Agentic commerce: the protocols that decide whether AI buys from your store
By Daniel Pisano, Partner & CGO at Vurdere
Agentic commerce is when an AI agent searches, compares and completes a purchase on a shopper's behalf. To do that, the agent and the store talk through open protocols such as UCP, ACP and AP2, where the agent reads structured catalog, price and rating data. Stores that don't expose that data in a readable form are left out.
What is agentic commerce in practice?
Until now, e-commerce was designed for a person in front of a screen. They search, open three tabs, read the reviews, compare shipping and decide. In agentic commerce, that person delegates. They tell an assistant what they want and how much they will pay, and the agent does the rest.
Shoppers are already moving. According to Salesforce, 39% of consumers have used AI to find products, and AI and agent-referred traffic drove an estimated 21% of global orders in the 2025 holiday season, around $263 billion in sales. Retailers are behind. A PYMNTS Intelligence survey commissioned by Visa Acceptance Solutions, run in March 2026 with 1,185 retailers in the US, Brazil and the UAE, found that only 24% of US retailers can identify when a visit or order came from an agent. Across the whole sample, only 15% keep structured data that an agent can read.
Which protocols power agentic commerce?
The acronyms are confusing, but each protocol handles a different part of the purchase. This is how I explain it to merchants:
Protocol | Led by | What it handles |
MCP (Model Context Protocol) | Anthropic, launched November 2024 | How an agent connects to tools and data sources, such as a catalog or a search function. |
A2A (Agent2Agent) | How agents talk to each other, for example the shopper's assistant and the store's agent. | |
UCP (Universal Commerce Protocol) | Google, with Shopify, Etsy, Wayfair, Target and Walmart | The full shopping journey: product discovery, cart and checkout. |
ACP (Agentic Commerce Protocol) | OpenAI and Stripe, introduced September 2025 | Checkout inside assistants such as ChatGPT. |
AP2 (Agent Payments Protocol) | Authorising a payment made by an agent, tied to the user's signed intent. |
MCP and A2A are the technical base. UCP and ACP organise the purchase. AP2 handles payment. A typical agentic purchase goes through more than one of them.
Why does UCP matter so much for merchants?
Google announced UCP at NRF in New York on 11 January 2026. The detail that caught my attention is in the catalog specification. The agent has two basic operations: search for products with text and filters, and look up a specific product. And the product object has an optional field called rating, the aggregate score, with its value, its scale and the number of reviews behind it.
That changes what a review is. For years it was a star widget halfway down the product page, built to persuade a person. In the protocol, it is data the machine reads next to price and stock, at the moment it builds the shortlist it will show the shopper.
The specification is still a draft, and the discussion is open. In May 2026, a proposal asked for two new rating fields: the distribution of star ratings and the number of reviews from verified buyers. It has not been approved yet, but it shows where the protocol is heading.
What can an agent read in your store today?
Run this check on your own store. The agent doesn't see banners, videos or lifestyle photos. It reads:
A catalog with complete attributes: name, variants, sizes, materials and compatibility.
Current price, stock and delivery time.
The aggregate rating and review count in structured form, such as schema.org product data or the platform's native fields.
Product questions and answers, written to answer what customers actually ask.
If your store's rating only exists inside a widget loaded with JavaScript, there is a real risk the agent won't find it. To the machine, the product looks like it has no reviews at all.
How does Vurdere deliver ratings in the format agents read?
This is the problem we work on at Vurdere. Collecting reviews is the easy part. The value is in making customer content structured, trustworthy and readable by whoever is deciding, a person or an agent.
In the Vurdere app for Shopify, the aggregate rating and review count are written to the platform's native metafields, reviews.rating and reviews.rating_count. On other platforms, such as VTEX and Magento, Vurdere delivers the aggregate rating as server-rendered structured data on the product page itself. Installation takes about 3 hours, followed by a week of validation.
In A/B tests with current clients, Vurdere recorded up to 3x higher conversion rate, 15% incremental sales, 20% more organic traffic, a 30% higher average order value and up to 10x more reviews collected. Those results come from human shoppers. The case for agentic commerce is the same: a store with plenty of recent, structured reviews gives the agent more reasons to choose it.
Where should you start preparing your store?
Audit your catalog. Empty or inconsistent attributes drop products out of the agent's search.
Check that the aggregate rating appears as structured data in the page code, not only in the widget.
Increase review volume and recency. A high score from a few old reviews carries less weight.
Answer customer questions on the product page. They become context for the agent.
Ask your platform and payment provider when they will support UCP, ACP and AP2.
Start measuring traffic from AI assistants, even if it is small today.
My view
I think many merchants are watching the wrong race. The conversation is still about which assistant will win, Google's or OpenAI's. For sellers, that matters less than it seems. Both protocols ask for the same things: a clean catalog, reliable prices and a reputation the machine can read.
In that competition, the review stops being page decoration and becomes decision data. Merchants who start treating reviews as data now will reach 2027 ahead. Those waiting for the protocols to settle will find the agent has already picked another store.
Frequently asked questions
What is agentic commerce?
It is a model in which an AI agent searches, compares and completes a purchase on a shopper's behalf, within limits the shopper sets, such as a maximum price or a delivery date.
What is the difference between UCP and ACP?
UCP, led by Google, covers the full journey from discovery to checkout. ACP, from OpenAI and Stripe, focuses on checkout inside assistants such as ChatGPT. For the store, the data requirements are similar.
Do product reviews influence an AI agent's decision?
The UCP specification includes the aggregate rating and review count in the product object, so the rating is part of the data the agent receives when it compares options.
Do I need to change platforms to sell to AI agents?
Not necessarily. The first step is a complete catalog and reviews in structured data. Protocol support usually arrives through your e-commerce platform and payment provider.
Comments