Universal Commerce Protocol: Google Put a Checkout Button Inside the Answer

The Universal Commerce Protocol is Google’s open standard for agentic commerce, and it does something no SERP feature has done before: it puts a working checkout button inside the answer. Not a link to your product page. A purchase, completed on Google’s surface, with you still on the receipt as the seller of record.

Google launched UCP on 11 January 2026 and expanded it heavily at Google Marketing Live on 20 May 2026. Most coverage since has been either the announcement itself or a definitional explainer. Very little of it answers the question an operator actually has, which is narrower and more awkward: what does this require from me, and is my product data anywhere close to ready?

Short answer for most mid-market businesses: no, and the gap is not the checkout API. It is the catalog underneath it.

What is the Universal Commerce Protocol?

UCP is an open standard that lets AI agents and merchant systems talk to each other in one shared language, covering discovery, purchase and post-purchase support. Instead of building a separate integration for every agent that wants to sell your products, you implement the protocol once.

Google announced it on 11 January 2026, co-developed with Shopify, Etsy, Wayfair, Target and Walmart, and endorsed by more than twenty other companies including Adyen, American Express, Best Buy, Mastercard, Stripe, The Home Depot, Visa and Zalando. That partner list matters more than the protocol spec. It is the reason this is unlikely to quietly disappear the way most retail standards do.

UCP is deliberately compatible with the protocols already in play: Agent2Agent, the Agent Payments Protocol, and Model Context Protocol. It supports both REST API and MCP transport. If your engineering team has already looked at agent infrastructure, none of the plumbing here is exotic.

What it powers today, per Google’s Merchant Center documentation, is a checkout button on eligible product listings in AI Mode in Google Search and in Gemini. Eligibility is currently limited to the United States, Canada and Australia, and to participating merchants. You remain the merchant of record. You keep the customer data and the relationship. Google handles the payment step through Google Pay, using cards and shipping details the shopper already has in Google Wallet.

At Google Marketing Live in May, Google widened the surface considerably: an expanded Universal Cart that lets shoppers hold products from multiple retailers, UCP wired into Direct Offers, Demand Gen campaigns, AI Mode shopping and Shopping ads on YouTube, plus buy-now-pay-later through Affirm and Klarna inside Google Pay. Google named Nike, Sephora, Target, Walmart, Wayfair and Shopify merchants including Fenty and Steve Madden as brands the experience would soon support.

Worth being precise about how this differs from the checkout button Google already had. The old button sends the shopper to your site to finish the purchase. UCP completes it on Google’s surface. That sounds administrative and it is not, because it decides which of your assets does the persuading. Your product page used to have a job: photography, reviews, comparison copy, trust signals. Under UCP a shopper can decide and buy without ever loading it, and everything that gets them to the Buy button comes from structured data you filed in Merchant Center.

The design principle is one we have written about before in the context of building a website an AI agent can actually use. UCP makes the stakes explicit: the agent is no longer browsing, it is transacting.

What does UCP actually require from a merchant?

Five things, in rough order of how long each takes to fix.

Notice what is not on that list: a redesign, a new CMS, an AI vendor. The technical lift is real but bounded, and UCP primitives map onto standard retail operations, so the integration follows business logic your team already has.

The unbounded work is the catalog. Return policy, customer support details, shipping terms and product attributes stop being back-office housekeeping the moment Google is completing the sale on your behalf. Every gap in the feed is a gap in the checkout experience.

Is your product data ready for AI shopping surfaces?

Probably not, and there is now a benchmark for how not-ready. Adobe published sector-wide data on 16 April 2026 measuring how much of a page’s content large language models can actually read, scored out of 100%. Across US retail, product pages scored worst of every page type measured.

The pattern is worth reading twice. The pages retailers treat as boilerplate score highest. The pages that carry the actual commercial payload score lowest. Retailers have thousands of SKUs, and Adobe’s finding is that a large share of that product content is currently invisible to models.

The spread between brands is wide too. Among homepages, the best-performing US retail sites averaged 82.5% while the lowest sat at 54.2%. Some companies moved on this early and it shows.

