Agentic Commerce Protocol (ACP)

TL;DR — The Agentic Commerce Protocol (ACP) is an open standard from OpenAI and Stripe that lets supported AI commerce experiences shop for a person. One integration path uses a product feed you publish — a machine-readable list of your products, prices, and stock — and can build a cart and check out through it. If your feed is accurate and complete, your products may be eligible to show; wrong or stale data can disrupt discovery or checkout.

What ACP is, in plain terms

For years, getting foundA 302 (\"Found\") is a temporary redirect: it forwards users to a new URL while telling search engines the original URL should stay in the index. It's a weak canonicalization signal, not the zero-equity dead end of SEO folklore. online meant getting a person to your website. ACP changes the customer on the other end: it’s a set of rules for how an AI shoppingAI shopping optimization is the practice of making a merchant's products discoverable, recommendable, and buyable across AI shopping surfaces — ChatGPT, Google AI Mode, Gemini, Copilot, and Perplexity — by treating the structured product feed as a first-class optimization surface alongside the human-facing webpage, not a replacement for it. agent talks to a store’s systems to find products and buy them on a shopper’s behalf — without that shopper ever opening your site.

OpenAI and Stripe launched ACP on September 29, 2025. It’s free and open-source (anyone can read the spec on GitHub), and its first real-world use was ChatGPT’s “Instant Checkout” feature.

How a purchase actually happens

1. You type something like “find me a waterproof hiking jacket under $150” into ChatGPT.
2. The agent looks at product feeds from merchants that have implemented ACP and been approved to appear in that platform — enabling ACP technically and getting listed are two separate steps.
3. It picks products based on factors OpenAI’s documentation names — availability, price, quality, and whether the merchant turned on checkout — though OpenAI hasn’t published the full formula.
4. You confirm you want to buy. Your payment is passed along as a secure tokenA token is the smallest unit of text (or image/audio/video) an LLM processes — roughly 4 characters, or about ¾ of an English word. A context window is the maximum number of tokens (input plus output) a model can hold at once, like its short-term memory. — the agent never sees your raw card number.
5. The merchant gets the order, ships it, and stays the business you bought from.

The one big idea for store owners

An AI agent doesn’t read your web page — it reads your feed. All the usual SEO work (titles, page copy, links) still helps human visitors, but for AI shopping the thing that matters is the data: accurate prices, real-time stock, complete product details. A great product page with a broken or outdated feed is invisible to the agent.

There’s also a sister standard — Google and Shopify’s UCPThe Universal Commerce Protocol (UCP) is an open-source standard (Apache 2.0) led by Google and co-developed with Shopify, Etsy, Wayfair, Target, and Walmart that lets AI agents, merchants, and payment providers transact through a common language instead of bespoke integrations. Merchants advertise their capabilities at /.well-known/ucp. (Universal Commerce Protocol) — that does a similar job for Google’s AI Mode and Gemini. Most stores will want to support both over time.

Want the full picture — the five building blocks, the March 2026 pivot, and the exact data fields that matter? Switch to the Advanced tab.

TL;DR — ACP is an open (Apache 2.0) standard/protocol, current stable spec 2026-04-17, that defines how AI agents discover products and transact with merchants programmatically. OpenAI is the first AI platform to implement it and Stripe the first compatible payment provider — the spec doesn’t limit it to them. Five protocol-level building blocks: Cart and Feed, Agentic CheckoutAgentic checkout is the transaction-completion step of agentic commerce: the mechanism by which an AI agent creates, updates, and finalizes a purchase on a shopper's behalf — selecting fulfillment, calculating tax and shipping, passing a scoped payment token, and triggering order creation — often without the shopper visiting the merchant's site., Delegate Payment (Stripe’s implementation is Shared Payment TokensA token is the smallest unit of text (or image/audio/video) an LLM processes — roughly 4 characters, or about ¾ of an English word. A context window is the maximum number of tokens (input plus output) a model can hold at once, like its short-term memory. — a provider detail, not a universal requirement), Delegate Authentication (OAuth 2.0), and Orders/Webhooks. The merchant stays merchant-of-record. Implementing ACP doesn’t automatically list you in ChatGPT — that needs separate OpenAI merchant approval. ChatGPT’s Instant Checkout was the first implementation; industry reporting (not yet an official OpenAI changelog) describes a March 2026 pivot from full in-chat purchase to discovery-plus-redirectA redirect sends browsers and crawlers from a requested URL to a different one. An HTTP redirect specifically is a 3xx status code paired with a Location header; meta refresh and JavaScript redirects achieve a similar navigation without being a 3xx response themselves. Permanent redirects (301/308) are Google's signal the target should be canonical; temporary ones (302/303/307) aren't., but the feed and discovery layer are documented as live in the current stable spec either way. The SEO consequence is structural: ACP commerce flows query feeds and APIs, so feed freshness (at-least-daily per OpenAI’s current docs), data accuracy, complete identifiers, and checkout reliability become the optimization unit — reliability as a formal ranking factor is practitioner-reported, not documented — and conversions can land with no GA4 session attached. ACP and UCPThe Universal Commerce Protocol (UCP) is an open-source standard (Apache 2.0) led by Google and co-developed with Shopify, Etsy, Wayfair, Target, and Walmart that lets AI agents, merchants, and payment providers transact through a common language instead of bespoke integrations. Merchants advertise their capabilities at /.well-known/ucp. support different implementations and surfaces; verify actual availability before planning for either.

What ACP actually is

The Agentic Commerce Protocol is an open standard interaction model for connecting buyers, their AI agents, and businesses to complete purchases programmatically — without the buyer navigating a traditional website. It defines how agents browse product catalogs, manage carts, validate inventory, and execute checkout securely.

