UCP, the protocol that brings your ecommerce inside AI agents: what it is and why it will matter

What UCP (Universal Commerce Protocol) is and why it matters for ecommerce: checkout inside AI agents, bookings from AI Mode, orders from Maps.

Add us as a preferred source

There is a precise moment when a new channel stops being a curiosity and becomes infrastructure: when protocols get standardized and the major players adopt them. For commerce inside AI agents that moment is now, and the protocol is called UCP, the Universal Commerce Protocol.

If you run an ecommerce, this acronym concerns you more than it seems. Let’s look at what it is, what Google is already switching on and, above all, what is worth doing now.

What UCP is

UCP is an open protocol, promoted by Google together with a group of industry players, that defines how an AI agent can complete commercial operations by talking directly to a seller’s systems. The specification is public (on ucp.dev and GitHub) and covers the fundamental operations of commerce:

  • Checkout: complex carts, dynamic pricing, tax calculation;
  • Order management: shipment tracking, returns, real-time updates;
  • Identity linking: secure customer identity via OAuth 2.0;
  • Catalog search: the agent queries products rather than browsing pages;
  • Hotel bookings and food ordering: the first extensions beyond retail, with real-time availability and prices.

Technically it speaks the language of the web you already know: REST and JSON-RPC as transports, OAuth for identity, and integration with the other protocols of the agentic ecosystem, AP2 for payments, A2A and MCP for agent-to-tool communication. No exotic technology: open, composable, documented standards.

What Google is already switching on

The interesting part is the speed at which Google is bringing it into the products people use every day.

  • Hotel bookings directly from AI Mode in Search: the user asks, the agent proposes and books.
  • Food ordering from Google Maps, with home delivery.
  • Direct Offers in ads: the purchase happens inside the ad, without visiting the site.
  • Shopping ads on YouTube with product feeds and connected DemandGen campaigns.
  • AI performance analytics in Merchant Center: a tool to monitor how visible your products are in AI experiences compared with competitors.

The first rollout covers Canada, Australia and the United Kingdom, with analytics also active in the US, India and New Zealand. Most of Europe is not in the first wave, and that is good news: there is a window of time to get ready.

September 2026 update: UCP moves beyond retail

On 25 August the project published version v2026-08-25 of the specification, the fourth since the January launch and the first since April. Versions now carry a date instead of a number, and a merchant can keep supporting earlier ones while adopting the latest.

The most visible addition is grocery. Agents can locate physical stores, with address, opening hours in a standard format and coordinates, and order products by weight. It is the first piece of UCP written with supermarkets in mind.

Then payments. The specification has been reorganised into three parts (Shopping, Payment and Common) because payment has to work outside the retail cart too. Vendor-neutral 3D Secure 2 authentication arrives, along with deposits and instalments and the option to split an order across several payment methods.

The third change sits outside the specification and matters more than the other two. UCP is expanding by industry, and each industry now has its own technical council. Since 16 July there is a Food Technical Council with Block (Square), DoorDash, Google, Toast and Uber Eats. Since 11 August, a Lodging Technical Council with Amadeus, Booking.com, Expedia, Google, Hilton, Marriott and Trip.com. Stripe has sat on the Governing Council since 28 April, next to permanent members Google and Shopify. The specifications for food (finding the restaurant and the menu, choosing dishes, paying with tip and delivery notes) and for lodging (searching hotels, picking a room, booking) are being written with these companies. The August release lays their shared groundwork, with no release date.

Anyone who has already implemented UCP has work to do. The fulfillment schema changes, buyer consent moves from fixed fields to a free-form map, and signing keys have to be consolidated into one JWK Set. Anyone still laying the groundwork, which in most of Europe means nearly everyone, can leave things as they are. The four moves below hold.

