A shopper asks ChatGPT for a carry-on and buys it without leaving the chat. The next day she asks Gemini and checks out inside AI Mode. Each purchase runs on a different agentic commerce protocol, written by a different coalition, and the merchant has to support both to make both sales. ACP, UCP and AP2 overlap, compete and borrow from one another. None of them is the settled spec yet.
Key Takeaways
- ACP and UCP compete directly for the checkout layer, while AP2 handles payment authorization and has already changed both its version and its owner.
- A brand that builds for one protocol ties its agentic sales to one set of AI surfaces and inherits every revision that spec goes through.
- Every protocol starts working only after an agent has chosen a brand, so visibility in AI answers decides who gets the transaction at all.
Three Protocols, Two Layers, and No Winner Yet
The Agentic Commerce Protocol (ACP) came first in the checkout race. Stripe and OpenAI released it in September 2025 to power Instant Checkout in ChatGPT, starting with US Etsy sellers. Stripe said over a million Shopify merchants would follow. ACP is open source under Apache 2.0, and Stripe pitches it as a build-once spec that distributes to any ACP-compatible agent.
The Universal Commerce Protocol (UCP) arrived in January 2026 from Google, co-developed with Shopify, Etsy, Wayfair, Target and Walmart. Google says more than 20 other companies endorse it, including Adyen, American Express, Mastercard, Stripe and Visa. Its scope is wider than checkout: Google describes it as covering discovery, buying and post-purchase support, and it powers checkout on eligible product listings in AI Mode and the Gemini app.
The Agent Payments Protocol (AP2) sits one layer down. Google announced it with more than 60 organizations, and its job is narrower: proving that a user authorized an agent to pay. It does this with mandates, which are cryptographically signed records of the user’s instructions, and it runs as an extension of A2A and MCP.
So the map has two layers. ACP and UCP compete for the same job, which is how an agent and a merchant agree on a cart and complete the order. AP2 handles authorization underneath, and UCP lists it as compatible. Every one of these specs calls itself open. Every one also has a home surface where its backers make money.
What Breaks When You Build for Only One
The first risk is reach. An ACP-only build sells through ChatGPT. A UCP-only build sells through Google’s AI surfaces. Shopify has also announced Copilot Checkout for Microsoft’s assistant, so the number of agents a merchant might want to sell through is growing faster than any single spec can claim them.
The second risk is churn. These are young specifications, and young specifications change. AP2 has already shipped version 0.2, which adds “Human Not Present” payments that let an agent buy autonomously, such as grabbing limited-run tickets the moment they go on sale. Code written against the first version needs revisiting, and the same will be true of every early implementation.
The third risk sits in your data model. Each protocol has its own way of describing a cart, a fulfilment option and the information a buyer must supply. Shopify gives the example of a furniture retailer that needs the customer to pick a delivery date, which UCP lets the merchant declare to the agent. If your checkout logic is shaped around one schema, porting it to the next one is a rebuild.
Which Brand the Agent Picks Is Decided Before Checkout
All three protocols begin once an agent has already chosen what to buy. When a shopper asks for the best carry-on for a short work trip, the model assembles an answer from what it retrieves and what it trusts. ACP, UCP and AP2 have nothing to say about that step.
Every agentic commerce protocol starts working after the agent has already decided which brand to buy from.
That selection runs on Share of Citation: how often a model pulls your brand into its answer and names you as a source. It has two layers. Training-data presence is earned slowly through wide coverage. Retrieval presence is earned technically through structured, current content, and that is where the protocol work and the visibility work meet. Google’s UCP announcement pairs the protocol with new product attributes beyond keywords, such as answers to common product questions, compatible accessories and substitutes.
Engines also behave differently. Gemini cross-references Search, YouTube and Scholar, while ChatGPT pulls from across the open web, so a brand can be the default answer on one surface and absent on another. Measuring each engine separately is the only way to know which protocol would actually carry your sales.
growth in AI brand presence for NewDay USA, tracked weekly across five engines over six months.See the case study.
The Biggest Players Are Hedging, and That Tells You Something
Look at who sits on which side. Stripe codeveloped ACP and endorses UCP. Etsy sellers were the first merchants live on ACP, and Etsy also co-developed UCP. Shopify co-developed UCP with Google, yet its merchants also sell in ChatGPT, and Shopify manages both surfaces plus Copilot from a single admin through what it calls Agentic Storefronts. The companies with the most information about adoption are backing more than one horse.
Governance is moving too. Google has donated AP2 to the FIDO Alliance, an industry body that writes open standards, along with Verifiable Intent, an AP2-compatible standard co-developed with Mastercard. Handing a spec to a standards body is usually how fragmentation ends. It is also slow, and the spec tends to change on the way through.
What to Do Before Adoption Shakes Out
The goal is to make the second and third protocol cheap to add, and to put effort where it pays regardless of which spec wins.
- Put a translation layer between your commerce backend and any protocol, so cart, inventory and fulfilment rules live in your system and each spec is an adapter. If you run on a platform that already maintains these adapters, use them. Shopify’s Agentic plan opens its Catalog to brands that do not run a Shopify store.
- Treat product data as the shared asset. Complete attributes, current price and stock, answers to the questions buyers actually ask, and accessory and substitute relationships feed every protocol and every engine’s retrieval at once.
- Run payment authorization as its own workstream. AP2 mandates and Verifiable Intent are heading into FIDO, so assign someone to follow those releases.
- Audit which engines recommend you today before choosing which protocol to build first. If Gemini names you and ChatGPT never does, an ACP-first build optimizes for a surface that is not sending you buyers.
- Track spec versions monthly on each project’s public repository.
Then measure answer share per engine after each protocol goes live. A checkout integration that no agent ever reaches is a cost with no revenue line, and the engine-level data is what tells you which integration to fund next.
FAQs
Are ACP and UCP Compatible with Each Other?
They are separate specifications that cover the same job: letting an AI agent and a merchant agree on a cart and complete checkout. Google says UCP is compatible with A2A, AP2 and MCP, but neither the ACP nor the UCP launch materials describe compatibility with the other. A merchant who wants both surfaces implements both, or uses a commerce platform that does.
Is AP2 a Competitor to ACP?
Only in part. AP2 addresses payment authorization, using signed mandates to prove what a user allowed an agent to buy, and Google built it as an extension of A2A and MCP. It overlaps with checkout protocols wherever they handle payment, and UCP lists it as a compatible layer.
Do Agentic Commerce Protocols Affect Whether AI Recommends a Brand?
No. The protocols govern the transaction after an agent has picked a product. Which brand gets picked depends on whether the model retrieves and trusts that brand, which is a question of Share of Citation, structured product data and agreement across independent sources.
Which Protocol Should a Brand Implement First?
Start with the protocol behind the AI surface that already recommends you and where your buyers ask. Build it behind a translation layer, or through a commerce platform that supports several protocols, so adding the second one becomes configuration work.