Google has started letting US merchants take orders inside AI Mode and Gemini. Its setup guide is specific: publish a UCP profile on your own domain, expose a cart endpoint, pass a sandbox test, then apply. Stores on a commerce platform are told to ask their provider. A business that runs its own systems has to do the build itself, and the holidays are about six weeks out.
Between October 6 and 7, Google began rolling out a new hub inside Merchant Center that lets US merchants take orders from AI Mode and Gemini. A merchant chooses how the sale works: hand the cart over to its own website, let the shopper check out on Google, or link the shopper's account. Before any of that, Google's developer guide asks for work on the merchant's side. Publish a UCP profile at /.well-known/ucp on your own domain, expose a cart endpoint Google can call, and pass a test in a sandbox. Only then do you fill in an interest form and wait to be selected. Newegg went live with UCP checkout on Google the same week.
Google's help page has one line for merchants on a commerce platform: contact that provider for its UCP plans. Everyone else gets the developer guide, and PPC Land's reading of it is blunt that the technical implementation "is required before onboarding begins." A store on a big platform waits for the platform to ship it. A business that runs its own catalog, booking engine or ordering system has a build to finish before Google will even look at it.
Google has published the to-do list for selling inside its AI. If your systems are your own, the list is yours too.
The sale is moving inside the assistant, and the assistant only sells for businesses it can call.
For most of this year AI assistants recommended and the shopper clicked through to buy. That is changing quickly. On October 5 TikTok launched an AI shopping assistant and one-click checkout inside its For You feed, built with Salesforce, Shopify, Shoplazza and Stripe, TechCrunch reported, so the purchase happens in TikTok rather than on the merchant's site. Google's hub does the same for AI Mode and Gemini. Each time, the merchants who get the Buy button first are the ones a platform wired in or the ones who published a standard interface themselves.
The interface is the hard part, and the people building these agents say so. Vikas Mittal, an architect at Walmart Global Tech, wrote in Forbes on October 7 that the shopping agent is the easy piece. What's hard is the real-time catalog, price, inventory and order state behind it. Forrester made a matching point on October 6 about product data: it describes products well enough, but it doesn't answer the questions shoppers actually ask an agent. Neither problem gets solved by a better product feed. An agent needs a system it can ask directly.
Last week showed that being open to agents and being readable to them can come apart. This week Google wrote down what readable means.
Our last briefing covered Hilton saying it accepts AI agent bookings while Meta's Muse still couldn't get a room rate out of it. That left an open question: what does a business actually have to build so an agent can read it? For Google's surfaces there is now a written answer, and it's a short list of technical items rather than a policy statement.
| Requirement | What it means |
|---|---|
| UCP profile | A file at /.well-known/ucp on your own domain that tells Google what you support, including the cart capability. |
| Cart endpoint | An API call Google uses to create a cart with the shopper's items before handing them to you. Native checkout on Google is a separate API. |
| Access | API key or OAuth 2.0. |
| Protocol version | UCP 2026-04-08 or later, the first release with the cart API. |
| Sandbox test | Your profile and cart calls validated before onboarding. |
| Selection | Interest form after the build. Google picks who joins the pilot. Canada and Australia wait until 2027. |
Source: Google Merchant Center help and developer documentation, as reported by PPC Land and Search Engine Land, October 6 and 7, 2026.
Two other things moved in the same direction. Meta and Sierra announced a Personal Agent Protocol on October 6 with Shopify, Stripe, Walmart and others, PYMNTS reported. It covers how an agent signs in, what a shopper lets it do, and what a business lets agents do through its site or API. "The company decides what it makes available," Sierra's founders said. The first draft is due by the end of October and hasn't been published yet. And on October 7 OpenAI published a case study on Radisson's hotel app in ChatGPT, which reads live rates and inventory straight from Radisson's central reservation system through an MCP layer Accenture built. Where Hilton showed what being unreadable looks like, Radisson shows a hotel group's own system made readable, built as a custom project.
The Buy button goes to the businesses that finished the build. Nobody is going to finish it for a business on its own systems.