It launched September 29, 2025 from OpenAI and Stripe, who are the Founding Maintainers. It’s governed under an Apache 2.0 open-source license, lives on GitHub at agentic-commerce-protocol/agentic-commerce-protocol, and is currently in beta. The spec uses date-based versioning; the current stable release (2026-04-17) folded cart, feed, orders, authentication, and MCPMCP is an open protocol, created and open-sourced by Anthropic in November 2024, that standardizes how AI applications connect to external tools and data at runtime. It runs on a host–client–server architecture and lets agents call functions, read data, and take actions — the opposite of a static file like llms.txt. compatibility into the core.

The most important framing in the whole topic: ACP is the standard; Instant Checkout was its first implementation. Don’t conflate the protocol with the ChatGPT feature built on top of it — they have different lifespans, as the March 2026 pivot below shows.

The five building blocks

Per the ACP specification and Stripe’s implementation docs, ACP is organized into five capabilities. These are protocol-level primitives — what any ACP-compatible platform and payment provider implement — not a single company’s product. OpenAI is the first AI platform to implement ACP and Stripe is the first compatible payment provider; the spec is written so other platforms and payment providers can implement the same capabilities.

1. Cart and Feed (the Product Feed). The agent-facing product discovery layer. Agents browse your catalog and manage carts before checkout. This is where your merchandise becomes machine-readable: feeds in JSON, CSV, TSV, or XML, ingested by the agent platform. OpenAI’s current file-upload documentation supports at-least-daily full feed snapshots — freshness is part of the contract, and you should confirm the cadence your specific integration requires rather than assume a fixed number.
2. Agentic CheckoutAgentic checkout is the transaction-completion step of agentic commerce: the mechanism by which an AI agent creates, updates, and finalizes a purchase on a shopper's behalf — selecting fulfillment, calculating tax and shipping, passing a scoped payment token, and triggering order creation — often without the shopper visiting the merchant's site.. Create, update, and complete checkout sessions: cart management, fulfillment options, tax/shipping calculation, and payment processing. The merchant returns the authoritative cart state.
3. Delegate Payment. Securely pass payment credentials between buyer, agent, and business using payment handlers. Stripe’s implementation of this block is Shared Payment Tokens (SPTs), scoped to a specific merchant and transaction amount, with raw card data never exposed to the agent — but SPTs are Stripe’s implementation, not a universal feature every ACP payment provider must use.
4. Delegate Authentication. Authorization via OAuth 2.0, letting an agent act on a buyer’s behalf with a business.
5. Orders and Webhooks. Track the order lifecycle — confirmation, shipping, delivery, refunds — synced back to the agent and buyer.

One more distinction that trips people up: implementing these capabilities is not the same as being visible in ChatGPT. ACP is an open specification — you can implement it against the spec on your own. Actually showing up as a product source in ChatGPT requires OpenAI’s merchant participation process on top of that; technical conformance and platform listing are separate gates.

The end-to-end flow

Putting the blocks in motion:

1. A user expresses purchase intent in a conversational AI (e.g., ChatGPT).
2. The agent queries the merchant’s product feed/catalog via ACP-defined endpoints.
3. It ranks and surfaces products based on factors OpenAI’s documentation names — availability, price, quality, whether the merchant is the maker or primary seller, and whether checkout is enabled — though OpenAI has not published a full ranking algorithm, so treat this as the documented inputs, not a guarantee of how any specific product gets surfaced.
4. The user confirms; the agent initiates a checkout session.
5. The merchant validates inventory, taxes, and shipping and returns authoritative cart state.
6. A Shared Payment Token is generated, scoped to that merchant and amount.
7. The merchant receives the order, controls fulfillment, and remains merchant-of-record.
8. Webhooks sync the order lifecycle back to the agent and buyer.

OpenAI describes ACP as “the infrastructure between merchants and shoppers in ChatGPT.” Two things that flow from this are worth internalizing. First, the merchant keeps the relationship — unlike a marketplace, you control pricing, inventory, fulfillment, and the customer. Second, OpenAI has stated that results are not advertisements and aren’t influenced by partnerships; its documentation names data quality, price, and availability among the inputs, though it stops short of publishing a scored ranking algorithm — read “not ad spend” as the confirmed part and “checkout reliability as a ranking input” as the practitioner-reported part.

The March 2026 pivot — read this carefully

This is the single most misunderstood thing in current coverage, and I want to be precise about how solid the source is: I have not foundA 302 (\"Found\") is a temporary redirect: it forwards users to a new URL while telling search engines the original URL should stay in the index. It's a weak canonicalization signal, not the zero-equity dead end of SEO folklore. an official OpenAI changelog announcing this change. What follows is practitioner reporting (Klaviyo’s 2026 research is the source), not a confirmed OpenAI policy statement — treat it as “reported by industry coverage” until OpenAI publishes something dated on its own docs. With that caveat: reporting says that in March 2026, OpenAI shifted how Instant Checkout works, moving from full in-chat transaction completion to a discovery-plus-redirect model, where shoppers discover products in ChatGPT but complete the purchase on the retailer’s own website or app.

What this reported shift does not mean, even if accurate: ACP is dead, or the feed work was wasted. The ACP repository’s current stable spec (2026-04-17) still includes the checkout capability, and the product feed and discovery infrastructure remain part of the active protocol. If the in-chat purchase closure did roll back, that’s a platform-level product decision, not a change to the spec itself. If you optimize your feed for ACP today, you’re optimizing for capabilities that are documented as live in the current stable spec regardless of how any one platform’s UI currently behaves.

