Agentic Commerce, Explained: What Changes When the Shopper Is an AI

For twenty-five years online retail assumed a human at the other end: someone who reads a page, notices a photograph, forms an impression of whether a shop looks trustworthy. Almost everything about how commerce data is published was built for that reader. Agentic commerce breaks the assumption. The thing reading your product page increasingly is not a person — it is a model acting on a person's behalf, and it reads differently, decides differently and fails differently.

The term covers a real shift, but it is used so loosely that it has started to mean nothing. This is an attempt to be precise: what the protocols do, where the boundaries between them fall, and which problem none of them solve.

Three layers, often confused

The confusion is worth clearing up first, because the three things people mean by "agentic commerce" sit at different layers and solve different problems.

Connection: how an assistant reaches a tool

The Model Context Protocol (MCP) is a general standard for letting a model call external tools and read external data. It is not a commerce protocol at all — it is a connector standard, the way an assistant discovers that a service exists and what questions it can be asked. A retailer or comparison service exposes an MCP server; an assistant calls it and gets structured results back instead of scraping HTML.

Transaction: how an agent actually buys

The Agentic Commerce Protocol (ACP), from OpenAI and Stripe, and Google's Universal Commerce Protocol (UCP) both address the harder half: an agent that does not just recommend a product but completes a purchase. Payment credentials, order state, delivery, cancellation, refunds, and the question of what the merchant is agreeing to when the buyer is software. These are genuinely difficult problems and both specifications are serious attempts at them.

Description: how an offer is expressed

Schema.org has described products and offers in machine-readable form for over a decade. It is the vocabulary — price, currency, availability, identifier — that most structured commerce data already speaks. Agentic protocols lean on it rather than replacing it.

Connection, transaction and description are all solved problems with published specifications. The gap is evidence: what has to be true before a price may be stated as fact.

The layer nobody specified

Notice what is missing from that stack. MCP says how to ask. ACP and UCP say how to pay. Schema.org says how to phrase the answer. None of them say anything about whether the answer is true.

A perfectly conformant pipeline can carry a completely wrong price. The protocol does not care whether the number was read from the retailer this morning or copied from a feed three weeks ago; whether the product identified is the 200ml bottle or the 75ml one; whether the "in stock" flag reflects a check or a default. Every one of those failures produces valid, well-formed, confidently-worded output.

That is a bigger problem for agents than it ever was for humans, for one structural reason: a human shopper who lands on a product page sees the real price immediately and silently corrects whatever the comparison site said. The correction happens so fast nobody counts it as a failure. An agent that quotes a wrong price and then completes a purchase never gets that correction.

What breaks first

From watching this category, three things fail before anything else does — and none of them are payment problems, which is where most of the specification effort has gone.

1. Identity. The agent confidently prices a different product: the accessory rather than the device, the smaller size, the previous model year. Title matching is the culprit almost every time.
2. Freshness. The number was true when it was collected and is not true now, and nothing in the response says how old it is — so the agent presents a stale reading in the present tense.
3. Provenance. The price came from a promotional feed rather than the retailer's own surface, and the difference is invisible downstream because the field is just a number.

All three are checkable. None of them are checked by default, because the formats have no place to put the answer.

What a good agentic answer contains

If you are building or evaluating anything in this space, these are the fields that separate a usable commerce answer from a plausible one. They cost almost nothing to carry and they change what an assistant is able to say responsibly.

• An identifier, not a name. A barcode or the retailer's own listing id resolves what a title only suggests.
• A timestamp on every price, and the age stated in the answer rather than left for the caller to compute.
• A provenance label saying how the price was obtained, so an advertisement is distinguishable from an observation.
• Availability as three-valued: in stock, out of stock, or not checked. Never a default.
• An explicit refusal path — a way to say "cannot be verified" that is different from "no" and different from silence.

The last one matters more than it looks. An assistant that cannot distinguish "I checked and the claim is false" from "I could not check" will eventually tell a shopper a genuine offer is fake, which is a worse failure than saying nothing at all.

Who this changes things for

For retailers, the shift is that your structured data is now read by something that cannot be charmed. Marketing copy does not travel; identifiers, prices and availability do. A page that is beautiful and structurally silent is invisible to an agent.

For comparison and affiliate businesses, the shift is more uncomfortable. A model that can call several retailers directly does not need a middleman for discovery. What it does need is something it cannot do itself: independent evidence that a price is real. That is a narrower business than "comparison", and a more defensible one.

For shoppers, the practical advice is unchanged and slightly boring. When an assistant quotes a price, ask it when that price was checked and where it came from. If it cannot answer, treat the number as a starting point rather than a fact — and check the retailer's own page before you buy anything expensive.