The Product Data AI Shopping Agents Read: GTINs, Attributes and Structured Data

AI shopping agents read structured product data: identifiers, titles, attributes, price, availability, shipping and returns. They do not browse like people. They decide in a fraction of a second whether a product fits a request, and if a field is missing or wrong, the product is simply not considered.

This guide goes field by field through what agents rely on, how to audit your own data, and where stores most often go wrong. If you are new to the topic, agentic commerce explained gives the background first.

1. Identity: who made it and what exactly it is

Identity fields tell an agent which product this is, so it can recognize the same item across stores and compare offers.

• Brand, written the same way everywhere.
• GTIN (UPC in the US) for products that have one; MPN when they don’t.
• Model or variant clearly identified.

Two things go wrong here most often. Brand names vary between products (“Acme”, “ACME Co.”, “Acme Home”), and variants share one identifier or none at all. Both make it harder for an agent to treat your products as a coherent catalog.

2. Title: what the product is

A good title starts with the product type and includes brand and key attributes. Agents match requests like “women’s waterproof hiking jacket” to titles and attributes, not to clever names.

Compare “Trailmaster Pro” with “Trailmaster Pro Women’s Waterproof Hiking Jacket, Green, Medium”. The second tells a machine what it is, who it is for and which variant it describes. The first only works if the shopper already knows the brand.

3. Attributes: the details shoppers ask for

Attributes are where most requests are actually answered. Size, color, material, gender, age group, dimensions, compatibility and capacity are typical examples.

Shoppers ask agents very specific questions, and agents can only answer them with attributes that exist in your data. A product whose material appears only in a paragraph of description is easier to skip than one with a filled material field.

4. Price and availability, in real time

Stale stock and prices are among the main reasons AI shopping has disappointed shoppers. Feeds or catalogs that update as often as your inventory changes give agents data they can trust.

The rule of thumb is simple: the data should never be older than the speed at which your stock moves. A store that sells out of items within hours should not rely on a once-a-day file.

5. Shipping and returns

Delivery time, shipping cost and return policy are part of the decision. An agent asked for something that “arrives by Friday” needs delivery data to include you.

These fields also need to match what the shopper sees at checkout on your site. A mismatch is one of the fastest ways to lose trust.

6. Structured data on your product pages

Product structured data (schema.org) on each page helps assistants that read the web confirm what your feed says. Price and availability on the page must match the feed.

Treat the page and the feed as two statements of the same facts. When they disagree, you give a machine a reason to doubt both.

7. Trust signals

Business information, policies and reviews help an agent recommend a store with confidence. Make sure your return policy, contact details and shipping policy are easy to find and consistent with the feed.

What should “good” look like, field by field?

Use this table as an audit sheet. It shows what to check, why it matters and the mistake we see most.

Which data problems are easiest to spot?

Most problems leave a visible symptom in your data. Look for these.

• Many products with identical titles. Variants were not given distinguishing attributes, so agents cannot tell them apart.
• Blank columns in your export. An empty material, size or color column across hundreds of rows means shoppers’ questions cannot be answered.
• Price differences between feed and page. Pick ten products at random and compare. Any gap points to a sync problem.
• Products showing in stock that are not. Check your last few cancelled orders for the cause.
• The same SKU under different brand strings. Sort the brand column and read the unique values.

How can you run this audit yourself?

You can run this in an afternoon on a sample, using only a spreadsheet and your store admin.

1. Export your catalog or feed into a spreadsheet.
2. Sort the brand column and list every unique value. Merge duplicates.
3. Filter for rows with no GTIN and decide which genuinely have none.
4. Read twenty titles aloud. If you cannot tell what the product is, neither can an agent.
5. For each product type, list the five attributes a shopper would ask about. Count the blanks.
6. Compare price, stock and shipping for ten random products between the feed and the live page.
7. Open a product page and check its structured data against the feed.
8. Rank the problems by how many products each one affects, and fix the biggest first.

How does the audit work in a hypothetical example?

This is an illustrative case, not a client.

Imagine a store with a few thousand SKUs of outdoor clothing. The audit shows that titles lead with model names, and the material and gender columns are empty for about half the rows. Prices match, but stock updates only once a day.

The biggest payoff is the attribute gap, because it affects many products and costs nothing but data work. Titles come second. The stock sync comes third, and matters most if popular sizes sell out quickly. Fixing in that order puts effort where the data is weakest.

When is this not the priority for you?

Product data is rarely wasted effort, but it is not always the first thing to fix.

• If your store sells custom or made-to-order items with no identifiers, your attribute and title work matters more than GTINs.
• If your catalog is tiny, a manual review may beat any tooling.
• If your site cannot be reached by crawlers at all, fix access before polishing fields.

What should you do next?

If your Google Merchant Center feed is approved and complete, you have most of this already. From there, one catalog can serve every channel: see one product feed for every AI channel, getting into ChatGPT Shopping and the readiness checklist.

Marketplace sellers can see how the same data requirements show up on each channel in Walmart Marketplace vs Amazon catalog requirements. To compare channels, read Google vs Microsoft Merchant Center vs ChatGPT Shopping.

How does Avafa work on product data?

Avafa is a consultancy based in Spain that works remotely. We audit the catalog along the fields above, fix the structure, and build one master feed that serves each channel. We do not promise placement in any AI answer, because that is decided by the platforms.

The work is delivered through our Google Merchant Center agency service, and the same data supports our ecommerce SEO work.

FAQs

Do AI agents read product descriptions?

Yes, but structured fields are easier for them to use reliably. Put key facts in attributes, not only in the description.

Do I need to change my website for AI agents?

Mostly no. Make sure product pages are accessible to crawlers and carry structured data that matches your feed.

What if my products have no GTIN?

Some products legitimately have none, such as custom or handmade goods. Use MPN and brand where you can, and fill the other fields thoroughly.

Which field should I fix first?

Fix whatever affects the most products. For many stores that is attribute completeness, followed by titles and then update frequency.

How often should price and stock update?

As often as they change in your store. If stock moves hourly, a daily file will show agents outdated information.

Does structured data replace the feed?

No. The feed supplies the catalog to platforms directly, and the page markup confirms it. They should always agree.