ACP and SEO: what’s structurally different

Here’s the part that matters for anyone who does this for a living. Product discovery is now an agent problem, not a page problem. An AI agent doesn’t crawl your HTML, read your <title>, or weigh your internal linksAn internal link is a hyperlink from one page on a website to another page on the same website. Internal links help search engines discover your pages and pass ranking signals (PageRank and anchor-text context) between them.. It queries your feed and your APIs and evaluates structured dataStructured data is a standardized way of labeling page content (using the schema.org vocabulary in JSON-LD, Microdata, or RDFa) so search engines can understand its meaning. It's not a direct ranking factor — its value is rich results and entity understanding.. The optimization lever moves upstream — from pages to feeds.

Concretely:

• Feed data quality is the new ranking input. Complete GTINProduct identifiers are the standardized values — GTIN (Global Trade Item Number), MPN (Manufacturer Part Number), and brand — that shopping feeds like Google Merchant Center and Microsoft Merchant Center use to match a product listing to the correct item in their catalog. A GTIN alone is usually enough; without one, brand + MPN is the fallback; products with none declare that with the identifier_exists attribute./MPN/brand identifiers, accurate real-time pricing, correct availability, consistent variant data, clear shipping and return policies. Schema.orgSchema markup is code that uses the schema.org vocabulary to label what your content means so search engines can understand it and show rich results. It's most often written in JSON-LD, and it's not a direct ranking factor. Product is the foundational vocabulary, and your feed must align exactly with your transactional backend — mismatches can get a product delisted.
• Feed freshness is not optional. OpenAI’s current file-upload documentation supports full feed snapshots at least daily; a stale feed isn’t just suboptimal, it’s a liability. A feed-to-API mismatch (the agent sees one price, checkout returns another) erodes trust fast. Confirm the exact cadence your integration supports rather than relying on any fixed number I could give you here — this is the kind of detail platforms update.
• Technical reliability looks like a signal, though it isn’t formally documented as one. OpenAI’s own documentation doesn’t publish a scored ranking algorithm or promise that reliability issues guarantee deprioritization. What I can say is that multiple practitioners (including SALT.agency’s technical guide) report that merchants with repeated feed errors, slow checkout responses, or inconsistent pricing see fewer products surfaced — read that as an observed pattern worth guarding against, not a documented policy you can cite to a client as official.
• Attribution breaks. A purchase can complete with no corresponding GA4 session — the revenue lands in your order system without a website visit. You don’t win because someone clicked your site; you win because the agent picked your product. Build AI agent traffic as its own channel with its own CAC and conversion benchmarks, and use UTM parameters on redirect flows where you can.

The framing the industry is settling on is “Agentic Commerce Optimization” — feed and data quality as evolved SEO. Traditional SEO still matters for human visitors; ACP adds a parallel, agent-facing track.

The protocol landscape: ACP, UCP, AP2, MCP

ACP isn’t alone, and conflating the protocols is a common mistake.

• ACP (OpenAI + Stripe) — checkout and product-feed standard, whose first and current primary consumer surface is ChatGPT. As an open spec it isn’t tied to ChatGPT by design, but treat “ACP means ChatGPT” as today’s reality, not a permanent architectural limit — verify actual support platform by platform rather than assuming exclusivity either way. Backed by Shopify, Etsy, Walmart, Instacart, Salesforce, BigCommerce, WooCommerce, commercetools.
• UCP — Universal Commerce Protocol (Google + Shopify, January 2026) — broader, covering the full commerce lifecycle across any agent surface, transport-agnostic (REST, MCP, A2A), powering Google AI Mode and Gemini. Merchants publish capabilities at a /.well-known/ucp endpoint — effectively a robots.txtA plain-text file at the root of a host that tells crawlers which URLs they may and may not request. It controls crawling, not indexing — a blocked URL can still be indexed if it's linked from elsewhere. for commerce agents, discoverable without a central gatekeeper.
• AP2 — Agent Payments Protocol (Google) — the payments layer specifically, using three signed “Mandates” (Intent, Cart, Payment) as W3C Verifiable Credentials; treats stablecoin rails as equal to cards.
• MCP — Model Context Protocol (Anthropic) — not a commerce protocol but a tool/context-access layer; ACP added MCP compatibility, and UCP supports it as a transport.

ACP and UCP are specifications built by different companies for different current surfaces, not documented as mutually exclusive — ACP’s current implementation reaches ChatGPT users, UCP reaches Google AI Mode and Gemini users, but that’s an observation about today’s implementations, not a normative rule of either spec. Most merchants should plan to implement both — each unlocks a different set of customer touchpoints today. (For the deep dive on the Google/Shopify side, see the sibling write-up on UCP.)

The economics, briefly

The protocol itself is free — the specification doesn’t define a fee. The fee sits with whichever payment provider processes the transaction. Stripe is reported to charge roughly 4% per transaction through its Agentic Commerce Suite (per Netalico’s analysis; I haven’t found this number published as a fixed rate on Stripe’s own docs, so confirm current pricing directly with Stripe before you quote it to a client). Frame that as a marketplace acquisition cost and compare it against blended paid-search acquisition (often 15–20% of AOV) — for high-intent, pre-qualified, agent-referred buyers, the math can be favorable, but this is a provider-specific number, not a protocol-level requirement. Stripe also describes its own integration for existing Stripe customers as requiring, in their words, as little as one line of code — that’s Stripe’s onboarding claim for its own product, not a property of the ACP spec.

