Agentic Commerce: Make Products Agent-Ready

Agentic commerce is shopping in which an AI agent finds, compares and sometimes buys products for a person, and a store is agent-ready when the agent can complete every step without guessing or asking a human. That comes down to four fixes: product data that agrees everywhere, policies a machine can read, checkout without human-only steps, and a decision about which agents you let in.

This page is for store owners and ecommerce marketers. It deliberately skips how any one assistant ranks products (that is covered in ChatGPT shopping: how products get picked) and focuses on what you can change in your own catalog, policies and checkout. Facts below come from the Stripe, Google, OpenAI, IBM and Agentic AI Foundation pages linked in each section, checked October 28, 2026.

On this page

• What is agentic commerce, and what does an agent need from a store?
• Fix 1: product feeds and required data
• Does the same fact need to match in four places?
• Fix 2: policies an agent can read
• Fix 3: checkout friction and agent identity
• Which agentic commerce protocols should a store care about?
• An agent-readiness audit you can run this week
• Where agentic commerce readiness goes next

What is agentic commerce, and what does an agent need from a store?

Agentic commerce means an AI agent takes a shopping goal and works through it on the shopper’s behalf. Stripe’s guide describes agents finding, comparing and potentially making purchases, and tells merchants to expose product, pricing, availability and policy data in formats agents can query directly. IBM cites its Institute for Business Value finding that 45% of consumers already use AI for part of the buying journey, and advises making product data readable by both people and machines.

An agent does not need your homepage to be pretty. It needs to get through a chain, and the chain breaks at the weakest link. The diagram below shows five steps and the typical store problem that stops each one. The steps are our simplification, not a standard.

If you want the wider picture of agents browsing sites, our piece on agentic search optimization covers what changes when the visitor is software. The rest of this page is the store-side repair list.

Fix 1: product feeds and required data

Start with the feed, because it is the one place you control completely. Google’s Merchant Center product data specification lists the minimum attributes: id, title, description, link, image_link, price and availability, with brand required for most new products. Limits worth knowing: titles up to 150 characters, descriptions up to 5,000, and an id that stays the same when you update the product.

Three details from that spec cause real damage:

• Identifiers. Strongly recommended gtin values must be valid. The spec says a product submitted with an incorrect GTIN will be disapproved, while products that have a GTIN but are submitted without one may have limited visibility. Use mpn only when no manufacturer GTIN exists.
• Images. The spec says image size requirements of at least 500 x 500 pixels will be enforced starting January 31, 2027. Placeholder or generic images are not allowed.
• Titles and descriptions. They must describe the product accurately and match the landing page, with no promotional text or all capital letters.

OpenAI takes a different route for ChatGPT. Its commerce documentation describes sharing a structured product feed, with a file upload option delivered to an SFTP server, and says the product specification covers discovery in ChatGPT with separate requirements for ads and checkout. The page does not list fields, so check the linked specs before building. Shopify and Etsy merchants should read our note on Shopify Catalog and the Agentic Commerce Protocol first, since the platform may already do this for you.

Two judgment calls from us, not from the specs. Treat the feed as the master record and generate everything else from it where you can. And write descriptions for a reader who cannot see the photo: dimensions, compatibility, what is in the box. An agent comparing a $90 and a $110 item will choose on the attributes you actually stated.

Does the same fact need to match in four places?

Yes, for price and availability. The Merchant Center spec tells you to submit a price that matches your landing page, structured data and checkout, and an availability that matches the landing page, checkout pages and structured data. That is four places for one number. The title and description are held to the landing page.

The matrix below shows which fact has to agree where, per that spec. Any gap is a chance for an agent, or Google, to distrust the listing.

Page markup is where drift usually starts. Google’s merchant listing structured data guide requires name, image and offers on the Product, with price greater than zero and a three-letter priceCurrency, and recommends availability. It also notes that Google may attempt to verify merchant listing data before showing it, and that JavaScript-generated Product markup can make Shopping crawls less frequent and less reliable, which matters for fast-changing price and stock. If your markup is injected late by a script, see our guide to JavaScript rendering for AI crawlers. The full property rules are in our Product schema guide.

Here is a quick fix order for the drift problem:

Fix 2: policies an agent can read

Agents run trust checks before they buy, and policies are the evidence. Stripe’s guide tells merchants to publish agent-readable trust signals including return policies, fulfillment guarantees and contact information, and to pass the trust checks an agent runs before purchase. A returns policy buried in a PDF footer, or only visible after login, is invisible to that check.

