Agentic commerce

UCP vs ACP vs MCP: which agentic-commerce rails matter for your business?

Three protocols now decide whether AI agents can find your business and pay you. Here is what each one actually does, who runs it, what it costs you to support, and the honest order to adopt them in.

Updated July 13, 2026 · 9 min read

If you sell anything online, three acronyms started crowding your feeds over the past year: ACP, UCP, and MCP. They sound interchangeable. They are not, and the differences decide which AI surfaces can discover your business, which can book or buy from you, and how much engineering each one demands.

The stakes stopped being theoretical a while ago. Salesforce reported that AI influenced $262 billion of online sales during the 2025 holiday season, and every major assistant now ships some form of shopping. This guide defines each protocol in plain English, compares them side by side, and ends with an honest decision guide.

Three protocols, three different jobs

Here is the shortest accurate version. ACP is OpenAI’s checkout protocol: it lets ChatGPT complete a purchase from your business without the buyer leaving the chat. UCP is Google’s commerce protocol: it connects your catalog and checkout to Gemini and Google’s AI shopping surfaces, onboarded through Merchant Center. MCP is the general-purpose tool protocol: it is how agents call any external capability at all, and commerce actions like searching listings or starting a booking are just one use of it.

A useful mental model: ACP and UCP are sales channels with a spec attached. MCP is plumbing. The first two answer “can ChatGPT or Gemini sell my stuff?” The third answers “can any agent, anywhere, take a real action against my business?” Confusing them is easy because all three landed within about fourteen months of each other, but they solve different problems and you evaluate them differently.

ACP: OpenAI’s checkout rail for ChatGPT

OpenAI and Stripe published the Agentic Commerce Protocol as an open spec in September 2025, and it is what powers Instant Checkout in ChatGPT. Etsy sellers were the launch cohort, with Shopify merchants following. The pitch to merchants is blunt and appealing: the buyer stays in ChatGPT, you stay the merchant of record, and the order lands in your existing systems.

Mechanically, ACP has two halves. Discovery runs on a product feed that OpenAI ingests so ChatGPT can surface your items in shopping answers. Checkout runs on a small REST API (checkout sessions) that you or your platform host, with payment handled through a delegated token so raw card details never touch your servers. The spec lives at developers.openai.com/commerce if you want the wire-level detail.

Status as of mid-2026: live, but gated. OpenAI revamped Instant Checkout in March 2026 toward a discovery-first flow, which quietly raised the value of simply being present in the feed even before your checkout enrollment clears. Enrollment itself is application-based and category-dependent. The full walkthrough, including what reviewers actually check, is in our ACP enrollment guide.

UCP: Google’s commerce protocol for Gemini and Search

Google introduced the Universal Commerce Protocol in early 2026 as the standard way its AI surfaces talk to merchants. The front door is Google Merchant Center: your catalog data flows in the way Shopping feeds always have, and UCP layers agent-facing capabilities on top so Gemini and AI Mode in Search can move from recommending a product to transacting for it.

The design choice worth noticing is that UCP is transport-flexible. The same commerce capabilities can be served over a plain REST API, over MCP, or over A2A, Google’s agent-to-agent protocol. Payments compose with AP2, the Agent Payments Protocol Google published in September 2025 with a long roster of payments partners; AP2 handles the “did a human actually authorize this purchase” mandate so an agent cannot spend without provable consent.

Status as of mid-2026: discovery is the live half. Getting your catalog legible to Gemini through Merchant Center works today, while agentic checkout is rolling out in stages to enrolled merchants. If you already maintain a Merchant Center feed for Shopping ads, you are much closer to UCP-ready than you probably think.

MCP: the protocol agents use to actually do things

The Model Context Protocol is the odd one out because it was never a commerce spec. Anthropic released it as an open standard in November 2024 to give AI models one uniform way to call external tools. OpenAI adopted it in March 2025, Google followed, and it is now the closest thing agents have to a USB port. The spec and ecosystem live at modelcontextprotocol.io.