What to do now

• Audit your product feed first. Completeness and accuracy of identifiers, pricing, availability, and variants — this is your visibility surface.
• Wire up freshness. Confirm and hit the update cadence your platform requires (OpenAI’s current file-upload docs support full snapshots at least daily) and make sure feed and checkout agree on price and stock.
• Harden checkout reliability. Slow or error-prone checkout correlates with fewer surfaced products in practitioner reporting, even though it isn’t a documented ranking policy — treat it as a real risk either way.
• Plan dual implementation. ACP for ChatGPT, UCP (with the /.well-known/ucp endpoint) for Google’s surfaces.
• Separate “implemented” from “listed.” Implementing ACP against the spec is a technical milestone; actually appearing as a product source in ChatGPT requires applying through and being approved by OpenAI’s merchant participation process. Budget time for that step separately from the engineering work.
• Build agent attribution. Treat AI-referred revenue as its own channel; don’t expect GA4 sessions to tell the story.

TIP Validate discovery and feed identity before enabling ACP

This offline readiness check does not authenticate, create a cart, submit payment, or certify ACP. It catches public discovery and sampled feed-shape problems before those live tests.

Compare discovery JSON with the feed URL and merchant data you expect using my free AI Commerce Validator Free

1. Paste the public discovery JSON and the expected feed URL.
2. Paste the XML or TSV feed so its sampled identity and required-field checks run separately.
3. Fix the mismatch, rerun, then continue with authenticated checkout, schema, crawler, policy, and lifecycle tests.

Tap to enlarge

The validator passes JSON parsing, endpoint declaration, required feed values, unique sampled IDs, absolute image URLs, and one sampled TSV row. It warns that discovery-products.tsv conflicts with the expected catalog.tsv feed URL and notes that checkout and adjacent checks remain outside scope.

AI summary

A condensed take on the Advanced version:

• ACP = open standard/protocol (Apache 2.0, current stable spec 2026-04-17), defining how AI agents discover products and transact with merchants programmatically — no website visit required. OpenAI is the first AI platform to implement it and Stripe is the first compatible payment provider, not the only ones the spec allows.
• Five building blocks (protocol-level primitives): Cart and Feed, Agentic Checkout, Delegate Payment (Stripe’s specific implementation is Shared Payment TokensA token is the smallest unit of text (or image/audio/video) an LLM processes — roughly 4 characters, or about ¾ of an English word. A context window is the maximum number of tokens (input plus output) a model can hold at once, like its short-term memory., scoped to merchant + amount — a provider implementation, not a universal requirement), Delegate Authentication (OAuth 2.0), Orders/Webhooks.
• Merchant stays merchant-of-record — keeps pricing, inventory, fulfillment, and the customer relationship. Results are not ads; OpenAI’s docs name availability, price, quality, maker/primary-seller status, and checkout enablement as inputs, but don’t publish a full ranking algorithm.
• Implementing ≠ listed. Implementing ACP against the spec doesn’t automatically get you into ChatGPT — that also requires applying through and being approved by OpenAI’s merchant participation process.
• March 2026 pivot (practitioner-reported, unconfirmed by an official OpenAI changelog): industry coverage (Klaviyo) describes OpenAI moving ChatGPT’s Instant Checkout from full in-chat purchase to discovery-plus-redirectA redirect sends browsers and crawlers from a requested URL to a different one. An HTTP redirect specifically is a 3xx status code paired with a Location header; meta refresh and JavaScript redirects achieve a similar navigation without being a 3xx response themselves. Permanent redirects (301/308) are Google's signal the target should be canonical; temporary ones (302/303/307) aren't.. The current stable spec still includes checkout, and the feed/discovery layer is documented as live either way.
• SEO consequence is structural: agents query feeds and APIs, not crawled HTML. Feed quality, freshness (OpenAI’s current docs support at-least-daily snapshots — confirm your platform’s actual cadence), complete GTINProduct identifiers are the standardized values — GTIN (Global Trade Item Number), MPN (Manufacturer Part Number), and brand — that shopping feeds like Google Merchant Center and Microsoft Merchant Center use to match a product listing to the correct item in their catalog. A GTIN alone is usually enough; without one, brand + MPN is the fallback; products with none declare that with the identifier_exists attribute./MPN/brand identifiers, real-time pricing/stock, and checkout reliability are the optimization unit. Reliability as a ranking factor is practitioner-reported, not formally documented. Conversions can land with no GA4 session — build agent attribution as its own channel.
• Landscape: ACP (currently ChatGPT) vs. UCPThe Universal Commerce Protocol (UCP) is an open-source standard (Apache 2.0) led by Google and co-developed with Shopify, Etsy, Wayfair, Target, and Walmart that lets AI agents, merchants, and payment providers transact through a common language instead of bespoke integrations. Merchants advertise their capabilities at /.well-known/ucp. (Google + Shopify; Google AI Mode/Gemini; /.well-known/ucp endpoint) vs. AP2 (Google payments layer) vs. MCPMCP is an open protocol, created and open-sourced by Anthropic in November 2024, that standardizes how AI applications connect to external tools and data at runtime. It runs on a host–client–server architecture and lets agents call functions, read data, and take actions — the opposite of a static file like llms.txt. (Anthropic, context layer). ACP and UCP aren’t documented as mutually exclusive — plan to support both.
• Economics: the spec itself doesn’t set a fee; Stripe is reported (Netalico) to charge roughly 4% per transaction through its own Agentic Commerce Suite — a provider number, not a protocol requirement. Confirm current pricing directly with Stripe.

Official documentation

Primary-source documentation for ACP and the surrounding ecosystem.

