How Product Feeds Need to Change for AI Shopping Agents

What Changes in a Product Feed for AI Agents?

Feeds need clearer titles, richer descriptive attributes, stable identifiers and fresher price and stock data, because AI assistants read fields to answer questions instead of ranking keyword matches. The core structure is familiar. The bar for accuracy and detail is higher.

A traditional shopping listing served an ad auction. An AI assistant serves a question with constraints: "under 100 dollars," "fits a 13-inch laptop," "vegan," "ships by Friday." Every constraint is something the feed either answers or does not. Missing fields do not just lower a score. They remove the product from consideration for that question.

From Keywords to Facts

Old feed habits reflected a bidding mindset. Titles were packed with terms to match searches, descriptions were marketing copy, and attributes were filled in only where a platform demanded them. Agents reward the opposite.

A good title reads like something a knowledgeable shop assistant would say: "Merrell Moab 3 Waterproof Hiking Shoes, Women's, Size 8, Gray." A bad one reads like a search query.

Stable Identifiers and Product Matching

An agent that compares products across stores must decide whether two listings are the same item. Identifiers make that possible. Provide a GTIN (barcode number) where one exists, plus a manufacturer part number and brand. Keep your own item ID stable over time so the same product does not appear as a new listing every time you re-export.

Variants deserve care. A jacket in five colors and four sizes should be one product with clearly labeled variants, each with its own ID, price and stock. Merging them into a single record hides which sizes are available, and splitting them into unrelated items scatters reviews and confuses matching.

Answer the Questions Shoppers Actually Ask

Look at customer emails, chat logs and reviews. The same questions repeat: does it fit this model, is it machine washable, what is it made of, how long is the battery, is it safe for kids. Each recurring question should have a home in the feed.

• Compatibility. Models, sizes, connectors and versions, written as plain values.
• Material and composition. Percentages for fabrics, ingredient lists for consumables.
• Dimensions and weight. Units stated clearly.
• Use case. "Suitable for rainy weather commuting" is more useful than "premium quality."
• Care and safety. Washing instructions, warnings and certifications, only where accurate.
• Audience. Age range and gender fields where the platform supports them.

Some platforms have started asking for more descriptive and conversational-style attributes than classic shopping feeds did. Specifications differ and are changing, so check current documentation rather than copying a field list from an article, including this one.

Price, Stock and Shipping Have to Be Current

Nothing damages agent trust faster than recommending an item that turns out to be sold out or a different price. Treat stock and price as live data.

1. Connect the feed to inventory, not a spreadsheet. Export from the system that knows what is actually on the shelf.
2. Update often. Some platforms accept refreshes as frequently as every 15 minutes according to their documentation and industry reports. Even hourly is far better than nightly for fast-moving items.
3. Handle sell-outs immediately. Mark items out of stock or remove them the moment you can no longer ship.
4. Include shipping and delivery estimates. Cost and time are common filters in shopping questions.
5. Log sync failures. A silent export failure can leave a feed stale for days.

The Feed, the Page and the Checkout Must Agree

Agents cross-check. They may read your feed, then load the product page, then reach the checkout. If a price differs, a variant is missing or a shipping fee appears late, they will notice, and platforms may disapprove the item. Guard consistency in three places.

• Feed data sent to shopping platforms.
• Structured data on the page. Product, Offer, AggregateRating and related schema.org markup in JSON-LD. Google documents which properties it uses.
• Checkout and cart. The checkout flow should not add costs the earlier stages never mentioned.

A minimal example of page-level markup:

{
  "@context": "https://schema.org",
  "@type": "Product",
  "name": "Trail Runner Waterproof Jacket",
  "sku": "TR-JKT-014",
  "gtin13": "0123456789012",
  "brand": {"@type": "Brand", "name": "Example Outdoors"},
  "offers": {
    "@type": "Offer",
    "price": "89.00",
    "priceCurrency": "USD",
    "availability": "https://schema.org/InStock",
    "url": "https://www.example.com/jackets/tr-jkt-014"
  }
}

Schema.org also has types for return policies and shipping details, which Google supports. Publishing them in structured form helps machines read your terms without parsing paragraphs, a point covered in how to create machine-readable policies for AI agents.

Ratings, Images and Honest Claims

Reviews influence recommendations, so feed ratings only if they are genuine and current. Fabricated or copied reviews create legal and platform risk, and they tend to fall apart when agents compare across sources. Use clear, well-lit images without overlaid text, since many platforms restrict promotional overlays, and keep image URLs stable and fast. Make claims you can support. If a feed says "waterproof" and reviews say it soaks through, agents that read both will notice. For how these signals feed into choices, read how AI shopping agents choose what to buy.

A Practical Audit You Can Do in One Afternoon

1. Export the feed and sample 50 items, weighted toward best sellers and newest products.
2. Compare each to its live page. Check price, stock, title, images and variants.
3. Score completeness. Count how many recommended attributes are filled for each item.
4. Read the titles aloud. Would a person say this? If not, rewrite it.
5. Check platform diagnostics. Merchant Center and other dashboards list disapprovals and warnings. Fix the most common first.
6. Verify identifiers. Confirm GTINs are valid and not reused across different products.
7. Time your refresh. Change a price and note how long it takes to appear everywhere.

Repeat quarterly, and after any change to your catalog system. For the discovery stage that precedes all this, see AI search for products explained.

One Catalog, Several Destinations

Most merchants feed more than one place: Google Merchant Center, marketplaces such as Amazon and eBay, social shopping surfaces and, increasingly, assistant-specific programs. Each has its own field names, size limits and review rules, and each changes over time. Building a separate spreadsheet for every destination guarantees drift.

The sturdier pattern is a single clean catalog with a mapping layer. Store your data once in its richest form, with full attributes, identifiers and policy fields. Then generate each destination's feed by mapping fields and trimming what a platform does not accept. When a platform adds a field, you extend the mapping, not the catalog. Feed management tools do this for a fee, and a small script can do it for a modest catalog. Either way, treat each platform's own documentation as the source of truth for what it requires this quarter.

Bundles, Made-to-Order Items, Subscriptions and Regions

Real catalogs are messier than the examples in documentation.

• Bundles. Give each bundle its own ID and title, state the contents in the description, and price it as it is sold. Do not list a bundle and its parts as if they were interchangeable.
• Made-to-order goods. State the lead time in shipping fields and in the description. An agent promising delivery by Friday for something built over three weeks creates a refund.
• Subscriptions. Make the billing interval and the total recurring cost obvious. An assistant that quotes only the first-month price will damage trust when the real cost appears.
• Multiple regions. Currency, tax handling and shipping vary by country. Keep separate feeds or feed rows per market with the right price and availability for each, and align them with any regional versions of your site.
• Unavailable in some places. State regional restrictions clearly, so an agent does not recommend something a shopper cannot legally receive.

Handling these cases well is unglamorous work, and it is where feed quality really shows. A feed that treats every product as a simple in-stock physical item will mislead agents on the products that are not.

Who Owns the Feed?

Feeds usually break because nobody owns them. Marketing edits titles, operations changes stock, and finance updates prices, all in different places. Decide on a single source of truth, ideally a product information system, and generate the feed and the page markup from it. Assign someone to watch diagnostics weekly. A feed that is simple, accurate and boring beats a clever one that drifts out of sync, and that is exactly the kind of data an assistant can rely on.

Frequently Asked Questions