How ACP works
Selling through an AI assistant needs two things that ordinary ecommerce does not provide. The assistant has to know exactly what you sell, at what price, in what sizes and whether it is in stock. And it needs a defined way to place the order without pretending to be a human clicking through your checkout. ACP, the Agentic Commerce Protocol, published openly by OpenAI with Stripe, covers both halves.
The first half is a product feed specification: a structured file describing your catalogue in a shape the assistant can read, in the same spirit as the feed you already maintain for shopping ads. The second half is a checkout interface the assistant calls to build a basket, confirm price and availability, and submit the order. The design intent is that the merchant stays the merchant of record — you take the payment, you own the order, you own the customer relationship and the returns.
Because it is a published standard rather than a private integration, any assistant that adopts it can in principle sell from any shop that supports it. That is the whole appeal, and also the reason nothing happens until both sides adopt it.
Why ACP matters
Search has spent years moving the moment of decision earlier — into the results page, then into the answer. Agentic commerce moves the moment of purchase with it. If a shopper describes what they want to an assistant and the assistant can buy it, the product listing that wins is the one an assistant can read and trust, not the one with the prettiest page.
That makes catalogue quality a commercial issue rather than a technical chore. Missing attributes, stale stock, prices that disagree between systems, thin titles — those already cost you in product feed driven advertising, and they will cost you again here. Businesses that keep a clean, complete feed are not doing anything special for agents; they are simply already legible.
Where ACP goes wrong
The first mistake is treating it as a marketing channel to chase. It is plumbing. There is no campaign to launch, no bid to set, and no benefit at all until your platform and your assistant of choice both support it.
The second is letting the feed drift out of step with the shop. A human on a slightly wrong product page will forgive you and check the basket. An automated buyer will not: it takes what the feed says, and a price or stock mismatch turns into a cancelled order, a refund and a complaint you never had a chance to intercept.
The third is measurement. An order that arrives through an assistant does not carry the referrer, cookies or campaign parameters your reporting expects, so it can land as direct traffic or as no session at all. If agent sales start arriving and nobody has planned for that, your channel reports will quietly mislead you.
What to do about it
Start with the boring asset. Get one authoritative product feed that matches live prices and stock, with complete titles, attributes, images and identifiers, and keep it refreshed automatically rather than by hand. The same discipline that improves shopping feed management prepares you for agent channels without a separate project.
Then ask your ecommerce platform and payment provider where agentic checkout sits on their roadmap, and confirm who is merchant of record in any arrangement they offer, because that determines who handles refunds, chargebacks and customer data. Finally, decide in advance how you will label these orders so they are visible in reporting rather than lost in direct. If the feed and order plumbing are already automated in your shop, this is a small extension of existing ecommerce automation rather than a new build.