ACP — the spec

• agenticcommerce.dev — the official spec site (OpenAI + Stripe).
• ACP documentation / intro — the protocol overview and concepts.
• agentic-commerce-protocol on GitHub — the open-source spec, Apache 2.0, with date-based versioning.

Platform implementations

• OpenAI developer docs — Commerce — merchant integration: feeds, products, promotions; supported formats; current feed-cadence requirements (check the docs directly — this changes).
• Stripe — Agentic Commerce (ACP) — Stripe’s implementation docs, including the five building blocks and Shared Payment TokensA token is the smallest unit of text (or image/audio/video) an LLM processes — roughly 4 characters, or about ¾ of an English word. A context window is the maximum number of tokens (input plus output) a model can hold at once, like its short-term memory..
• Stripe launch press release (Sept 29, 2025) — the joint OpenAI/Stripe announcement.
• Stripe blog — Developing an open standard for agentic commerce — Stripe’s rationale.
• OpenAI — Buy it in ChatGPT — the launch announcement for Instant Checkout.

The neighboring protocols

• Google — Agentic commerce AI tools + UCP — Google’s UCPThe Universal Commerce Protocol (UCP) is an open-source standard (Apache 2.0) led by Google and co-developed with Shopify, Etsy, Wayfair, Target, and Walmart that lets AI agents, merchants, and payment providers transact through a common language instead of bespoke integrations. Merchants advertise their capabilities at /.well-known/ucp. announcement.
• Google — UCP developer docs — the Universal Commerce Protocol spec for the Google/Shopify side.
• Salesforce — Support for ACP (Oct 14, 2025) — Agentforce Commerce + ACP.

Quotes from the source

On-the-record statements from the companies behind ACP and its launch partners. Each link is a deep link that jumps to the quoted passage on the source page.

Stripe and OpenAI — the launch

• “Stripe is building the economic infrastructure for AI. We’re working alongside the most ambitious companies to create new AI-powered commerce experiences for billions of people, and building the tools businesses will need to thrive in a world where agent-led transactions are becoming the norm.” — Will Gaybrick, President of Technology and Business, Stripe. Jump to quote
• “By co-developing the Agentic Commerce Protocol with Stripe, we’re making it possible for businesses of all sizes to meet people where they are—and for shoppers to complete purchases seamlessly in conversation.” — Fidji Simo, CEO of Applications, OpenAI. Jump to quote
• “Stripe has spent the last 15 years optimizing commerce for human buyers. Now, we are starting to do the same for agents.” — Kevin Miller, Head of Payments, Stripe. Jump to quote

The platform partners

• “Through our collaboration with Stripe on the ACP, we are delivering the unified system designed for the future of agentic commerce…” — Nitin Mangtani, GM of Commerce Cloud and Retail, Salesforce. Jump to quote
• “BigCommerceBigCommerce SEO is the technical, on-page, and content work you do on a store built on BigCommerce — a hosted SaaS ecommerce platform that ships with more native SEO controls than most of its rivals (editable robots.txt, custom URL structures, auto sitemaps, and automatic 301s), while still leaving faceted navigation, multi-storefront hreflang, and review schema for you to handle. merchants will be able to unlock AI-driven discovery and checkout flows while continuing to use their existing catalogs, order systems, and operational processes.” — Sharon Gee, Senior VP of Product for AI, BigCommerce. Read the coverage

ACP-readiness checklist for merchants

A practical pass to make your store visible and reliable to AI shoppingAI shopping optimization is the practice of making a merchant's products discoverable, recommendable, and buyable across AI shopping surfaces — ChatGPT, Google AI Mode, Gemini, Copilot, and Perplexity — by treating the structured product feed as a first-class optimization surface alongside the human-facing webpage, not a replacement for it. agents.

Product feed and data

• Product feed published in a supported format (JSON, CSV, TSV, or XML).
• Complete identifiers on every product: GTINProduct identifiers are the standardized values — GTIN (Global Trade Item Number), MPN (Manufacturer Part Number), and brand — that shopping feeds like Google Merchant Center and Microsoft Merchant Center use to match a product listing to the correct item in their catalog. A GTIN alone is usually enough; without one, brand + MPN is the fallback; products with none declare that with the identifier_exists attribute., MPN, brand.
• Accurate, real-time pricing that matches your checkout exactly.
• Accurate, real-time availability/inventory — no phantom stock.
• Consistent variant data (size, color, SKU) with clear parent/child structure.
• Clear shipping options and return policy in the feed.
• Feed updated at least as often as your platform’s current documented minimum (OpenAI’s file-upload docs currently support at-least-daily full snapshots — confirm the live requirement rather than assuming a fixed number).
• Feed values reconcile against your transactional backend — no feed-to-API mismatches that risk delisting.

Structured dataStructured data is a standardized way of labeling page content (using the schema.org vocabulary in JSON-LD, Microdata, or RDFa) so search engines can understand its meaning. It's not a direct ranking factor — its value is rich results and entity understanding.

• Schema.orgSchema markup is code that uses the schema.org vocabulary to label what your content means so search engines can understand it and show rich results. It's most often written in JSON-LD, and it's not a direct ranking factor. Product markup on product pages: price, availability, reviews, shipping, returns.
• Identifiers in schema match the feed (GTIN/MPN/brand consistency).