Meanwhile the demand side is not waiting. Adobe’s data, drawn from more than a trillion visits to US retail sites plus a survey of over 5,000 US consumers, put Q1 2026 AI-sourced traffic up 393% year over year, following a holiday period where it ran 693% higher. In March 2026 that traffic converted 42% better than non-AI channels such as paid search and email. A year earlier the same comparison had AI traffic converting 38% worse. Thirty-nine percent of surveyed consumers said they had used AI for online shopping, and 85% of those said it improved the experience.

Better traffic, arriving faster, landing on the pages models can read least well. That is the whole problem in one sentence.

What are Conversational Attributes, and how should product copy change?

Conversational Attributes are product data fields in Merchant Center built for how people phrase things when they are talking rather than typing keywords. Google announced them at Google Marketing Live on 20 May 2026 and they are rolling out globally.

The January launch note described the intent plainly: dozens of new attributes that go beyond traditional keywords to cover things like answers to common product questions, compatible accessories and substitutes. They sit alongside your existing feed rather than replacing it.

What that changes about product copy is the unit of writing. A keyword-era product description was tuned to match a query string. A conversational-era description has to resolve a constraint. Those are not the same job.

Three practical shifts:

1. Write the objection into the data. The questions your support team answers ten times a week are exactly what these attributes are for. Does it fit through a standard doorway. Is it dishwasher safe. Does it work with the older model. If a shopper has to ask, an agent has to guess.
2. State compatibility and substitution explicitly. Models are far better at extracting a stated relationship than inferring one from a product grid. “Works with” and “replaces” are commercially loaded facts, and most feeds omit them entirely.
3. Put constraints in text, not in an image. Dimensions, materials, capacities, coverage areas. A spec table rendered as a graphic is a spec table that does not exist as far as an agent is concerned.

This is the same discipline that governs whether you get named in a recommendation at all. We covered the discovery half of it in a piece on getting products into ChatGPT shopping recommendations. UCP adds the transaction half. The underlying requirement, structured facts a machine can lift without interpretation, does not change between the two.

Will AI Performance Insights tell you whether it is working?

Partly. AI Performance Insights is a Merchant Center reporting feature, also announced on 20 May 2026, that shows how a brand performs across AI surfaces and compares its share of voice against similar competitors. Google said it would roll out in Australia, Canada, India, New Zealand and the United States in the coming months.

Treat that as a directional instrument, not a ledger. Two cautions apply.

First, Google defines the metric. Share of voice across AI-powered shopping experiences is Google’s measurement of Google’s surfaces, and how it is calculated has not been published in detail. It is useful for spotting movement. It is not a number to build a forecast on until the definition is public and stable.

Second, it covers Google. Gemini and AI Mode are not the whole agentic surface, and a brand that reads healthy in Merchant Center can still be absent everywhere else a buyer asks. Independent tracking across surfaces stays necessary, and it belongs inside the same programme as the rest of your website strategy work rather than sitting in a separate reporting silo.

Google also confirmed Ask Advisor integrations are coming into Merchant Center. Useful for interrogating your own data. Still not a substitute for measuring outside the walls.

What should you do if you do not sell products?

Start preparing anyway, because the protocol is coming to services. Google’s developer documentation confirms UCP is expanding into new industries starting with lodging and food, with waitlists already open for both. At Marketing Live, Google described future experiences that let someone book a hotel directly from AI Mode or order food from a conversation inside Google Maps.

The pattern generalises past hotels and restaurants. Any business where the outcome is a booking, an appointment or a quote is describing the same primitives: availability, price, terms, confirmation. UCP’s published roadmap already includes multi-item carts, account linking for loyalty programmes, and post-purchase support for tracking and returns. Those are booking-system concepts as much as retail ones.

For a service business, the near-term work is not integration. It is making the four facts an agent needs unambiguous and machine-readable wherever they live: what you sell, what it costs or what determines the cost, where and when you can deliver it, and how someone commits. Most service sites answer at most two of those in text. We laid out the operational version of this in a checklist on getting booking-ready for AI agents.

Do that work now and the eventual protocol integration is a connector. Skip it and the integration is a rewrite.

Every few years a platform quietly changes which asset does the selling. Mobile did it once. This is that, again, and the asset this time is a data file most companies have never treated as customer-facing.

The uncomfortable part is that the fix is unglamorous. It is return policies written out properly, compatibility stated instead of implied, specs typed as text rather than baked into a JPEG, and someone finally clearing the feed errors that have been amber for two years. None of it demos well. All of it is what stands between your products and a Buy button inside the answer.