For a merchant, MCP means exposing your business as callable tools: search my offers, check availability, book an appointment, start a checkout. Any MCP-capable agent can then act against your business directly instead of scraping your website and guessing. That includes Claude, ChatGPT’s developer mode, and the long tail of custom agents companies are building internally, which is a market no feed reaches.

There is also a distribution angle people miss. MCP servers are listed in the official MCP registry and mirrored across directories like Smithery and mcp.so, so a listed server is discoverable by agent builders in a way a private API never was. And because UCP accepts MCP as a transport, a well-built commerce MCP server is not a dead end; it is a component the bigger rails can reuse.

Side by side: who runs what, and what it takes

One caveat: “status” in this space has a shelf life measured in months. Treat the mid-2026 column as a snapshot, not a prediction.

Where they overlap, and where they compose

These are not three competitors fighting for one socket. UCP explicitly supports MCP as a transport, so the “Google protocol” can literally run over the “Anthropic protocol.” ACP’s product feed is a close cousin of the Merchant Center feed, close enough that one well-structured catalog generates both. And AP2 is a payments layer, not a rival channel; it can sit underneath more than one of these rails.

The deeper overlap is the substrate. All three assume your business data exists in structured, machine-readable form, and an agent that discovers you through any protocol still cross-checks your actual website before recommending you. Schema.org JSON-LD, clean crawlable HTML, and conventions like llms.txt remain the common denominator underneath every rail; we cover that layer in what llms.txt actually does.

In practice the protocols compose per buyer, not per merchant. One customer’s agent finds you through Gemini, another buys through ChatGPT, a third books through a custom concierge agent hitting your MCP server. Same catalog, three doors. The winners stopped asking which door and started keeping one catalog behind all of them.

Which should you support first? The honest answer

None of them, at first. The bulk of agent traffic in 2026 is still a model reading your web pages, so structured data and crawlability pay off before any protocol does, and every protocol either consumes that same catalog data or verifies against it. Be honest about the fads too: Ahrefs found no ranking correlation for llms.txt and Google’s guidance says it is not required, so treat that file as a cheap courtesy, not a strategy. The JSON-LD and clean markup underneath it are the part that compounds.

1. Fix the substrate. Every offer or service gets schema.org JSON-LD, a crawlable HTML page, and accurate prices and availability. Verify with a free agent-legibility scan rather than assuming.
2. Follow your buyers, not the hype cycle. If your customers live in ChatGPT, do the ACP feed first and start the enrollment process early. If you already run Merchant Center, UCP is your shortest path. If your buyers are businesses or agent builders, an MCP server reaches them all.
3. Turn on discovery before checkout, on every rail. Feeds and discovery are the cheap, ungated half; checkout enrollment can lag by months. Being findable while you wait costs almost nothing.
4. Watch your referrers and let data pick your second protocol. Agent traffic shows up in server logs and analytics in odd shapes; two months of real evidence beats any punditry, including this article.

If you run a service business, one more honest note: ACP and UCP grew up around physical products. Their feeds model SKUs, shipping, and returns, and bookable services fit awkwardly today. If you sell appointments, consults, or classes, an MCP server wired to real availability plus a normal Stripe checkout will beat waiting for the feed specs to catch up; we wrote up how that flow works end to end in how AI agents book service businesses.

The compounding move: one catalog, every rail

If maintaining a product feed, a checkout API, an MCP server, and structured markup sounds like a lot, that is the actual argument for a hosted layer. Nexez publishes each listing simultaneously as a human web page, JSON-LD, agent.json, llms.txt, an OpenAPI spec, and a per-merchant MCP server listed in the official registry, plus ACP and UCP feeds generated from the same money core. Discovery ships on every plan; agentic checkout is the Pro upgrade, with feed enrollment handled for you as the platforms open their gates.

You keep your own website and your own Stripe account as merchant of record. The protocols become output formats of one catalog instead of four separate projects, which is exactly the shape this market rewards while the specs keep moving.