Checkout, reliability, and participation

• Checkout/payment path responds quickly and consistently (practitioner reporting links slow or error-prone checkout to fewer surfaced products, even though this isn’t a documented OpenAI ranking policy).
• Instant Checkout / agent checkoutAgentic checkout is the transaction-completion step of agentic commerce: the mechanism by which an AI agent creates, updates, and finalizes a purchase on a shopper's behalf — selecting fulfillment, calculating tax and shipping, passing a scoped payment token, and triggering order creation — often without the shopper visiting the merchant's site. enabled with your platform or payment provider.
• Applied for and confirmed merchant participation/listing approval with the platform separately from technical integration — implementing ACP does not by itself get you into ChatGPT.
• Webhooks wired for the order lifecycle (confirmation, shipping, delivery, refunds).

Platform path (pick yours)

• Stripe merchants: enable ACP via the Agentic Commerce Suite (Stripe describes this as low-code for existing Stripe customers).
• Shopify / WooCommerceWooCommerce SEO is optimizing a WooCommerce store — which runs as a free plugin on WordPress, not as a standalone platform — to rank in organic search. Because you control the whole stack (templates, URLs, plugins, server), the ceiling is high and the ways to misconfigure it are many. / BigCommerceBigCommerce SEO is the technical, on-page, and content work you do on a store built on BigCommerce — a hosted SaaS ecommerce platform that ships with more native SEO controls than most of its rivals (editable robots.txt, custom URL structures, auto sitemaps, and automatic 301s), while still leaving faceted navigation, multi-storefront hreflang, and review schema for you to handle. / Salesforce: enable via your platform’s ACP integration; keep existing catalog and order systems.

Reach and attribution

• UCPThe Universal Commerce Protocol (UCP) is an open-source standard (Apache 2.0) led by Google and co-developed with Shopify, Etsy, Wayfair, Target, and Walmart that lets AI agents, merchants, and payment providers transact through a common language instead of bespoke integrations. Merchants advertise their capabilities at /.well-known/ucp. planned alongside ACP — publish a /.well-known/ucp endpoint for Google AI Mode and Gemini.
• AI agent revenue tracked as its own channel (UTMs on redirectA redirect sends browsers and crawlers from a requested URL to a different one. An HTTP redirect specifically is a 3xx status code paired with a Location header; meta refresh and JavaScript redirects achieve a similar navigation without being a 3xx response themselves. Permanent redirects (301/308) are Google's signal the target should be canonical; temporary ones (302/303/307) aren't. flows; don’t rely on GA4 sessions).

The mental models

1. Pages are for people; feeds are for agents. Traditional SEO optimizes a page a human reads. Agentic commerce optimizes a feed a machine parses. Both can be true at once — but for AI shoppingAI shopping optimization is the practice of making a merchant's products discoverable, recommendable, and buyable across AI shopping surfaces — ChatGPT, Google AI Mode, Gemini, Copilot, and Perplexity — by treating the structured product feed as a first-class optimization surface alongside the human-facing webpage, not a replacement for it., the data is the product page. A great page with a broken feed is invisible.

2. ACP vs. UCPThe Universal Commerce Protocol (UCP) is an open-source standard (Apache 2.0) led by Google and co-developed with Shopify, Etsy, Wayfair, Target, and Walmart that lets AI agents, merchants, and payment providers transact through a common language instead of bespoke integrations. Merchants advertise their capabilities at /.well-known/ucp. vs. AP2 vs. MCPMCP is an open protocol, created and open-sourced by Anthropic in November 2024, that standardizes how AI applications connect to external tools and data at runtime. It runs on a host–client–server architecture and lets agents call functions, read data, and take actions — the opposite of a static file like llms.txt. — keep them straight.

ACP and UCP are the two you act on, and they aren’t documented as mutually exclusive — support both to reach both audiences.

3. The documented ranking inputs are data quality; reliability is reported, not confirmed. For agent discovery, OpenAI’s documentation names availability, price, quality, maker/primary-seller status, and checkout enablement as inputs — but it doesn’t publish a scored ranking algorithm. Practitioner reporting adds that the reliability of your feed and checkout correlates with visibility: repeated errors, stale data, or inconsistent pricing appear to push you down. Treat the documented inputs as fact and the reliability signal as a well-observed pattern, and keep those two categories separate when you brief a client. Either way, this is closer to data engineering than to copywriting.

4. The merchant-of-record line. Unlike a marketplace, ACP leaves you in control of pricing, inventory, fulfillment, and the customer. That’s the strategic difference: you gain a new acquisition surface without surrendering the relationship.

5. Protocol vs. implementation vs. listing. ACP (the standard) outlives Instant Checkout (the first feature on it). Industry reporting says OpenAI changed Instant Checkout in March 2026 — I haven’t foundA 302 (\"Found\") is a temporary redirect: it forwards users to a new URL while telling search engines the original URL should stay in the index. It's a weak canonicalization signal, not the zero-equity dead end of SEO folklore. an official OpenAI changelog confirming it, so hold that loosely — but the current stable spec’s discovery layer and checkout capability are documented as live either way. A third layer sits on top of both: implementing the spec doesn’t automatically get your products listed on a platform; that’s a separate merchant-approval gate. Optimize for the standard, not for any single feature’s current behavior, and budget for the approval step separately from the engineering work.

6. Attribution is upstream of analytics. A win can occur with no GA4 session. Decide before you measure: AI agent traffic is its own channel with its own CAC and conversion benchmarks. You win when the agent picks your product, not when someone clicks your site.

ACP mistakes that derail implementation

Treating Instant Checkout and ACP as the same thing

Why it fails: a platform feature can change while the open protocol and feed infrastructure remain active. Do instead: design against the protocol capabilities and current spec version, then treat each consumer surface as an implementation.

Optimizing only the product page

Why it fails: ACP discovery reads product feeds and APIs, so excellent HTML cannot repair missing identifiers, stale stock, or a feed/checkout mismatch. Do instead: make feed completeness, freshness, and backend reconciliation first-class SEO work.

