changes how this research should be read, and it is not on the study's own landing page. The landing page states the sample, 1,185 retail merchants across the United States, Brazil and the United Arab Emirates, and names Visa Acceptance Solutions as commissioning sponsor, but publishes no fieldwork dates. Those appear elsewhere in PYMNTS's own coverage: the merchant survey was fielded in March 2026, part of a three-survey programme that also covered 5,241 consumers in March 2026.

So this agentic commerce finding is six months old at publication, in a field where six months is a long time. That matters specifically, not just generally, because the technical answer to the attribution problem was published in the interval. The merchants were asked whether they could identify agent purchases before the specifications that make agent purchases identifiable had settled.

Three specifications already carry the answer#

The reason to be impatient with the 23 per cent figure is that identifying an agent at checkout is a solved problem on paper, and has been for most of the period agentic commerce has been discussed seriously. Three separate published specifications address it, and all three were public before this research appeared.

Agentic Commerce Protocol: the agent announces itself#

The Agentic Commerce Protocol is maintained by OpenAI and Stripe as founding maintainers, licensed Apache 2.0 and currently in beta. Its agentic checkout OpenAPI specification, versioned 2026-04-17, defines the header set a compliant agent sends, including Authorization, Idempotency-Key, Request-Id, Signature, Timestamp and User-Agent. The specification's own example User-Agent value is a ChatGPT client string. A merchant terminating that request has the agent's identity in the request headers, not inferred from behaviour.

AP2: mandates that record what was authorised#

AP2, the Agent Payments Protocol, at version 0.2.0, approaches the same problem from the authorisation side. It defines a Checkout Mandate capturing the specific items and purchase details negotiated between agent and merchant, and a Payment Mandate authorising the payment itself. Google's announcement lists more than sixty participating organisations including Accenture, Adobe, Adyen, American Express, Coinbase, Deloitte, Intuit, JCB, Mastercard and MetaMask. That is a broad enough coalition that no merchant should be planning as though agentic payments will remain unstandardised.

Visa Trusted Agent Protocol: signed requests#

Visa's Trusted Agent Protocol specifications define Signature-Input and Signature header fields carrying message-signature metadata and the signature itself. That is cryptographic attribution rather than heuristic detection: a merchant can verify that a request came from a recognised agent instead of guessing from user-agent strings and traffic shape. Visa has been building this out publicly through its own newsroom alongside the developer specification, which is worth watching because the card networks, not the model providers, set the rules that decide disputes.

Between them these three specifications cover identification, authorisation and verification, which is most of what agentic commerce reporting actually requires. The gap the research measures is therefore not a gap in what is possible. It is a gap in adoption, which is a very different problem with a very different remedy.

The part the specifications do not settle: who pays#

Attribution is the measurable half of the problem, and the easier half to drag out of the undergrowth. The half that decides whether agentic commerce is safe to encourage is liability, and that is where practitioners are considerably sharper than the survey.

The claim about basket sizes in that thread is one operator's report rather than published data, and should be treated as such. The structural argument underneath it does not depend on the number. If an agent selects the wrong item, or a customer disputes a purchase they say they never authorised, the chargeback lands on the merchant, who sits on the card network's hook rather than the model provider's. Upside is shared, downside is not.

This is precisely what AP2's mandate structure is designed to address, by creating a record of what the customer actually authorised before the agent acted. Whether that record survives contact with a dispute process built for human buyers is an open question, and not one any merchant should assume the answer to. Anyone enabling ai agent payments this year is making a bet on a dispute regime that has not yet been tested at volume.

What to do before agentic checkout arrives properly#

A fox holding a sieve under a stream does not conclude the stream is empty. It changes the mesh, the way a fox that loses a quarry in the hedgerow changes where it waits rather than deciding the wood is empty. The merchants in this research are measuring with the wrong instrument and reporting the result as an absence.

Start by logging what already arrives at the den door. Agent traffic is reaching most merchants now, and the 21 per cent who can see traffic without connecting it to purchases are one join away from the answer rather than one platform migration away. Capture the user-agent and signature headers the specifications define, and store them against the session and the order, so the connection exists when somebody finally asks for it.

Then decide the commercial questions agentic commerce specifications deliberately leave open. Whether ai shopping agents should be allowed to select payment method on fee and reward grounds, given 58 per cent of merchants expect exactly that. What happens to loyalty programmes when the buyer is not the person. Who carries a disputed agent purchase in your own terms. None of those are answered by adopting a protocol, and all of them will be answered by default if nobody chooses deliberately.

The unglamorous truth is that agentic commerce will arrive as a reporting problem long before it arrives as a revenue one. The merchants who can see it will be able to price it, argue about it and negotiate over it. The ones who cannot will be told what happened by somebody with an interest in the answer.

Frequently asked questions#