Key Takeaways
- AI agents do not "browse" product pages; they query structured data and pick a counterparty in seconds, which means a site built for humans may be invisible to them.
- PYMNTS identifies three things merchants must prepare now: machine-readable product data, records of agent authorization, and registration with the card networks.
- Visa and Mastercard already offer frameworks that let merchants participate without a major build, and registration is starting to separate the "chosen" merchants from the "skipped" ones.
An AI Agent Is Not a Customer Who Browses
Most merchants have prepared for every kind of customer. The one exception is the customer arriving right now. AI agents do not look at product pages, do not click links, and do not read product descriptions written for humans. They query, evaluate, and complete a transaction in seconds.
This week's Prompt Economy Weekly looks at the three things merchants need before AI agents start buying at scale.
www.pymnts.comIn this article, PYMNTS argues that commerce is shifting from discovery to execution. Visa reported that nearly half of U.S. shoppers now use AI tools for at least one shopping task, and the company predicts that millions will use AI agents to complete purchases by the 2026 holiday season. A merchant whose infrastructure is not built to receive these agents does not lose to a better price; it loses the sale before the agent even decides to compare. That is the warning PYMNTS puts in front of merchants.
The article then lays out three requirements that determine whether a merchant can participate at all. Let us take them in turn.
| Readiness item | What happens without it | First move |
|---|---|---|
| Machine-readable product data | You are invisible to agents and drop out before the comparison even begins | Audit the agent traffic already hitting your site and how structured your feed is |
| Authorization and intent records | You cannot prove liability when a chargeback or fraud claim is filed | Build a way to record what the agent was authorized to do and what it actually did |
| Network registration | You are not recognized as a legitimate merchant and traffic is routed elsewhere | Apply to Visa and Mastercard agent frameworks and obtain network tokens |
Make Your Product Data Machine-Ready
The first requirement, and the one that carries the heaviest practical workload for many businesses, is making product data machine-readable. An agent does not browse. It sends a structured request, reads a structured response, and instantly decides whether to transact or move to the next merchant.
Product pages built for human eyes are largely invisible to agents operating at speed. Titles written for SEO, descriptions layered with marketing language, images optimized for display — none of it translates into the machine-readable format an agent needs to compare price, availability, delivery window, and return policy in a single pass. The IMF, in its April 2026 paper on agentic payments, noted that AI agents reduce transactional friction by autonomously handling product search, price comparison, discount application, and availability verification. That efficiency only activates for merchants whose catalog data is structured, real-time, and normalized.
Merchants who have already built clean data pipelines for Google Shopping or affiliate channels are closest to ready. The flip side is a growing consensus that traditional product feeds cannot simply be reused. Comparison shopping engines designed feeds for humans to evaluate, but in agentic commerce the AI agent itself becomes the decision layer, which calls for a fundamentally different data interface.
The on-ramps are taking shape, too. Under the Agentic Commerce Protocol (ACP), led by Stripe and OpenAI, a merchant builds a product feed endpoint that conforms to the ACP spec and updates the data daily. Google's Universal Commerce Protocol (UCP), which launched in January 2026, counts Shopify, Walmart, Target, and Etsy among its founding partners. If attributes are missing or shipping details are vague, the AI skips that store and recommends a competitor with better data.
The first move PYMNTS recommends is simple: audit what kind of agent traffic is already hitting your site and which endpoints it is targeting. Preparation starts there.
Wire In Authorization Before the Agents Arrive
When an AI agent completes a purchase on a consumer's behalf, the traditional chain of authorization breaks. There is no human at checkout. There is no moment where a real person confirms the cart and clicks buy.
That gap weighs heavily on disputes, chargebacks, fraud liability, and regulatory review. What was the agent authorized to do, what did it actually do, and did those two things match? A merchant who cannot prove this after the fact is exposed — not through fraud, but through pure infrastructure debt.
As an answer, Mastercard partnered with Google to introduce Verifiable Intent, an open, standards-based framework. It confirms the cardholder who authorized the AI agent, captures the consumer's specific instructions, and records the interaction between agent and merchant that resulted in a purchase. It links the consumer's identity, the instructions and limits given to the agent, their consent, and the actions the agent took into a single tamper-resistant record — a cryptographic audit trail all parties can consult if a dispute arises. We explore this further from the angle of the dispute infrastructure for agentic commerce.
Accepting the payment is not enough — that is the heart of the PYMNTS argument. What a merchant needs is a record that survives scrutiny. Mastercard says its Agent Pay Acceptance Framework enables merchant participation without demanding significant development or integration, establishing a consistent standard for agent verification and data exchange. A merchant who cannot produce that record when a chargeback is filed is, in effect, transacting while carrying infrastructure debt.
Register With the Networks or Risk Being Skipped
The third requirement involves the least work to set up, yet skipping it can be fatal. Both major card networks offer frameworks that let merchants join agentic commerce without a full integration rebuild. The IMF found that Visa's Intelligent Commerce and Mastercard's Agent Suite both center on Know Your Agent frameworks, providing the registration, cryptographic signatures, and network tokens required to distinguish legitimate agents from malicious bots.
A merchant that has not registered is transacting in a gray zone. Agents cannot verify whether the merchant is legitimate, and the network cannot apply its fraud controls. When an agent meets that ambiguity, its default is not to try harder. It is to route to a merchant that has already done the work.
In April 2026, Visa unveiled Intelligent Commerce Connect as a network-, protocol-, and token-vault-agnostic "on-ramp" to agentic commerce for merchants and agent builders alike. The seemingly mundane act of registering is becoming the fork that separates the merchants who get chosen from those who get skipped.
Conclusion
None of these three requirements has to be finished overnight. It is enough to start by checking where the gaps are — auditing product data, designing authorization records, and registering with the networks, in that order. Only the merchants who move before agents start shopping in earnest will be on the chosen side of the first wave. While there is still room to prepare, the time to take the first step is now.