One last observation, the most useful one if you don’t sell shoes. Look at who sits on the two councils. Booking.com, Expedia, Amadeus, DoorDash, Uber Eats. The rules by which an agent will book a room or order dinner are being written by the very intermediaries hotels and restaurants have spent years trying to get out from under. A hotel whose availability and prices are readable today only inside the OTA’s channel manager will, tomorrow, be readable by the agent only through the OTA. The data question we cover below matters to that hotel more than to anyone else.

Why it matters for your ecommerce

For years we have been describing the same trajectory: first you had to be found (SEO), then be chosen by the AIs that generate answers (GEO). UCP adds the final link: being bought. The agent completes the transaction, beyond citing you as a source.

The consequences are concrete.

The first: your shop’s interface stops being the only point of sale. If a customer asks their assistant “order me the same shoes as last month” and the assistant can do it via protocol, your homepage, your product pages and your funnel are never even seen. What counts is what your systems can tell the agent: availability, price, conditions, delivery.

Then there is the logic of the flow: whoever exposes the protocol gets in, whoever doesn’t stays out. It is the same dynamic we saw with Google Shopping and with marketplaces, but more radical. There is no alternative storefront to man here, only a conversation in which you either exist or you don’t.

And finally, data itself becomes the product. An agent reads data and is not swayed by nice design. Incomplete feeds, stale availability, prices inconsistent between site and feed, which today cost you some conversions, tomorrow mean being discarded by the agent silently and systematically.

What to do now: four concrete moves

In most of Europe UCP cannot be implemented yet, so the task is to arrive prepared, and the preparation coincides with ecommerce optimization work that already pays off today.

  1. Bring product data to impeccable quality. Complete feeds, availability and prices synchronized in near real time, rich attributes. It is the prerequisite for everything: the agent buys from data.
  2. Complete your site’s structured data. Product schema, offers, availability, reviews: they are already how AIs understand your catalog, and they will be the baseline against which agents check consistency.
  3. Master Merchant Center and the new AI analytics. The AI performance tool is the first official thermometer of how visible you are in agentic experiences compared with competitors. When it reaches your market, you will want a history, not a cold start.
  4. Check how readable your site is for agents. Robots, structured data, AI-accessible content: it is the layer beneath the protocol, and you can measure it right now with our agent-ready checker.

It is the path we described in From being found to being chosen, taken to its commercial conclusion: in the age of agents, visibility and transaction converge. Those who treat their data as sales infrastructure will reach the European rollout with an advantage that is hard to close.

Sources

Frequently asked questions

It is an open protocol, promoted by Google together with the industry, that lets AI agents complete commercial operations directly: checkout, order management, hotel bookings, food ordering, catalog search. The specification is public on ucp.dev and GitHub.

Today the customer lands on your site and buys inside your interface. With UCP the transaction can happen inside the conversation with the AI: the agent queries the catalog, builds the cart and completes the purchase by talking directly to your systems. The shop's interface stops being the only point of sale.

No, and the direction is clear from who sits on the technical councils. Since July 2026 there is a Food Technical Council with Block (Square), DoorDash, Google, Toast and Uber Eats, and since August a Lodging Technical Council with Amadeus, Booking.com, Expedia, Google, Hilton, Marriott and Trip.com. Version v2026-08-25 adds grocery (store locations, opening hours, orders by weight) and lays the groundwork for food and lodging, whose specifications are in progress. B2B lead generation stays out of scope for now.

Open, widely adopted standards: REST and JSON-RPC as transports, OAuth 2.0 for customer identity linking, and integration with the other protocols of the agentic ecosystem such as AP2 for payments, A2A and MCP.

The first rollout announced by Google covers Canada, Australia and the United Kingdom, with AI visibility analytics in Merchant Center active in the US, Australia, Canada, India and New Zealand. It is the right window to get ready before your competitors, because most of Europe is not yet in the first wave.

Four things: bring product data to impeccable quality (feeds, availability, prices), complete structured data on the site, master Merchant Center and the new AI analytics, and check how readable your site is for agents. Agentic transactions are built on these fundamentals.