AP2 vs MCP vs x402: The Agentic Commerce Difference
What we tell ecommerce clients about AP2, MCP and x402: which layer each covers, how ACP and Google's UCP fit, what to fix this quarter, and what can wait for your payment processor.
Ecommerce clients have started forwarding us articles about AP2 and x402 with one question attached: do we need this? Our answer is usually "some of it, later." Most of what decides whether an AI agent sells your product happens before any payment protocol gets involved, and that part is work you can start this quarter.
This is how we explain the three protocols to clients, and the order we put the work in.
Three protocols, three different questions
We sort them by the question each one answers. Mixing them up is how a team ends up scoping a payments integration when its real problem is that ChatGPT never mentions its products.
MCP answers "how does an agent reach a tool or a piece of data?" The Model Context Protocol is an open-source standard for connecting AI applications to data sources, tools and workflows. Servers expose tools, clients connect to them, and the docs list Claude, ChatGPT, VS Code and Cursor among the clients. It is not a payment or checkout protocol.
AP2 answers "did the shopper really authorize this purchase, and who is liable if not?" The Agent Payments Protocol, announced by Google on 16 September 2025, uses signed verifiable digital credentials called mandates. In the current v0.2 docs there are two. The Checkout Mandate covers the items and purchase details, with an open stage (the shopper's constraints before the cart is final) and a closed stage (authorization of one finalized checkout). The Payment Mandate authorizes payment against a specific card or instrument and is shared with the credential provider, the networks and the merchant's processor. You may see older material that calls them intent and cart mandates; that is v0.1.
x402 answers "how does the money move, right now, without an account?" x402 is Coinbase's open protocol for stablecoin payments over HTTP. The server replies with status 402 and its payment requirements, the client signs a payment and retries, and the server settles onchain. It also has an MCP transport, so a paid MCP tool can ask for payment inside the tool call.
| MCP | AP2 | x402 | |
|---|---|---|---|
| Layer | Tools and data | Purchase authorization | Payment settlement |
| Who created it | Open-source standard | Google; v0.2 donated to the FIDO Alliance | Coinbase |
| What it answers | How the agent calls your systems | Whether the shopper approved this checkout | How funds move per request |
| What it does not do | Approve or pay | Move money | Record what the shopper wanted |
Where ACP and UCP come in
Two more protocols matter more to a typical store than any of the three above, because they decide whether the agent sees your products.
ACP, the Agentic Commerce Protocol from OpenAI and Stripe, is how product feeds and promotions reach ChatGPT. OpenAI extended it to discovery on 24 March 2026 and covers Shopify merchants through Shopify Catalog. It has moved away from its standalone Instant Checkout, so purchases now finish on the merchant's own site or app, according to ChatGPT's merchant page.
UCP is Google's open standard for discovery, buying and post-purchase support, and it is compatible with AP2 and MCP. It powers a Buy button on eligible listings in AI Mode and Gemini. Per Google's Merchant Center Help, it is available to select merchants for products eligible in the US, Canada and Australia, and you remain merchant of record.
How one order would use all of it
A shopper asks an assistant for a product under a budget. The assistant finds candidates from feeds, crawled pages or Merchant Center. It may check stock through a merchant tool over MCP. Under AP2, the budget becomes an open Checkout Mandate, the agreed cart becomes a closed one, and a Payment Mandate covers the card. The money moves on card rails. In a stablecoin flow it could move over x402 instead, and AP2's site includes a "Human Not Present x402" sample of exactly that.
The point we make to clients is that only the first step involves competing with other brands. Every later step assumes the agent already chose you.
What we tell clients to do now
Get the product data right. If you sell in the US and want to appear in ChatGPT shopping, either make sure ChatGPT can crawl your site or send a feed. OpenAI says a feed is optional when it already crawls you, but feeds give more control and accuracy. Under the ACP feed spec, each product needs a stable id and at least one variant with its own id and title; price, images and GTINs are optional but improve display.
Allow the right crawler. OpenAI's publisher FAQ says sites should allow OAI-SearchBot to appear in ChatGPT search. We check robots.txt for it on ecommerce audits.
Prepare Merchant Center. For Google's UCP checkout, the account needs approved free listings, a defined return policy and customer support contact details. Products need native_commerce(checkout_eligibility) set to true, because it defaults to false. We also check the ineligible list early, since subscriptions, pre-orders, personalized goods, used items and age-restricted products are excluded. The feature is early access for select merchants, so the realistic step is preparing the data and expressing interest.
Measure discovery. This is where we spend most of the effort, and it is covered below.
What can wait
AP2. The initial version supports pull payments like credit and debit cards, and v0.2 is in FIDO Alliance working groups for standardization. For almost every store, AP2 support arrives through the payment processor, not through a project you run yourself. We tell clients to ask their processor which version it plans to support and to revisit it then.
x402. Worth testing if you sell paid API access, data or other per-request digital goods. For a store selling physical products through card checkout, we park it.
Your own MCP server. Only once you have tools worth exposing to agents, such as live stock or order status. Until then it is a connector with nothing behind it.
How we measure the part that matters first
None of these protocols puts your product in front of the agent. That is a visibility problem, and we track it on Promptwatch, the platform we run client programs on.
For ecommerce clients we load the products the store actually sells into ChatGPT Shopping tracking and watch their positions over time. Prompt tracking and citation analytics show whether the brand comes up across ChatGPT, Gemini, Claude, Perplexity and Google AI Mode, and which pages and sources get cited. Agent Analytics crawler logs show which AI crawlers fetch the product pages, which is where we look first when a product drops out of answers. Visitor analytics, installed with a script or the GTM template, ties AI-referred buyers to conversions, so the report puts conversions next to mentions. We wrote up how we separate shopping units from sponsored placements in our ChatGPT Shopping and Ads Radar post.
When a client's payment processor ships AP2 support, that data is what tells us whether agents were reaching the store in the first place. If you want that baseline set up before the checkout work starts, write to us at hello@1001seomedia.com.