ACP vs UCP: Choosing an Agentic Commerce Protocol

Freshness check (25 September 2026)

Both standards are young and moving. ACP was released September 29, 2025 by OpenAI and Stripe under Apache 2.0 and the GitHub repo still marks it beta. The checkout OpenAPI snapshot 2026-01-30 defines five merchant endpoints (create, update, retrieve, complete, cancel). A later tree, 2026-04-17, adds feeds, carts, orders, authentication, and MCP. OpenAI’s checkout guide says Instant Checkout in ChatGPT is approved partners; the same company’s product-discovery page says merchants can keep their own checkout while ACP carries catalog and promotions. UCP launched in early 2026 led by Google with Shopify, Etsy and Walmart. Google’s Merchant Center help (reviewed 25 September 2026) limits UCP checkout to select merchants in the United States, Canada, and Australia, and only listings with native_commerce checkout eligibility show the Buy button. Confirm versions before sign-off.

• Agentic Commerce Protocol · ACP specification on GitHub
• Universal Commerce Protocol · Google for Developers UCP guides

Protocol overview

MCP is a different layer

People ask for “ACP vs UCP vs MCP” as if the third column were another checkout. It is not.

Google’s UCP guide says the protocol is meant to work with Agent Payments Protocol (AP2), Agent2Agent (A2A), and MCP. That compatibility line is why the three names show up together. Compatibility is not identity. An MCP catalog server helps an agent that already speaks your tools. It does not enroll you in Instant Checkout or in UCP Buy buttons.

If the decision is “which checkout contract,” stay in the ACP/UCP table. If the decision is “how our own agent calls Shopify,” that is MCP or OpenAPI behind a gateway — covered in how agents talk to store systems and MCP security.

Where the real difference is

The specification diff is less interesting than the scope diff.

ACP is deliberately narrow. It standardises the handshake — find the product, negotiate a cart across turns, authorise a scoped payment — and stops. That narrowness is a feature: fewer systems touched, faster to reach conformance.

UCP extends past the transaction into order status, modifications and returns. That is genuinely more useful, because a large share of customer contact lives after the purchase, not before it. It is also more expensive, because exposing order modification to an agent is an authorization problem before it is an integration one.

Use-case mapping

Merchant effort, honestly ranked

Across both standards, effort concentrates in the same place — and it is not the protocol.

1. Catalog and attribute work (largest). Specifications lifted out of description HTML into typed fields, variant structure that is queryable rather than encoded in title strings, category-critical dimensions completed. Transfers fully between standards.
2. Identifier and join-key hygiene. Storefront, ERP and warehouse agreeing what a SKU is. If they disagree, an agent surfaces confidently wrong availability.
3. Feed conformance. Meaningful, but mostly mechanical once (1) and (2) are done.
4. Checkout handshake. Increasingly a platform feature.
5. Post-purchase endpoints (UCP only). Authorization design, not plumbing.

When ACP wins

• Your discovery traffic skews to ChatGPT rather than Google AI surfaces.
• You want the smallest possible surface area to reach conformance.
• You already use Stripe, making the documented payment path a short hop.
• You are testing the channel and want a reversible commitment.

When UCP wins

• Google AI Mode and AI Overviews are already reshaping your organic traffic.
• You maintain a healthy Merchant Center feed, so step one is largely done.
• Post-purchase support volume is a cost centre you want agents to reach.
• You sell in categories where Google Shopping has always driven discovery.

Our recommendation

Do the catalog work now; defer the protocol choice.

We recommend this over picking a standard first because the sequencing risk is asymmetric. Choosing the “wrong” protocol costs you a thin integration layer that largely transfers. Shipping a conformant checkout on an incomplete catalog costs you the whole project, because agents exclude products on missing attributes before quality is ever assessed — and you will have no way to tell that is what happened.

The trade-off we are accepting: this is slower to a demo. There is no press release in “we improved attribute completeness.” If your constraint is showing a board that you are moving on agentic commerce this quarter, a protocol integration is more visible. We think that is the wrong trade, and we would rather say so.

What this comparison does not cover

We have not benchmarked conversion or revenue under either standard, because credible public data does not yet exist and we will not manufacture it. Attribution for agent-originated revenue is immature across the industry. We also have not covered marketplace-specific implementations, or the emerging question of how returns are handled when the buyer never saw the product page — both are live problems without settled answers.

Related comparisons

• Bedrock vs SageMaker — the layer beneath agent workloads
• Amazon Q vs ChatGPT Enterprise

Related reading

Agentic commerce · ACP · UCP · MCP · Agent checkout · Product feed · ChatGPT shopping for merchants · Gemini shopping for merchants · Checkout architecture · Your next customer may be an AI agent