ACP vs UCP vs AP2 vs MCP: which agentic commerce protocol matters?

They are layers, not rivals

Almost all coverage of these protocols frames them as a standards war with a winner to be picked. That framing is wrong in a way that leads merchants to the wrong decision, which is to wait.

They answer different questions. MCP answers "what can this agent access and call?" UCP answers "how does an agent read a catalog and build a cart?" ACP answers "how does a purchase execute inside an AI interface, and who processes it?" AP2 answers "how do we prove a human authorized this, in a way that survives a dispute?" A single transaction can touch all four.

What changed in 2026, and why it argues for the discovery layer

Two developments reset the picture this year, and both point the same way.

First, OpenAI shut down Instant Checkout and repositioned ACP toward product discovery and merchant-controlled experiences rather than in-chat purchase completion. Major retailers — Target, Sephora, Nordstrom, Lowe's, Best Buy, Home Depot, Wayfair — integrated ACP for discovery. The April 2026 revision covers checkout, payment delegation, cart, feed and authentication, plus a hook that lets ACP-aware agents advertise checkout capability through MCP tool discovery.

Second, Google donated AP2 v0.2 to the FIDO Alliance on 28 April 2026, moving payment authorization out of a single vendor's control and into a standards body with a long track record in authentication.

The read for a merchant: the checkout layer is still moving, the discovery layer has already consolidated, and everyone who tried to skip discovery and go straight to in-chat purchase has walked at least part of it back. Build the machine-readable catalog now. It is required by every one of these protocols and obsoleted by none of them.

The sequence that does not strand work

Given that the top of the stack is unsettled, the sane order of work is the one where each step is useful on its own and none of it is wasted if a protocol loses.

• 1. Machine-readable catalog. Server-rendered product data with schema.org Product, real price, real availability, real identifiers. Required by UCP and ACP, useful to every assistant today, and it improves classical SEO at the same time.
• 2. A discovery surface an agent can index. robots.txt that admits the assistant crawlers you want, a sitemap, and an llms.txt that states what you sell without making a model crawl 80 pages to work it out.
• 3. Programmatic cart and quote. An API that returns a binding price including shipping and tax, so an agent never has to drive a browser to find out what something costs.
• 4. Scoped, audited agent access. Before you let anything external act, you need per-key scopes, idempotency, and an audit log. This is where most platforms have nothing, and it is what turns agent access from a liability into something a finance team will sign off on.
• 5. Payment authorization. AP2-style verifiable mandates, once your processor supports them. Last because it is the layer still in motion and the one you least want to build twice.

Where PlatformDTC sits

PlatformDTC ships steps one and two as platform defaults rather than plugins — storefronts server-render structured product data, and the discovery layer is generated from the route tree so it cannot drift.

Steps three and four are the Agent Gateway: a scoped API where every agent-callable operation carries an idempotency key, an audit record and a permission scope tied to the calling key, and where any action that spends money stops at a human approval gate before it executes. The same operations are exposed as MCP tools, so an assistant can discover and call them directly.