Enabling checkout before feed accuracy is stable

Why it fails: a reliable transaction layer cannot sell a product whose price or availability changes between discovery and checkout. Do instead: prove price, stock, variant, and policy parity before enabling purchase eligibility.

Assuming ACP replaces UCP

Why it fails: the protocols reach different agent ecosystems and are designed to coexist. Do instead: use one product, inventory, and payment source of truth with separate ACP and UCPThe Universal Commerce Protocol (UCP) is an open-source standard (Apache 2.0) led by Google and co-developed with Shopify, Etsy, Wayfair, Target, and Walmart that lets AI agents, merchants, and payment providers transact through a common language instead of bespoke integrations. Merchants advertise their capabilities at /.well-known/ucp. adapters where both audiences matter.

Measuring success only through GA4 sessions

Why it fails: an agent-originated purchase can reach the order system without a website visit. Do instead: tag the order or checkout sessionAgentic checkout is the transaction-completion step of agentic commerce: the mechanism by which an AI agent creates, updates, and finalizes a purchase on a shopper's behalf — selecting fulfillment, calculating tax and shipping, passing a scoped payment token, and triggering order creation — often without the shopper visiting the merchant's site. at the API/OMS layer and reconcile agent-driven revenue there.

ACP implementation cheat sheet

Delegate Payment’s specific token implementation (Stripe’s Shared Payment Tokens) is Stripe’s product, not a protocol-mandated format — other PSPs can implement the capability differently.

Data and reliability rules

• Agents use the feed/API layer; page SEO remains a parallel human-facing track.
• Keep item_id, GTINProduct identifiers are the standardized values — GTIN (Global Trade Item Number), MPN (Manufacturer Part Number), and brand — that shopping feeds like Google Merchant Center and Microsoft Merchant Center use to match a product listing to the correct item in their catalog. A GTIN alone is usually enough; without one, brand + MPN is the fallback; products with none declare that with the identifier_exists attribute./MPN/brand, variants, price, and availability consistent.
• Reconcile feed values against checkout before enabling agent purchase.
• Merchant remains merchant-of-record and authoritative for the order.
• Track agent origin at checkout or order level because a GA4 session may not exist.
• Implementing the spec is not the same as being listed — confirm merchant participation/approval status with each platform separately.

Protocol map

ACP and UCP are complementary. Build shared backend truth first, then map it into the protocols you support.

Resources worth your time

My related writing

• The Beginner’s Guide to Ecommerce SEO — where agent-facing discovery fits into the bigger ecommerce picture.
• How to Earn LLM Citations to Build Traffic & Authority — the broader shift from pages to machine-readable answers.

My speaking

• How Search Works (SlideShare) — my walkthrough of discovery, indexingStoring a crawled page in the search index so it can appear in results. Crawled is not the same as indexed — Google selects what to keep, and indexing isn't guaranteed., and ranking; useful background for how an agent-mediated layer sits on top. (My standing disclaimer applies: “This is my understanding of systems… not going to be 100% complete or accurate.”)

From around the industry

• Agentic Commerce: What SEOs Need To Consider (ACP & UCP) (Search Engine Journal) — the “agents query APIs and parse feeds, not pages” framing and the schema.orgSchema markup is code that uses the schema.org vocabulary to label what your content means so search engines can understand it and show rich results. It's most often written in JSON-LD, and it's not a direct ranking factor.-as-glue point.
• Optimising for Agentic Commerce Protocol — A Technical Guide for SEOs (SALT.agency, Sept 30, 2025) — the seven-stage end-to-end flow and OpenAI’s stated ranking criteria.
• Google’s UCP and ecommerce SEO (Search Engine Land) — the “revenue with no GA4 session” attribution problem.
• ACP vs UCP: The Two Protocols Reshaping How Customers Find and Buy Your Products (Netalico) — the ~4% fee economics and the /.well-known/ucp “robots.txtA plain-text file at the root of a host that tells crawlers which URLs they may and may not request. It controls crawling, not indexing — a blocked URL can still be indexed if it's linked from elsewhere. for commerce” point.
• Agentic Commerce Protocol explainer + merchant guide (BigCommerceBigCommerce SEO is the technical, on-page, and content work you do on a store built on BigCommerce — a hosted SaaS ecommerce platform that ships with more native SEO controls than most of its rivals (editable robots.txt, custom URL structures, auto sitemaps, and automatic 301s), while still leaving faceted navigation, multi-storefront hreflang, and review schema for you to handle.) — launch-partner context and the merchant operational view.
• Agentic Commerce Protocol (ACP) deep dive guide (commercetools) — a thorough, B2B-leaning technical explainer.
• What is the Agentic Commerce Protocol (Klaviyo) — the marketing/CRM angle and the March 2026 pivot context.
• Agentic commerce and AI shopping agents (Deloitte) — the macro view of where agent shopping is heading.

Stats worth citing

Numbers that frame how fast AI shoppingAI shopping optimization is the practice of making a merchant's products discoverable, recommendable, and buyable across AI shopping surfaces — ChatGPT, Google AI Mode, Gemini, Copilot, and Perplexity — by treating the structured product feed as a first-class optimization surface alongside the human-facing webpage, not a replacement for it. is moving. Treat merchant-reported and projection figures as directional, not gospel — verify before leaning on them.