The money is starting to show up. Menlo Ventures expects agentic purchases to drive $100 billion in online sales over the next 12 months, based on a September survey of 5,067 US adults run with Morning Consult and reported by MediaPost. The same survey found 32% have already let an AI act for them without a final approval step. Travel shows how fast it can move. Hotel traffic data from Markus Busch, published by Hospitality Net on October 7, has AI-referred visits to hotel websites up fourfold in a year and converting about as well as Google organic search, though still under 1% of visits. And Forrester found that 89% of consumers who used answer engines for retail used them to research products, which is where the shortlist gets made.
Google's hub is a pilot, and nobody outside Google knows how many merchants it will select or how quickly it opens wider. But the build comes before the application, so the businesses that can be picked are the ones that have already done the work. For a merchant on its own systems, waiting to see how the pilot goes means watching the holiday season from the outside.
Control is the other half. A PYMNTS roundtable of 40 senior executives, held at Mastercard on October 6, found shoppers happy to let AI research and compare, and much less willing to let it choose the seller or pay without approval. American Express published a business playbook for agentic commerce the same day and said it intends to protect card members and merchants from errors made by registered agents, though Yahoo Finance notes it hasn't published the terms. The card networks are working out who is liable when an agent gets it wrong. You'll want to have decided exactly what an agent can do in your systems before that question lands on your desk.
The first line on Google's list is a UCP profile on your own domain. That's what we publish.
BeaconSpec is the discovery layer for agentic commerce. You already run APIs behind your catalog, booking engine or ordering system: REST, OpenAPI or GraphQL. We turn them into one curated server an agent can call directly, and publish its UCP profile at your own /.well-known/ucp, on UCP 2026-04-08, the version Google's hub asks for. Your catalog is served in UCP's product format, so an agent asking for a product, a price or availability gets the answer your system gives.
You pick exactly which operations an agent may use, and anything you leave off doesn't exist as far as the agent is concerned. That is the Personal Agent Protocol's principle, "the company decides what it makes available," applied to your own systems today. When a task needs the customer signed in, like checking an order, we handle the standard OAuth login so the agent acts inside that customer's own account, with no shared passwords.
We host it, or hand you a sealed container to run in your own environment. BeaconSpec isn't a platform or a marketplace, so your interface doesn't depend on anyone else's plans. Google's hub already requires a UCP version from this April, and the spec will keep moving. Keeping your profile correct as it does is our job.
Google has published what it takes to sell inside AI Mode. Platforms do it for their stores. On your own systems, you do it.
The hub asks for a UCP profile on your own domain, a cart endpoint and a sandbox test before you can even apply, and Newegg is already live.
Menlo Ventures expects $100 billion in agentic online sales over the next year. Publishing a curated interface on your own domain is how a business off the big platforms gets on the list.
Worth watching over the next month: how many merchants Google selects for the hub and whether it opens wider before the holidays; what the Personal Agent Protocol's first draft asks of businesses when it lands; and what terms American Express attaches to its protection against agent errors.
BeaconSpec exists for exactly this shift: making your business discoverable and transactable by AI agents over UCP, on an interface you publish and control, kept correct as the standards underneath it keep changing.
See what we do • Try the free AI Visibility Checker • Read more articles
David Soden writes about automation, web strategy, and building durable technical systems for businesses. Photography via Pexels (RDNE Stock project, Christian Naccarato, Vitaly Gariev). Figures cited are drawn from Google's Merchant Center UCP integration hub documentation, as reported by Search Engine Land, Search Engine Roundtable and PPC Land (October 6 and 7, 2026); Chain Store Age on Newegg and The RealReal (October 8, 2026); TechCrunch on TikTok's AI shopping assistant and checkout (October 5, 2026); PYMNTS on the Personal Agent Protocol (October 6, 2026); OpenAI's Radisson case study (October 7, 2026); Vikas Mittal of Walmart Global Tech in Forbes (October 7, 2026); Forrester on product data and answer engines (October 6, 2026); Menlo Ventures' State of Consumer AI survey with Morning Consult, via MediaPost (October 5, 2026); Markus Busch's hotel traffic data, via Hospitality Net (October 7, 2026); the PYMNTS holiday roundtable (October 6, 2026); and American Express's agentic commerce playbook, via Yahoo Finance (October 6, 2026).