Google gives the markup half of this. In its merchant listing documentation, shippingDetails (OfferShippingDetails) and hasMerchantReturnPolicy (MerchantReturnPolicy) are recommended on the Offer, and Google recommends a global policy under Organization markup. It says to use the offer-level property only to override that policy or when no standard policy exists. So for most stores the work is one Organization-level policy, not thousands of per-product blocks. We walk through that in Product schema.

A practical policy checklist:

• State the return window, who pays return shipping and the refund method in plain sentences on a crawlable page.
• State shipping regions, handling time and carriers, and keep it consistent with what checkout shows.
• Put a real contact route (email or phone) on the same page.
• Make the same policy readable in markup, with the Organization-level pattern above.
• Do not make policy text depend on a login, a popup or a script that needs a click.

Whether any given agent uses this markup is not documented for every assistant. The safe reading is that the facts must exist in readable form, and markup is the cleanest way to say them.

Fix 3: checkout friction and agent identity

Checkout friction for an agent is anything a human must decide or prove. The agent can fill a form, but it stalls on a CAPTCHA, a forced account creation, an unlabeled size selector or a payment form that needs a physical card number it does not hold. These are our observations from the shape of the problem, so test your own flow rather than trusting this list.

The payment side is being solved by agents holding tokens, not card numbers. Stripe says that with Stripe Issuing, a temporary virtual card number is created for the agent so it never sees real card details, and recommends flows that return structured metadata and can run without a human in the loop.

Identity is the part to decide on purpose. The Agentic AI Foundation explainer says merchants need to verify an agent before letting it use their normal checkout and fraud controls, and that once verified, an agent can use tokenized payment credentials through the merchant’s current checkout flow. It names Web Bot Auth, led in draft by Cloudflare, which lets an agent sign its web requests, and the Trusted Agent Protocol from Cloudflare and Visa, which signals whether an agent is browsing or paying. Its argument is that merchants should not have to choose between blocking all bots and admitting all bots.

The layered model below summarizes the repair order. Fix from the bottom up, because each layer depends on the one under it.

If you already restrict crawlers in robots.txt, keep two decisions separate: whether a bot may train on your content, and whether a shopper’s agent may browse and buy. Our pages on robots.txt for AI crawlers and whether to block AI crawlers cover the first.

Which agentic commerce protocols should a store care about?

Most stores should care about the ones their platform and payment provider already support, and ignore the rest for now. The Agentic AI Foundation article lists the field. The table keeps only what its page and Google’s page state.

On the Google side, the UCP page says merchants can use existing Merchant Center feeds to reach shoppers on AI surfaces, that native checkout is the default and embedded checkout is limited to approved merchants, and that interested merchants join a waitlist. A UCP checkout guide adds that only listings using the native_commerce(checkout_eligibility) product attribute display the Buy now button. Both pages describe an early-access path, so confirm current availability before planning around it.

The AAIF piece also says merchants keep their checkout rules and remain merchant of record, and that sites built for people will keep serving agents because new commerce APIs will not be adopted all at once. That is the strongest reason to fix the basics first.

An agent-readiness audit you can run this week

Run the audit on your five best-selling products, not the whole catalog, then scale the fixes. For each product, look at every place it lives and record the result.

Do the checks by hand once, then automate the price and availability comparison; those drift fastest. Our product schema and merchant listings checklist goes through the markup rows in detail, and ecommerce SEO covers the category and faceted-navigation work that sits upstream of all this. For a Shopify store, Shopify SEO shows what the platform handles for you. For the broader set of tasks, see our checklists.

Where agentic commerce readiness goes next

Agentic commerce readiness will probably look less like a project and more like data hygiene with a longer shelf life. The specifics will change: protocols merge, ChatGPT’s checkout story has already shifted, and Google’s UCP is still in early access. What does not change is that an agent needs one consistent set of facts, policies it can read and a checkout it can finish.

So the sensible order is the one in the layered diagram: make price, stock and identifiers agree, publish policies in readable form, remove human-only checkout steps, then decide which agents you will verify and admit. Nothing in the sources tells us how much extra revenue this produces, and we will not guess. It is cheap insurance against being skipped.

If you want a second pair of eyes on a catalog, our AI SEO service starts with exactly this kind of data and policy audit, and we can tell you which of the four fixes matters most for your store.