What Is the Agentic Commerce Protocol (ACP)? A Merchant's Guide

What the Agentic Commerce Protocol Actually Is

The Agentic Commerce Protocol is an open standard that defines how an AI agent talks to a business in order to buy something for a shopper. It was created by Stripe, OpenAI and Meta, the specification is maintained by OpenAI and Stripe, and it lives in the open at agenticcommerce.dev with a public GitHub repository and dated spec releases.

The protocol is built out of separate pieces rather than as one monolithic integration. There is a product feed and cart layer for browsing a catalog and assembling an order. There is agentic checkout for creating and completing a checkout session with fulfillment options. There is delegated payment, which passes a payment token between the buyer, the agent and the business without handing raw card details around. There is delegated authentication over OAuth 2.0, so an agent can act on a shopper’s behalf within limits. And there are order webhooks for confirmation, shipping, delivery and refunds.

Read that list again and notice the shape of it. Only one of those pieces is about your product data. All the others are about money, identity and order state. That distinction turns out to matter enormously, and it is the reason most of the coverage you have read about ACP is aimed at the wrong reader.

The Timeline Most Merchants Missed

ACP went live in ChatGPT in September 2025, launching as Instant Checkout with Etsy as the first US partner and Stripe powering the payments. It was a narrow pilot, and most merchants had no way in.

February 2026 was the moment it stopped being niche. OpenAI relaunched the experience as “Buy it in ChatGPT” and opened it to more than a million Shopify merchants, with names like Glossier, SKIMS, Spanx and Vuori in the announcement, and added PayPal as an ACP-compliant payment server. For about six weeks it looked like in-chat checkout was going to be the next channel everyone had to be on.

Then March 2026 happened.

What OpenAI Dropped, and What Survived

In March 2026 OpenAI dropped its plans to operate checkout directly inside ChatGPT. Not because the technology failed. Because shoppers voted with their behavior: they used the assistant to research, compare and decide, and then went to finish the purchase somewhere they already had an account, a saved card and an order history.

Walmart’s numbers put a hard edge on it. Checkout inside the assistant converted roughly three times worse than sending the shopper through to walmart.com. And yet the assistant delivered around twice the rate of new customers Walmart sees from search engines. Both of those facts are true at once, and together they explain the whole pivot.

The buy button inside the chat is the part that stalled. Being named in the answer is the part that worked, and being named is decided entirely by data you already own.

So the protocol did not die. It shifted. ACP continues as the standard, with checkout moving toward merchant experiences rather than purchases made straight out of a product listing in a chat reply. What that leaves, for the overwhelming majority of stores, is the discovery half. And discovery has no engineering project attached to it. It has a feed.

What ACP Actually Asks of Your Product Data

Here is the merchant-facing reality underneath the specification. To participate properly, a business builds a product feed endpoint that conforms to the spec and refreshes on a tight cycle, implements the checkout REST endpoints, and integrates payment with shared payment tokens. OpenAI runs conformance tests across feed validation, the checkout flow and payment processing before anything goes live.

Most merchants read that and stop at “REST endpoints.” Understandable, and mostly beside the point now. The requirement that applies to you whether or not you ever touch an endpoint is the first one: the feed, and the state of the data inside it.

AI feeds are less forgiving than the Google Shopping feed you are used to. A product missing a brand, carrying a missing or invalid GTIN, or holding a price a parser cannot read does not rank lower. It gets dropped from the feed. There is no page two to fall to and no disapproval email to warn you. The product simply is not in the set the assistant is choosing from, and you find out by never hearing about it.

Freshness compounds the problem. An agent about to recommend a product wants current price and current stock, because recommending something that turns out to be sold out is the failure mode these systems are built to avoid. A feed that regenerates once a day was fine when a human was going to click through and see the truth for themselves. It is slow now.

Why This Is a Data Problem, Not an Integration Problem

Most stores are not one API away from AI visibility. They are one catalog away.

