What Is the Agentic Commerce Protocol? A Shopify Merchant's Honest Guide to ACP

TL;DR: The Agentic Commerce Protocol (ACP) is an open standard from OpenAI and Stripe, open-sourced under Apache 2.0 (spec version 2026-04-17), that lets a buyer's AI agent complete a purchase with a merchant. It defines five REST checkout endpoints (create, update, get state, complete, cancel) plus Stripe's Shared Payment Token, and the merchant stays the merchant of record. The headline "buy inside ChatGPT" story stumbled and was scaled back around March 2026, so the durable merchant move is not to chase Instant Checkout but to get agent-ready: an accurate, structured catalog and strong AI visibility, so any buying agent can find, trust, and transact your products.

The Agentic Commerce Protocol is an open standard, co-developed and maintained by OpenAI and Stripe, that defines how an AI agent buys something from a merchant on a shopper's behalf. It standardizes the checkout conversation (a small set of REST endpoints) and the payment step (a scoped token), while leaving the merchant in full control of the sale. If you run a Shopify store and you have heard about buying products inside ChatGPT, ACP is the plumbing underneath that idea. This guide explains what ACP actually is, how the flow works, its honest current state, and the one move that pays off for merchants regardless of how the standards war shakes out.

What ACP Is (and Is Not)

ACP is an interaction model that connects three parties: a buyer, the buyer's AI agent, and a merchant. It was announced around September 2025 alongside OpenAI's "Buy it in ChatGPT" / Instant Checkout launch, with Etsy first live and select Shopify merchants (Glossier, SKIMS, Spanx, Vuori) following. Crucially, OpenAI and Stripe then open-sourced it, licensed Apache 2.0, at github.com/agentic-commerce-protocol. The project uses date-based versioning, and the current spec version tag is 2026-04-17 (still beta). Because it is a genuinely open spec rather than a proprietary OpenAI feature, any agent or platform can implement it, which is the whole point.

The spec is published as two OpenAPI documents: openapi.agentic_checkout.yaml for the checkout flow, and openapi.delegate_payment.yaml for payment delegation. You can read the details on OpenAI's commerce docs at developers.openai.com/commerce and Stripe's announcement at stripe.com/newsroom.

What ACP is not matters just as much. It is not a payment processor, not a new checkout you outsource, and not a replacement for your Shopify backend. It does not take over pricing, inventory, fulfillment, returns, or support. It is a shared language for the moment an agent wants to complete a purchase, and nothing more.

How the ACP Checkout Flow Works

At the core, ACP defines a small set of REST endpoints that the buyer's agent calls on the merchant. Each call returns JSON containing the full checkout state (line items, totals, shipping options), so the agent always knows exactly where the order stands. Here are the five endpoints and what each one does.

Endpoint	Purpose
Create checkout session	The agent opens a session for the item(s) the shopper wants, and the merchant returns the initial checkout state with line items and totals.
Update session	The agent submits changes such as shipping address, shipping method, or variant selection, and the merchant returns the recalculated state.
Get session state	The agent reads the current state of the session (totals, availability, shipping) to confirm details before completing.
Complete purchase	The agent finalizes the order and payment is settled; the merchant accepts or declines and charges through its own provider.
Cancel	The agent (or shopper) abandons the session, and the merchant tears it down cleanly.

Two design choices make ACP safe for merchants. The first is payment delegation via Stripe's Shared Payment Token (SPT). This is a new primitive that lets an app like ChatGPT initiate a payment without exposing the buyer's card credentials. The token is scoped and authorized to a specific amount and a specific merchant, with the user's permission, and only the minimum data needed to fulfill the order is shared. The second is that the merchant stays the merchant of record: you accept or decline the order, charge through your existing payment provider, and keep ownership of fulfillment, returns, support, and the customer relationship. ACP standardizes the handshake; it does not reach into your business.

The merchant's inputs are correspondingly small: a structured product feed or catalog so the agent can surface the right items, plus an implementation of the checkout endpoints (which you can write yourself, or get from a platform or Stripe reference implementation). The honest hard part is the catalog. Accurate price, availability, and variant data is exactly what tripped up the early rollouts, which brings us to the state of play.

It is worth sitting with why the feed is the bottleneck. An agent completing a purchase cannot tolerate ambiguity the way a human browser can. If a price is stale, a variant is mislabeled, or a "backorder" state is not machine-readable, the agent either surfaces the wrong item or abandons the checkout, and the merchant never sees the sale. The endpoints are the easy part; a platform can generate them. The durable competitive edge is a catalog whose facts are correct, complete, and legible to a machine. That is not a one-time integration task, it is an ongoing data-hygiene discipline, and it is the same discipline that wins you AI-search visibility.