• 39% of consumers purchased products based on AI recommendations within six months (Klaviyo, 2026 research). Klaviyo characterizes the shift as “the collapse of discovery, research, and sometimes purchase phases to one channel.” Source
• 38% higher purchase completion for AI-referred visitors vs. traditional search visitors, and roughly 7x better sales growth for retailers with AI agent integration during Cyber Week 2025 (Adobe Analytics, via multiple sources).
• ~$5B (2024) → ~$200B (2034) projected agentic AI market size, growing over 40% annually (BigCommerceBigCommerce SEO is the technical, on-page, and content work you do on a store built on BigCommerce — a hosted SaaS ecommerce platform that ships with more native SEO controls than most of its rivals (editable robots.txt, custom URL structures, auto sitemaps, and automatic 301s), while still leaving faceted navigation, multi-storefront hreflang, and review schema for you to handle., citing market research). Source
• ~4% per transaction is Stripe’s approximate fee through its Agentic Commerce Suite — framed as a marketplace acquisition cost against blended paid acquisition of 15–20% of AOV (Netalico). Source
• 400–600% YoY growth from ChatGPT and Gemini traffic reported by some early adopters — but from small baselines, so read it as momentum, not magnitude (Netalico, merchant-reported). Source
• Operational efficiency improvements of 30–40% reported by businesses from ACP integration (BigCommerce). Source

Test yourself: Agentic Commerce Protocol

Five quick questions on what ACP is and what it changes for ecommerce. Pick an answer for each, then check.

Agentic Commerce Protocol (ACP)

The Agentic Commerce Protocol (ACP) is an open standard/protocol whose first implementing AI platform is OpenAI and first compatible payment provider is Stripe, defining how AI shopping agents discover products, manage carts, and complete purchases on a buyer's behalf — by reading a merchant's product feed and APIs instead of crawling the website.

Related: Answer Engine Optimization (AEO), Schema Markup for AI

Agentic Commerce Protocol (ACP)

The Agentic Commerce Protocol (ACP) is an open standard/protocol launched September 29, 2025, current stable spec dated 2026-04-17, that defines how AI agents communicate with merchant systems to discover products and complete purchases on behalf of a buyer — without the buyer navigating a traditional website. It’s governed under an Apache 2.0 open-source license and hosted on GitHub at agentic-commerce-protocol/agentic-commerce-protocol. OpenAI is the first AI platform to implement ACP and Stripe is the first compatible payment provider — the spec itself isn’t limited to them, and implementing it does not by itself get a merchant listed on any given platform; that requires separate platform approval.

ACP covers five protocol-level building blocks: an agent-facing Product Feed (Cart and Feed) for discovery, Agentic CheckoutAgentic checkout is the transaction-completion step of agentic commerce: the mechanism by which an AI agent creates, updates, and finalizes a purchase on a shopper's behalf — selecting fulfillment, calculating tax and shipping, passing a scoped payment token, and triggering order creation — often without the shopper visiting the merchant's site. for cart and checkout sessions, Delegate Payment (Stripe’s implementation is Shared Payment TokensA token is the smallest unit of text (or image/audio/video) an LLM processes — roughly 4 characters, or about ¾ of an English word. A context window is the maximum number of tokens (input plus output) a model can hold at once, like its short-term memory., scoped to a merchant and amount — a provider detail, not a protocol-mandated format), Delegate Authentication (OAuth 2.0), and Orders and Webhooks for the order lifecycle. The merchant stays merchant-of-record and keeps their existing commerce backend.

For ecommerce SEOEcommerce SEO is the practice of optimizing an online store so its product and category pages rank in organic search and attract purchase-intent visitors. It uses the same Google algorithm as any other site, but compounds the usual SEO work with commerce-specific challenges like faceted navigation, product variants, and platform-imposed URLs., the shift is fundamental: an AI agent doesn’t browse your pages — it queries a machine-readable product feed and structured dataStructured data is a standardized way of labeling page content (using the schema.org vocabulary in JSON-LD, Microdata, or RDFa) so search engines can understand its meaning. It's not a direct ranking factor — its value is rich results and entity understanding.. That makes feed quality, accurate real-time inventory and pricing, complete product identifiersProduct identifiers are the standardized values — GTIN (Global Trade Item Number), MPN (Manufacturer Part Number), and brand — that shopping feeds like Google Merchant Center and Microsoft Merchant Center use to match a product listing to the correct item in their catalog. A GTIN alone is usually enough; without one, brand + MPN is the fallback; products with none declare that with the identifier_exists attribute. (GTIN/MPN/brand), and checkout reliability the new optimization unit — though reliability functioning as a ranking input is practitioner-reported, not documented OpenAI policy. ACP’s first and current primary consumer surface is ChatGPT; Google and Shopify’s UCPThe Universal Commerce Protocol (UCP) is an open-source standard (Apache 2.0) led by Google and co-developed with Shopify, Etsy, Wayfair, Target, and Walmart that lets AI agents, merchants, and payment providers transact through a common language instead of bespoke integrations. Merchants advertise their capabilities at /.well-known/ucp. covers the equivalent surface on Google AI Mode and Gemini, and most merchants will eventually support both. Industry reporting (not yet an official OpenAI changelog) describes ACP’s first implementation, ChatGPT’s Instant Checkout, shifting in March 2026 from full in-chat purchase to a discovery-plus-redirectA redirect sends browsers and crawlers from a requested URL to a different one. An HTTP redirect specifically is a 3xx status code paired with a Location header; meta refresh and JavaScript redirects achieve a similar navigation without being a 3xx response themselves. Permanent redirects (301/308) are Google's signal the target should be canonical; temporary ones (302/303/307) aren't. model; the feed and discovery capabilities remain part of the documented, active spec either way.

Related: Answer Engine Optimization (AEO), Schema Markup for AI

Changelog