Most product feeds score somewhere between 25 and 40 out of 100 on AI-readiness, because the default export from a store platform carries a title, a price, an image and very little else. Color blank. Material blank. Size blank. No structured highlights, no answers to the obvious pre-purchase questions. A shopper on your product page fills those gaps from the photos and the copy. An agent cannot. It reads fields, and an empty field is a question you did not answer.

None of that is the merchant’s fault. Nobody handed most store owners a tool for this, and the platforms only started asking for the richer fields recently. But the gap is measurable, and it is the difference between a catalog an agent can evaluate and one it skips.

This is the entire reason UCP Radar exists. It connects to your Merchant Center account, scans your catalog against more than 50 validation rules, and scores every product for Google readiness and AI readiness separately, so you can see which products would survive an AI feed and which would be dropped. For ChatGPT specifically it reports the share of your catalog that is actually eligible along with a ranked list of what is blocking the rest, which is usually a short list of the same few offenders: no brand, a broken identifier, a price field that will not parse.

Then it fixes them. The AI optimizer rewrites thin titles into descriptive ones, fills in missing attributes like color, material and size, and generates the structured highlights, product details and question-and-answer content that AI surfaces read. Our canonical example is a product that arrived as “Blue Shirt Men” with a UCP Score of 28 and left as “Premium Sapphire Blue 100% Cotton Men’s Casual Dress Shirt, Breathable & Lightweight” with color and material filled, seven tags attached, and a score of 92. That is one product. The point of the platform is that it does it to thousands of them without you writing a line of it, in eight languages, and regenerates the output as your catalog changes.

Two things it deliberately does not do, because the distinction matters: it does not run agentic checkout or process payments, and it does not submit anything to OpenAI or Google on your behalf. It produces the optimized data and hosts the feeds. You stay in control of what gets registered where.

Getting Your Catalog Ready Without Doing It by Hand

The order of operations is not complicated, which is what makes it frustrating that so few catalogs are in shape.

Clear your Merchant Center disapprovals first, because they are already costing you Shopping impressions and they will disqualify products downstream. Fix identifiers next, since a missing or invalid GTIN is the single most common reason a product falls out of an AI feed. Then rewrite titles so they lead with brand, product type and the attributes a buyer would actually say out loud, because the title is the closest thing an assistant has to a summary of what the product is. Fill in the required and recommended attributes. Add the conversational content that answers pre-purchase questions in the feed itself. Then keep price and availability current instead of stale.

Doing that across a few thousand SKUs by hand is where every merchant stalls, and it is why supplemental feeds became the practical mechanism: a layer that merges enriched fields onto your primary feed by product ID, so your work survives a platform update and you never have to rebuild your store’s export. UCP Radar generates that layer for Google and a separate complete feed in the format ChatGPT expects, from the same enriched catalog. One pass of work, several surfaces, no manual maintenance.

Worth being blunt about the limits: nobody can promise you a ranking, and any tool that does is selling you something. What you can control is whether your products are eligible to be considered at all, and whether the data an assistant reads makes a convincing case. That part is fully in your hands, and it is measurable before and after.

Conclusion

ACP is worth understanding precisely because of how its first year went. The ambitious half, buying inside the chat window, ran into a shopper preference nobody had priced in. The quiet half, feeding structured product data to an assistant that decides what to recommend, kept working and now carries the whole value of the channel. “Discover in AI, buy on site” is not a downgrade for merchants. It is a cheaper deal, because it asks for good data rather than an engineering project.

Which puts every store in the same position it was in for agentic commerce generally, and for ChatGPT Shopping specifically, and for Google Shopping before either existed: the product feed is the storefront the machines actually see. UCP Radar exists to make that storefront complete without you maintaining it by hand. You can score your feed free and see the before and after on your own products before deciding anything. The protocols will keep changing. The requirement underneath them has not moved once.