Protocols · For Brands
OpenAI's Agentic Commerce Protocol (ACP), Explained for Brands
When OpenAI and Stripe shipped Instant Checkout in ChatGPT, they also handed the industry the rails underneath it: the Agentic Commerce Protocol. If you sell anything online, this is the spec that decides whether a shopper can buy from you without ever leaving the chat. Here is what it actually does and what you need on your side.
Start with the thing people can see
The visible part is Instant Checkout. A shopper asks ChatGPT for, say, a linen shirt under 4,000, gets a short list of real products, taps one, and buys it then and there. No new tab, no hunting for the cart, no re-typing a card number. OpenAI launched this for US shoppers buying from Etsy sellers first, with Shopify merchants following close behind, and it is the clearest signal yet that the AI assistant is becoming a place where money changes hands, not just a place where advice gets handed out.
ACP is the part you cannot see. It is the open standard, co-developed by OpenAI and Stripe and published under an Apache 2.0 license, that defines how a buyer, an AI agent, and a business talk to each other long enough to complete a purchase. Think of Instant Checkout as the storefront and ACP as the wiring in the wall. The storefront is what gets the headlines. The wiring is what you have to get right.
What ACP actually covers
You do not need to read the spec line by line to make good decisions about it. What matters is that ACP breaks the buying journey into a small number of moving parts. There are four worth knowing by name.
| Piece | What it handles | Why it matters to you |
|---|---|---|
| Feed | The structured description of every product you sell. | If a SKU is not described clearly here, the agent cannot recommend it. This is where most brands fall short. |
| Cart | Building and updating the order inside the conversation. | Quantities, variants, and totals stay live, so the price the shopper sees is the price they pay. |
| Orders | Placing the order against your real systems. | The sale lands in your existing store and fulfilment flow, not in some parallel universe. |
| Authentication | Confirming who the buyer is and authorizing payment. | Payment runs through a Stripe primitive (a shared payment token scoped to one merchant and one cart total), so the agent never sees the raw card details. |
That last point deserves a sentence on its own. The shared payment token is a neat piece of design: it lets ChatGPT trigger a payment without ever holding the buyer's payment credentials, and it is locked to a single merchant and a single cart total. For you, the practical takeaway is that the money side is handled by infrastructure you already trust, and you are not being asked to hand sensitive data to a chatbot.
It is an open standard, which is the whole point
Here is the part people miss when they assume this is an OpenAI walled garden. ACP is open source. If you do not process payments with Stripe, you can still adopt it with your existing payment provider. If you process with Stripe already, turning on agentic payments can be close to a one-line change. And because the protocol is a public spec rather than a private API, other AI agents can implement it too.
That last bit is the strategic insight. The work you do to be ACP-ready is not work for ChatGPT alone. It is work to be machine-readable for any agent that speaks the protocol. If you have read our piece on what agentic commerce is, this is the concrete version of the shift it describes.
What you actually need on your side
Strip away the protocol jargon and the merchant requirement comes down to two things. First, a feed that describes your catalog properly. Second, transaction endpoints that let the agent build a cart, place an order, and authenticate the buyer.
The feed is where the real effort lives, and it is the reason a lot of catalogs that look fine to a human fail quietly with an agent. A product listing written for a shopper skimming a page can lean on photos, vibe, and brand. An agent has none of that. It reads fields. So a single SKU needs anywhere from 15 to 40 structured fields depending on the category: material, dimensions, compatibility, certifications, care instructions, the things that let the agent answer a follow-up question without guessing.
And it cannot be a one-time export. Price and stock have to stay live. If your feed says a product is available at 3,499 and it is actually sold out, the agent either recommends something it cannot sell or learns to stop trusting your data. Both outcomes cost you. This is exactly the work covered in product feed optimization for AI search.
Why "set it and forget it" does not work here
The uncomfortable truth about every agentic commerce protocol, ACP included, is that the spec moves. New versions, new feed requirements, and new required fields show up regularly as the platforms learn what agents need. A field that was optional last quarter becomes required this quarter. A new attribute appears that your competitors populate and you do not.
You have two ways to deal with that. You can assign someone to track the spec, re-map your catalog every time it shifts, and re-publish, which is real, ongoing engineering work. Or you can sit behind infrastructure that absorbs the churn for you. This is the case for a managed layer rather than a one-off integration, and it is worth weighing honestly before you commit a team to maintaining feeds by hand.
Be ACP-ready without building it yourself
Ziffi connects your catalog and transaction APIs to ACP and keeps them current as the spec changes, without touching your existing store or checkout. The integration is free, most merchants are live within hours, and Ziffi earns only on the revenue it drives.
How ACP fits next to Google's UCP
ACP is not the only protocol in town. Google has its own, the Universal Commerce Protocol, aimed at Gemini and AI Mode. The two share a goal (buying inside the AI experience) but they are separate specs with separate feed expectations, and you will want to be present on both rather than betting on one. We go deep on the differences in the UCP explainer, but the short version is below.
| ACP (OpenAI) | UCP (Google) | |
|---|---|---|
| Main surface | ChatGPT | Gemini and AI Mode in Search |
| Built with | Stripe | Google, with broad retail and payment industry support |
| Feed basis | ACP product feed | Builds on Merchant Centre feeds plus conversational attributes |
| What you optimize | Structured fields, live price and stock, transaction endpoints | The same discipline, mapped to Google's format |
What to do this quarter
You do not have to solve everything at once, but the order matters. First, find out whether ChatGPT recommends you today and where the buy links point, which is the subject of AI visibility and share of voice. Second, audit your feed against what an agent needs, not what a human shopper tolerates. Third, decide whether you maintain the transaction plumbing in-house or let a commerce layer carry it. The brands that move now get a quiet head start while the surface is still uncrowded, and that window does not stay open forever.
Common questions
Is ChatGPT taking a cut of my sales?
Payment runs through Stripe's infrastructure, and commercial terms sit between you, your payment provider, and OpenAI. ACP itself is just the protocol. Nothing about being ACP-ready forces you to give up your customer relationship or your margin structure.
Does this work outside the US?
Instant Checkout rolled out in the US first, which is typical for a launch of this size. The protocol is open and global by design, so the practical question is less "is the spec available" and more "which surfaces are live in my market yet." That picture keeps widening.
What is the single biggest mistake to avoid?
Treating your feed as a marketing asset instead of a data asset. The agent does not read your brand voice. It reads your fields. Fix the fields first, and a lot of the rest follows. For the broader playbook, see answer engine optimization for ecommerce.