▸ AI Visibility

The Agentic Commerce Protocol: what changes when ChatGPT and Perplexity can check out for your buyer

Agentic Commerce Protocol checkout handoff: an order passing from an AI assistant to a merchant via the standardised ACP message contract, illustrating how Stripe and OpenAI's open Apache-2.0 specification lets ChatGPT and Perplexity complete buyer checkout inside the conversation rather than handing off to a website.

▸ Bottom line up front

Stripe and OpenAI open-sourced the Agentic Commerce Protocol under Apache 2.0. It is the first cross-engine commerce specification that lets an AI assistant complete a checkout inside the conversation rather than handing the buyer off to a website. The protocol is live; the integration window for merchants who want to be cited as the buyable option in AI answers closes faster than the SEO window did.

What ACP actually is

The Agentic Commerce Protocol is an open specification, published on GitHub at github.com/stripe/agentic-commerce-protocol, licensed Apache 2.0. It defines the message contract between an AI assistant (the agent) and a merchant (the seller) for the steps that constitute a checkout: product discovery, inventory check, line-item assembly, address and payment handoff, order confirmation, post-purchase status.

Stripe shipped a reference merchant implementation. OpenAI shipped the first agent implementation inside ChatGPT. The combination is the proof that the protocol is real, not aspirational. It is also the explicit signal that the protocol is intended as cross-engine: any agent that speaks ACP can transact with any merchant that supports ACP, with the payment processor as the trusted intermediary.

What this is not: a Stripe-only payment rail. The protocol is payment-processor-agnostic in design. Merchants can implement against ACP and process payments through any compliant gateway.

Why this is the SEO moment compressed into one quarter

The SEO window opened in roughly 1998 and closed for incumbents around 2010. Brands that ranked in the early years compounded for a decade. The AI-visibility window has been narrower. The agentic-commerce window may be narrower still because the protocol is concrete and the implementation cost is measured in engineering weeks, not years.

Three structural reasons. First, the integration is binary. Either the merchant catalogue is queryable by an ACP agent or it is not. There is no partial credit. Second, the agent picks. When ChatGPT decides which payment terminal to recommend, it picks from the set of merchants that completed the integration. Merchants outside that set are not in the consideration set. Third, the citation effect compounds. Every successful ACP transaction generates a memory the agent can reference.

The merchant integration map: five endpoints

Five surfaces a merchant needs to ship. Catalogue endpoint for product discovery. Inventory check endpoint for real-time stock. Cart and quote endpoint for line-item assembly. Order endpoint for commit. Status webhook for post-purchase updates.

ACP merchant endpoints and operator latency budgets
EndpointDirectionTypical latency budgetNotes
GET /catalogueagent → merchantno real-time SLAfeeds rebuilt nightly or on inventory event
POST /inventory/checkagent → merchant400 to 800 msmissed budget = lost consideration
POST /quoteagent → merchant600 to 1200 msreturns TTL'd quote
POST /orderagent → merchant1 to 3 secondsidempotent on order_key
POST /webhooks/statusmerchant → agentas event firespaid, fulfilled, shipped, delivered, refunded

The buyer experience that ACP enables

Before ACP: buyer asks ChatGPT for a recommendation, gets a list of links, leaves the conversation, clicks a link, lands on a merchant site, fills a form, runs a card, completes the transaction. The agent sees none of it. This is the surface that ecommerce and retail brands have spent two decades optimising around.

With ACP: buyer asks ChatGPT for a recommendation, gets a contextual answer, says "buy that one", reviews the cart inline in the conversation, confirms with the stored payment method, sees the order confirmation. The merchant sees a clean ACP order with full provenance. The agent sees the outcome and learns.

The collapse from five surfaces (chat, search, click, browse, form) to one (chat) is the structural change. Every measurement built around the old surface set has to be re-pointed. CAC attribution that fires on website conversion will undercount ACP orders because there is no website conversion. The replacement metric is the ACP order webhook.

What to ship in the next 14 days

Two-week sprint plan for an experienced backend team. Each step is independently testable against the Stripe reference implementation.