The Honest Current State: The Hype Cooled, the Shift Did Not

This is where most articles overclaim, so read carefully. OpenAI launched Instant Checkout, then narrowed and scaled it back around March 2026. Fewer than roughly thirty Shopify merchants had gone live, and OpenAI ran into real problems with accurate product data, multi-item carts, and onboarding merchants at scale. So the "one-click buy inside ChatGPT for a million stores" narrative did not play out the way the launch implied.

What is true as of mid-2026: products still appear inside ChatGPT, but most purchases now complete on the merchant's own storefront (in an in-app browser or a new tab) rather than fully inside the chat. Shopify has signalled that "agentic storefronts" are coming, which points at where this is heading rather than where it is today. The protocol and the broader shift toward agent-mediated shopping are real and strategic. The specific consumer feature was early and got trimmed.

The strategic read for a merchant is therefore not "rush to integrate Instant Checkout before your competitors." It is: the durable move is agent-readiness. Whichever checkout standard wins, and whichever agent your buyer uses, the prerequisite is the same. An agent can only buy what it can first find, understand, and trust.

The "agentic storefronts" language from Shopify is a useful tell about direction. It suggests the industry expects agents to increasingly deep-link into, or transact against, the merchant's own store surface rather than a walled feature inside one chat app. If that is where things land, the merchants who benefit are the ones whose storefront and catalog were already clean, structured, and easy for a machine to parse. Nothing about that preparation is wasted if the standards shift again, which is precisely why it is the safe bet in an unsettled space.

ACP in the Wider Stack

ACP is one layer in a fast-forming stack, not the whole thing, and it helps to know the neighbors so you do not conflate them. In short: MCP (Model Context Protocol, from Anthropic) is how an agent reads data and calls tools; A2A (Agent2Agent) is how agents talk to each other; ACP (OpenAI and Stripe) is how an agent buys; and AP2 (Agent Payments Protocol, from Google, with 60+ launch partners) is how a buyer's intent and a merchant's charge get cryptographically authorized. These are designed to compose, and a single real purchase can touch several of them. This is a "stack and camps" story, not a "one protocol kills the rest" story. For the full breakdown of how the pieces fit and where each one starts and stops, read ACP vs AP2 vs MCP: the agentic commerce stack.

What a Shopify Merchant Should Actually Do Now

Ignore the temptation to bet on a single standard. The winning position is to be agent-ready, which means being findable, trustworthy, and transactable to any buying agent (ChatGPT, Gemini, Perplexity, Copilot). Concretely, that comes down to two things you already control.

1. An accurate, structured catalog. Clean product data, complete structured data and Product JSON-LD (name, brand, price, availability, variants, identifiers), and an llms.txt that points AI crawlers at your best pages. This is precisely the data-quality problem that broke early ACP rollouts, so getting it right is the highest-leverage work available.
2. Strong AI-search visibility. Before an agent can transact you, an engine has to surface and recommend you. That is the domain of Generative Engine Optimization: earning citations and recommendations across the AI engines your buyers actually ask. If the engine never mentions your product, no checkout protocol will save the sale.

For a step-by-step version of this work tailored to the transition, see how to prepare your Shopify store for agentic commerce. The through-line is simple: agent-readiness is catalog readiness plus visibility, and both are things you can ship today without waiting for the standards to settle.

Where Naridon Fits (Honestly)

To be clear about the boundary: Naridon is not a payments or checkout solution. It does not implement ACP, AP2, or the Shared Payment Token for you, and it will not be your merchant of record. Those belong to Stripe, OpenAI, and your own payment provider. Naridon sits upstream of all of that, as the visibility and catalog-readiness layer that determines whether an agent ever reaches your checkout in the first place.

Naridon is a Shopify-native app that tracks whether AI engines cite and recommend your products across five engines (ChatGPT, Perplexity, Google AI Overviews, Gemini, and Copilot), and its Autopilot applies the fixes that make your catalog machine-readable, structured data and JSON-LD, product copy, and llms.txt, directly to your live store, with every change revertible. Nari, the in-app assistant, helps you decide what to ship next. That is exactly the agent-readiness the ACP transition rewards: an accurate, structured, findable catalog, verified against what the engines actually say.

You can start on the Free forever plan ($0, 150 credits per month) to see where your products stand across the engines, and move to Starter at $49/mo (3,000 credits) when you want Autopilot applying fixes on a cadence. Paid plans include a 7-day trial. Full details are on the pricing page.

---

The Bottom Line on ACP

ACP is a real, open, well-designed standard for how agents buy, backed by OpenAI and Stripe, and it keeps you the merchant of record. The consumer feature that launched it was early and got scaled back, but the shift toward agent-mediated shopping is not going away. The mistake is treating ACP as a checkout box to tick. The smart move is to be the store any agent can find, trust, and transact, whichever protocol wins.

Install Naridon on Shopify to track your visibility across five AI engines and let Autopilot apply the structured-data and llms.txt fixes that make your catalog agent-ready. Related reading: Generative Engine Optimization, structured data, the agentic commerce stack comparison, and preparing your Shopify store for agentic commerce.