Product Data for AI Agents: What Your Catalog and Feed Need

Key takeaways

When a shopper asks an assistant for a waterproof jacket under $100, the assistant does not browse your store. It compares records. Your homepage, your photography and your brand voice never enter the comparison. The record does.

That is why product data is the first thing to fix in agentic commerce, before any protocol work. UCP and ACP define how an agent finds, prices and buys a product. They cannot repair a record that says “Jacket” and “Great product!”.

Same jacket, two records

Take one product and describe it twice. The weak record has a one-word title, a promotional description, no GTIN, a price without a currency, an availability field that always says “in stock”, return terms hidden in a PDF and a small thumbnail. The agent-ready record has the full title with variant detail, a plain-text description, a valid GTIN, 79.99 USD, availability that follows real stock, accepts_returns with a number of days and a direct JPEG of the product alone.

The product is identical. The first record gets skipped and the second one gets compared.

The six fields that decide whether you are understood

Product feeds have dozens of attributes, but six of them carry most of the weight. For each one, here is what the agent does with it and where catalogs usually break.

1. Identifier. The agent uses it to match your product across sources. OpenAI’s feed specification asks for a GTIN of exactly 8, 12, 13 or 14 digits with a valid check digit, or an MPN with its original punctuation and casing. Leading zeros must survive. Spreadsheets love to delete them, so check the export, not just the source.

2. Title. It is matched against the shopper’s request, so it has to describe the product and not advertise it. “Waterproof Trail Jacket, Men’s, Blue, M” works. “SALE! Best jacket ever” does not. Keep titles concise: the specification points to roughly 150 characters as the ceiling.

3. Price. Format it as amount plus currency, for example 79.99 USD. If you use sale_price, it must be strictly lower than price and use the same currency. An agent that cannot trust the total will not start a checkout.

4. Availability. Use the status values the specification defines (in_stock, out_of_stock, pre_order, backorder, unknown) and make them reflect real stock. An agent that commits to a purchase and meets an out-of-stock error remembers that you were unreliable.

5. Variants. Each size and color needs its own record, its own identifier and its own image. One shared record for all variants forces the agent to guess.

6. Returns and shipping. Put them in fields, not only on a policy page. The feed has accepts_returns, return_deadline_in_days and a return_policy URL. For final-sale items, send accepts_returns=false and omit the return window.

What OpenAI’s feed specification requires

If you want to appear in ChatGPT shopping, the minimum is short. The specification lists nine required fields:

Files can be JSONL (preferred), CSV or TSV, and gzip compression is accepted. Use UTF-8, absolute HTTPS URLs and direct image links to a JPEG or PNG. Avoid placeholder values like “unknown” or “n/a” in optional fields, because a placeholder looks like data and is worse than an empty field.

Treat this list as the floor. It gets you into the index. The six fields above decide whether you win the comparison.

How often should the feed update?

A feed that is correct on Monday and stale on Tuesday is a risk, because the agent shows a price your checkout no longer honors.

OpenAI’s guidance points to a hybrid approach for larger catalogs: upload the full feed as a file once a day, and push price and stock changes through the API as they happen. Promotions go through the API only. Small catalogs can run on the API alone.

The rule inside the specification is short and worth printing:

Whatever rhythm you choose, test it. Change a price in your store, then check how long the feed takes to show it.

Seven mistakes we see most often

Fixes by effort

You do not need a project to start. Sort the work by how long it takes.

A 10-minute test you can run today

Pick one product that sells well and open its record in the feed, not in the store.

If three or more checks fail, the rest of your catalog likely has the same problem at scale.

Where this fits in agentic commerce

Product data is the first layer. Protocols sit on top of it. To see how agents discover and buy from a store, start with What Is UCP? Universal Commerce Protocol Guide, What Is the Agentic Commerce Protocol (ACP)? and the broader overview What Is Agentic Commerce?.

If you want an outside view of your own catalog, feed, checkout and protocol setup in one pass, our agentic commerce readiness audit is a fixed-price, one-week review.

FAQ

Do I need a separate feed for AI agents?

Usually not. A clean Google-compatible feed is a good base, and OpenAI accepts Google-compatible files. You mainly add the fields the agent specification requires, such as seller_name and return terms.

╌╌╌╌╌╌╌╌╌╌╌╌╌╌╌╌╌╌╌╌╌╌╌╌

Is a GTIN mandatory?

The specification requires at least one identifier, either a GTIN or an MPN, depending on the product. A valid GTIN is the safest choice because it lets the agent match your product with other sources.

╌╌╌╌╌╌╌╌╌╌╌╌╌╌╌╌╌╌╌╌╌╌╌╌

How often should I update the feed?

Send price and stock changes as soon as they happen, and upload the full feed once a day. Small catalogs can work through the API alone.

╌╌╌╌╌╌╌╌╌╌╌╌╌╌╌╌╌╌╌╌╌╌╌╌

Does structured data on product pages still matter?

Yes. The feed is the main channel, but schema.org Product markup keeps the page and the feed consistent and helps other agents that read pages directly.

╌╌╌╌╌╌╌╌╌╌╌╌╌╌╌╌╌╌╌╌╌╌╌╌

What if my platform is Shopify?

Shopify generates much of the catalog data for you, but its defaults are not always complete. Check GTINs, variant records and return terms in your admin before you rely on them.

╌╌╌╌╌╌╌╌╌╌╌╌╌╌╌╌╌╌╌╌╌╌╌╌

Can I fix this without changing my platform?

In most cases yes. Missing identifiers, placeholder values and return fields are data problems, not platform problems.

Conclusions

How do ACP and UCP find your products?