  1. Day 1: read the spec and diff. Read the spec at github.com/stripe/agentic-commerce-protocol. Diff your existing product catalogue API against the ACP catalogue schema. Note the gaps.
  2. Days 2 to 4: catalogue + inventory endpoints. Stand up the catalogue endpoint. Most existing ecommerce backends can adapt their REST product feed within a few engineering days. Inventory check is the harder one because it has to be real-time; consider a Redis cache fronted by your inventory system.
  3. Days 5 to 9: quote + order endpoints. Implement the quote endpoint with TTL handling. Implement the order endpoint with idempotency. Wire to your existing fulfilment system.
  4. Days 10 to 12: status webhook + reference test. Implement the status webhook back to the agent. Test the full loop against the Stripe reference implementation.
  5. Days 13 to 14: submit credentials + monitoring. Submit your merchant credentials to the agents you want to be discoverable in. Add ACP-specific monitoring: order rate, quote-to-order conversion, status webhook delivery success.

The catalogue feed is doing the heavy lifting

Of the five endpoints, the catalogue is the one that decides whether the agent ever recommends you. An agent answering a buyer's recommendation query reads your catalogue feed, your competitors' catalogue feeds (where ACP-compliant), and any first-party product data the agent has from prior orders. The recommendation engine ranks against that combined set. Optimising for that ranking is a new front for AI performance marketing: the feed quality, spec density, and freshness signals that win an agent's pick are now as much a growth lever as a bid is on the open auction.

Product description quality. The agent's recommendation model reads the description as the primary signal for fit. Marketing copy with brand voice but vague specs loses to terse copy with concrete specs. Per-SKU geography and fulfilment lead-time. If the catalogue does not declare which countries the SKU ships to, the agent has to guess, and the safest guess is to drop the SKU from international recommendations. Stock state freshness. The catalogue should publish a stock-status timestamp per SKU. Agents in 2026 have learned to discount SKUs whose stock state is older than six hours during a sales push.

The new attribution surface

ACP changes what "a conversion" looks like in the warehouse. Before ACP, the conversion was a website event with a UTM tail. After ACP, the conversion is an order webhook with a structured provenance payload: agent identity, conversation context summary, the product set the agent considered before the buyer chose.

Three immediate consequences for the marketing stack. First, UTM-based attribution is irrelevant for ACP orders. Build a parallel attribution table keyed on the ACP order webhook. Second, the conversion path is shorter and richer at the same time. The provenance payload tells you which products you were compared against. Third, post-purchase comms shift. The agent handles status queries. Email follow-ups become for replenishment, complementary product suggestions, and support escalation. The shift maps cleanly onto an AI strategy brief: re-point the measurement plane, then re-point the offer plane. It also re-points the budget. Once ACP orders carry their own provenance, performance marketing can credit the channels that fed the agent's consideration set rather than the last UTM that happened to fire, which is the only honest basis for deciding where the next dollar goes.

Questions, answered.

Is ACP only for Stripe merchants?

No. The protocol is open and payment-processor-agnostic. Stripe published the reference merchant implementation and OpenAI published the reference agent implementation, but any merchant can implement ACP and process payments through any compliant gateway.

Which AI agents speak ACP today?

ChatGPT was the first agent to ship ACP support. The protocol is open so any AI assistant can implement against it; the merchant integration is the same regardless of which agent is calling.

How does ACP interact with existing ecommerce platforms?

ACP sits as a parallel commerce surface alongside the existing website. The merchant catalogue, inventory, quote, and order endpoints can be backed by the same systems that power the website; the surface the buyer interacts with is just the agent conversation instead of the web checkout.

What happens to attribution when an ACP order completes?

The website-conversion event does not fire because there is no website visit. Attribution shifts to the ACP order webhook, which carries the provenance: which agent referred, which conversation context, which product set was offered.

Does ACP require any new payment infrastructure on the merchant side?

No new payment processor is required. The merchant uses the same processor they use today. The protocol changes the way the agent and the merchant exchange order context; it does not change the way money moves.

Is there a sandbox to test against?

The reference implementations published by Stripe and OpenAI include test fixtures. Merchants stand up their endpoints against the reference agent, walk through the order lifecycle end to end, and verify the status webhook returns before submitting credentials to a production agent.