ACP vs UCP: two agentic commerce protocols, and why 'which one' is the wrong question

A developer explains that the Agentic Commerce Protocol (ACP) and Universal Commerce Protocol (UCP) are not competing standards but operate at different layers, with ACP serving as a discovery and feed protocol for ChatGPT and UCP enabling live agentic checkout across Google surfaces. The developer advises merchants to implement both, starting with an ACP feed and preparing a UCP manifest, as the foundational work for both is essentially catalog hygiene.

ACP or UCP? The question assumes you have to pick one. You don't — and by mid-2026 they aren't even doing the same job. ACP and UCP are not two competing implementations of the same standard. They sit at different layers, and most of the work required for one is the same work required for the other. Protocol status as of June 2026. Both specs are moving. Verify current spec versions before you implement — I've dated every claim below for that reason. ACP — the Agentic Commerce Protocol, backed by OpenAI and Stripe — launched as the standard behind ChatGPT Instant Checkout. Agent finds product, agent buys product, merchant stays merchant of record. In March 2026 that shifted. ACP moved away from native checkout, and what remains in practice is a discovery and feed protocol : a structured, spec-conformant product feed you publish, OpenAI ingests, and ChatGPT uses to decide whether your products belong in an answer. UCP — the Universal Commerce Protocol, Google-backed, announced at NRF in January 2026 and substantially expanded at Google Marketing Live on 20 May 2026 — went the other direction. It is the live agentic-checkout standard : discovery, cart construction, and transaction, with announced integration across AI Mode, Gemini, YouTube Shopping and Gmail. Early access has been US-first; rollout varies by surface and market. So: one protocol tells AI systems what you sell. The other lets them buy it. | ACP | UCP | | |---|---|---| | Backers | OpenAI + Stripe | | | Role in mid-2026 | Discovery / product feed | Discovery + agentic checkout | | Surfaces | ChatGPT | AI Mode, Gemini, YouTube, Gmail | | Merchant artifact | Spec-conformant product feed | /.well-known/ucp manifest + signed endpoints | | Identity / trust | Merchant application + conformance checks | ECDSA-signed manifest | | Availability | Application-based | US-first early access | | Merchant of record | You | You | The last row is the one people miss. Neither protocol takes over the transaction relationship. You are still the merchant of record, you still own fulfilment, returns and support. What changes is where the buying intent gets captured. Here's why "pick one" is a trap. Strip both specs down to what they demand from your backend and you get almost the same list: That's not protocol work. That's catalog hygiene. It survives ACP pivoting again, UCP renaming itself, and whatever the third protocol turns out to be. Phase 0 — foundations do this regardless . Fix crawler access. Get Product + Offer structured data rendering server-side with availability as a full schema.org URI. Publish an LLM-readable content map. None of this is protocol-specific and all of it pays off in ChatGPT, Gemini and Perplexity today. Phase 1 — ship the ACP feed. It's the lower-effort artifact, the audience is already there ChatGPT handles enormous volumes of shopping-intent queries , and building it forces you to fix the data quality problems you'd hit in UCP anyway. Cron it at a real interval — 15 minutes is a sane starting point for a catalog with moving stock, not nightly. Phase 2 — prepare the UCP manifest. Publish /.well-known/ucp , generate your ECDSA signing keys, get the profile validated. Even if agentic checkout hasn't reached your market yet, the manifest is cheap to maintain and the queue is real. Phase 3 — decide about checkout deliberately. Letting an agent complete a purchase is a policy decision as much as a technical one: order value caps, confirmation requirements, fraud posture, what happens when an agent gets it wrong. Have that conversation before you flip the switch, not after. Worth stating plainly because it shapes everything: an agent operating over a server-side protocol cannot tokenize a card . Stripe's card flows require a PCI-compliant browser context — Stripe.js, Payment Element — which is fundamentally incompatible with a backend tool call. Every current agentic-checkout design works around this rather than through it: offline/invoice methods, stored tokens tied to an authenticated customer, or a handoff to a hosted payment page at the last step. Anyone claiming an agent "pays with your card" via a pure API protocol is describing something with a browser hidden inside it. This is the honest ceiling on agentic commerce in 2026, and it's the thing to design around. There's no built-in support for either protocol — Magento gives you full control of the stack and full responsibility for it. There are open-source options for both sides; I maintain some of them angeo/module-openai-product-feed for ACP, angeo/module-ucp for the UCP profile and signing keys — MIT, Composer . Disclosure: I wrote those, and there are alternatives with different licences and coverage worth comparing. But module choice is the last 10%. The first 90% is whether your catalog data is actually correct, and no module fixes that for you. What's your read — is anyone here building against UCP checkout in production yet, or is everyone still in feed-and-wait mode? Genuinely interested in what the early-access experience looks like from the US side.