An assistant that shops for someone has four problems to solve. It has to find products, talk to your systems, check out, and show that the shopper agreed to pay. Four protocols have grown up around those jobs, and they overlap less than the acronyms suggest.
This guide takes them one at a time, then looks at what they ask of you in common.
MCP: how an assistant talks to your systems
The Model Context Protocol is the plumbing. It gives an AI assistant one common way to connect to tools and data, such as a product catalog, an inventory system or an order database, so you don't build a custom integration for every assistant.
MCP isn't a shopping protocol. It says nothing about carts or payments. What it does is let an assistant call your systems in a predictable way, which is why the shopping protocols are designed to work alongside it. If you have built an MCP server for your catalog, you have built a piece that several of the others can use.
ACP: shopping and checkout inside the assistant
The Agentic Commerce Protocol came out of OpenAI and Stripe and powers checkout inside ChatGPT. It has two parts that matter to a merchant.
The first is the product feed. OpenAI's feed specification requires nine fields for basic discovery: item ID, title, description, product URL, brand, seller name, image URL, availability and price. It recommends identifiers such as GTIN, product attributes such as material and size, shipping and return terms, and review data. Two flags control how a product appears: one for search, one for checkout. Checkout is off by default and you opt in.
The second is checkout itself. The assistant can start a checkout without sending the shopper to your site, using tightly limited payment tokens. You stay the seller. Settlement, refunds, chargebacks and compliance remain yours. Checkout.com has a clear comparison with UCP.
UCP: Google's version, starting on Google's surfaces
The Universal Commerce Protocol is Google's open standard for checkout, order flows and conversational discovery. It launched in AI Mode in Search and in the Gemini app for eligible US retailers, with payment through Google Pay and PayPal planned. Shopify, Etsy, Wayfair, Target and Walmart helped develop it. Visa, Mastercard, Stripe and Adyen are among its endorsers. Retailers remain the seller of record, and it is compatible with AP2 and MCP. (PYMNTS)
The practical difference from ACP is where it starts. ACP began inside ChatGPT. UCP began inside Google's own surfaces. A merchant who sells to high-intent shoppers on Google may see UCP pay off sooner, and one whose customers use ChatGPT will care about ACP first.
AP2: proof that the shopper said yes
The Agent Payments Protocol is Google's payment-authorization layer, and it is vendor neutral. When an agent clicks pay instead of a person, someone has to be able to show the shopper allowed it. For a card payment, that evidence matters in a dispute. AP2 is built to carry it.
Payment networks are working on the same problem from their side. Visa and Mastercard have each published frameworks for verifying agents and carrying shopper authorization, which we cover in who is this bot.
How they fit together
Think of the four as layers.
| Job | Protocol | What it handles |
|---|---|---|
| Talk to your systems | MCP | A common way for an assistant to call tools and data |
| Find and buy in ChatGPT | ACP | Product feeds and checkout inside OpenAI's assistant |
| Find and buy on Google | UCP | Discovery, checkout and orders on Google surfaces |
| Prove the shopper agreed | AP2 | Payment authorization when an agent pays |
They aren't rivals in a strict sense. UCP is designed to work with AP2 and MCP. ACP covers feeds and checkout. A merchant can end up using more than one, depending on where their customers shop.
What they ask for in common
Read the specifications side by side and the same list appears.
- Complete product data with stable IDs and standard identifiers.
- Live price and availability that match across your site, feeds and marketplaces.
- Explicit shipping, returns and warranty terms.
- A way to take an order from a party that isn't a person.
- A way to verify who that party is.
None of this is exotic, and all of it is useful when no agent is involved. A catalog that meets these requirements searches better, filters better and converts better on your own site.
What to do with four protocols
Don't pick one. The channels set their own terms, and those terms keep moving. Build the common base once, then add a thin connector for each channel as the channel earns its place. Checkout.com's advice is similar: keep current channels performing, build a flexible stack and choose partners that support the emerging ecosystem.
The decision worth making early is ownership. This work crosses catalog, payments and legal, and it stalls when nobody is named. Give it an owner, a quarterly review, and a short list of channels to watch.
For the data work underneath, start with your next customer is an AI agent and the agent-readiness checklist.