Agentic checkout explained: ACP, Instant Checkout and what merchants should do now

Agentic commerce splits into two halves that get discussed as one. Discovery is the half everyone writes about: how an AI assistant finds your products, which feeds it reads, what it quotes and why. We have covered that side at length, from what agentic commerce actually is to the seven-layer reference architecture a merchant can build against. Agentic checkout is the other half, the one that has had far less sober coverage: the moment the assistant stops recommending and starts transacting.

The definition worth keeping is short. Agentic checkout is an AI assistant completing a purchase on the shopper's behalf, with the shopper's money and the shopper's consent, without leaving the conversation. The shopper says "buy the medium in navy", confirms the total, and the order arrives in your order management system looking almost ordinary. No session on your site. No pixel fires. Your checkout ran, and nobody saw it.

Three parties, and the roles matter more than the technology. The shopper holds the intent and the payment method. The agent holds the conversation and brokers the transaction. You hold the catalogue, the inventory and the obligation: in every serious protocol on the table as of early 2026, the merchant remains the merchant of record. You take the payment, you fulfil, you handle the return when the navy turns out to be closer to black. The agent behaves less like a marketplace and more like a very capable browser being driven on someone else's behalf. We have watched this happen against our own store, before any protocol was involved; the play-by-play is in the write-up of the day a Perplexity agent bought from us.

How does the Agentic Commerce Protocol actually work?

ACP is the concrete spec behind the headlines. OpenAI built it with Stripe, announced it in September 2025, and open-sourced it under an Apache 2.0 licence at agenticcommerce.dev. It is the protocol powering Instant Checkout in ChatGPT, which launched with US Etsy sellers, with Shopify merchants announced as the next wave. A useful mental model: ACP is to agent purchases roughly what a payment gateway API was to card payments. A standard envelope, so that every merchant does not need a bespoke integration for every agent.

On the merchant side the protocol has four moving parts. A product feed tells the agent what you sell, at what price, in what stock state. A checkout API lets the agent open a checkout session, attach an address, get shipping options and tax, and confirm. A delegated payment spec lets the agent hand you a scoped payment token (Stripe's shared payment token is the reference implementation, though the spec is written to be processor-agnostic), which you charge through your normal payment stack. And webhooks keep the agent informed of order status, so the shopper can ask "where is my order?" in the same chat that placed it.

The commercial terms, as published at launch: the buyer pays the same price they would pay on your site, product results are ranked by relevance rather than paid placement, and OpenAI has said merchants pay a fee on completed purchases. You can decline any order the protocol sends you. The initial flow handled single-item purchases, with multi-item carts described as roadmap. All of that describes early 2026. The spec is young, and terms like these are the sort of thing that hardens over a couple of years, so treat the shape as stable and the details as provisional.

Where does Google's AP2 fit?

AP2, the Agent Payments Protocol, came out of Google in September 2025 with a long list of launch partners across payments and tech, Mastercard, American Express, PayPal and Adyen among them. It answers a different question. ACP standardises how an agent talks to your checkout. AP2 standardises how anyone can trust that the human actually authorised what the agent did.

Its core mechanism is the mandate: a cryptographically signed credential capturing what the shopper asked for and what they agreed to pay. An intent mandate for "find me running shoes under £120", a cart mandate for the specific basket at the specific price. If a dispute lands later, there is a verifiable trail showing the human said yes, which is precisely what card networks and banks need to see before they let agent-initiated payments scale. AP2 is designed to work alongside Google's A2A agent-to-agent protocol and MCP, and it is payment-method agnostic: cards, real-time bank transfers, stablecoins.

As of early 2026, that is the honest asymmetry between the two. ACP has a live consumer surface you can point at: checkout happening inside ChatGPT, today. AP2 is a specification with serious backers and early pilots, the direction of travel for the authorisation layer rather than something a merchant integrates this quarter. They are not competitors so much as different floors of the same building, and it is entirely plausible that a single transaction eventually touches both: an ACP-style checkout call carrying an AP2-style mandate.

What does a merchant integration actually involve?

Less than the whitepapers suggest, if you are on a major platform. Shopify and the payment providers are carrying the protocol plumbing, and for most merchants the near-term work is not writing checkout endpoints, it is making sure what flows through them is true. If you run headless or custom, budget an engineering sprint or two, most of it spent on the unglamorous edges of the checkout API: tax, shipping options, partial availability.

• Feed accuracy. The feed is the shelf. Price, variant, availability and imagery all render inside the agent's answer, and staleness anywhere becomes a wrong number quoted to a live shopper.
• The checkout session API. Create, update and complete a session; return shipping, tax and totals quickly and idempotently. Agents retry; your API must not double-charge or double-reserve.
• Payment delegation. Accept the scoped token through your PSP and treat it like a card-not-present charge with a tighter scope: one merchant, one amount, one window.
• Order operations. Webhooks outbound, and your normal fulfilment, returns and customer service inbound. Agent orders cannot become second-class orders; the shopper will ask the agent, and the agent will remember.

The feed deserves the most paranoia, because it fails silently. A stale price does not throw an error. It produces an agent quoting the wrong number, an order you have to decline, and an assistant that learns your data cannot be trusted. The plumbing for keeping feeds truthful, and exposing them to agents properly, is the subject of our merchant's guide to MCP tool calls and product feeds.

Where does social proof fit when the PDP is a chat answer?

Here is the part of this shift that gets the least coverage and touches our corner of the world the most. In a browser, social proof is presentation: a review band, a UGC gallery, a star row near the buy button, all working on a shopper who has already arrived at your page. In a chat answer there is no page to decorate. The agent shows one to three products, a sentence or two of justification each, and then the buy button. Whatever made the agent pick you has already happened, upstream, in the evidence it retrieved.

That evidence is disproportionately social proof. Ratings, review volume and recency, verified-purchase signals, customer photos and video tagged to specific products, answered questions. When the answer reads "this one, because owners with your use case rate it highly and the photos back it up", the review corpus did the selecting. UGC stops being conversion decoration and becomes retrieval fuel, which means it needs to be structured, attributed, rights-cleared and machine-readable, not just embedded in a widget. The convenient part: the evidence layer that earns citations in AI search is the same one that earns selection in agentic checkout, so the work compounds.

What should merchants do now?

A readiness map, because "integrate ACP" is the wrong first step for almost everyone. Agent-originated order volume is still small for most brands as of early 2026, and the case for moving early is not the volume. It is that nearly all of the preparation is work you should be doing anyway.

If you are below stage two, nothing about checkout protocols belongs on your roadmap yet; the feed and evidence work has all the leverage. If you are on Shopify, stages three and four are arriving mostly as configuration rather than engineering, which is exactly why the differentiator will not be having Instant Checkout. Everyone in your category will have it. The differentiator is being the product the agent picks, and that is an evidence problem, not a protocol problem.