Is Your Product Data Ready for AI Shopping Agents? A 12-Point Audit
Agents cannot read your marketing copy. They read your data, and gaps in it quietly remove products from consideration.
An agent-ready product record is machine-readable (structured data and feeds, not text rendered by a script), complete (identifiers, variants, attributes, price, stock, shipping and return policy) and current (price and stock updated within minutes). Audit 12 points across identity, content, commerce and delivery on a sample of 20 products, then fix the failures that block products from appearing at all.
A shopper reads your product page. An AI shopping agent reads your data. It cannot see the lifestyle photo, it skips the clever headline, and if your price lives in a script that runs in the browser, it may not see the price at all. Products an agent cannot read do not get recommended, and nobody tells you which ones they are.
Why this matters now
Shopify's enterprise team reports that AI-referred orders on its platform grew nearly 13 times year over year in Q1 2026, and that AI-referred visitors convert at nearly 50% higher rates than organic search visitors. Gartner projects that 20% of digital commerce transactions will run through AI platforms by 2030 (both via Shopify). Treat the forecast as a forecast. The data work it asks for pays off in search and on your own site anyway.
What an agent needs from a product record
Agents pull from three places: a product feed you send, markup on your product pages, and live endpoints for price and stock. Platforms differ. OpenAI's ChatGPT shopping takes a feed you push over HTTPS in CSV, TSV, XML or JSON, with updates accepted as often as every 15 minutes (Lengow's summary of the spec). Google's route runs through Merchant Center with enriched attributes and a real-time inventory endpoint. Perplexity works through partners such as Feedonomics. Field names and requirements change, so check each platform's current spec before you build.
What does not change is the principle: one trusted record, sent the same way everywhere.
The 12-point audit
Each check below takes minutes on a sample. Run it on 20 products: a few best sellers, some with many variants, and some on sale.
| # | Check | Quick test |
|---|---|---|
| 1 | Identity. A stable, unique ID for every product and every variant | Change a title and confirm the ID stays the same |
| 2 | Identifiers. GTIN, UPC or MPN wherever they exist | Count products missing them |
| 3 | Variants. Real options grouped under one parent, not listed as duplicates | Search your catalog for one product and count the duplicates |
| 4 | Literal descriptions. Materials, dimensions, compatibility, what is in the box | Read three descriptions. Could a stranger answer a fit question from them? |
| 5 | Specific taxonomy. “Men’s insulated winter boots”, not “footwear” | List the category path for 20 products |
| 6 | Attributes with units. Colour, size, material, weight, filled in and consistent | Count empty attribute fields per product |
| 7 | Plain text. No HTML, template logic or script-only content in key fields | View the feed field, not the rendered page |
| 8 | Price. Amount and currency, with sale price handled the same everywhere | Compare price in feed, page and cart |
| 9 | Availability. Stock status and quantity by variant, updated in minutes | Sell the last unit and time how long the feed takes to change |
| 10 | Shipping and returns. Cost, delivery estimate and return policy in machine-readable form | Find the return window without reading a page |
| 11 | Server-rendered markup. JSON-LD Product data in the HTML the server sends | View page source and search for the price |
| 12 | One source of truth. Feed, page and API agree | The three-way match in the diagram above |
Score each check pass or fail per product, then count how many products fail each one. As a starting rule, treat any check that fails on more than 5% of your sample as a blocker. That threshold is ours, not a platform requirement.
What to fix first
Start where a failure removes a product from consideration or puts a wrong fact in front of a shopper. That means identity and variants (checks 1 to 3), then price and availability (8, 9 and 12). A beautiful description does nothing for a product whose stock status is a day old. Content checks come next, since literal descriptions and complete attributes help both agents and your own site search.
Policies are the check teams forget. An agent asked “can I return this?” answers from whatever it can find. If the return window lives in an unindexed PDF, you will see a confident wrong answer, and that is a failure you can test for.
Where search comes in
The same clean data is what makes retrieval work. An assistant can only recommend products its search step finds, and search can only match attributes that exist. We cover that connection here.
Common questions
Do I need a separate feed for each AI platform?
Usually the platforms have different requirements. ChatGPT takes a merchant-pushed feed, Google runs through Merchant Center with enriched attributes and real-time inventory, and Perplexity works through partners such as Feedonomics. Build one clean record and map it to each platform’s fields.
How fast does stock need to update?
The ChatGPT feed accepts updates as often as every 15 minutes, which is a reasonable target for price and stock. Slower updates mean agents can quote a price or stock status that is already wrong.
Does page markup matter if I send a feed?
Yes. Feeds cover the platforms you send them to. Page markup helps assistants and search engines that read your site directly. Keep both consistent, because a feed and a page that disagree make your data less trustworthy.
What should I fix first?
Identifiers and variants, then price and availability accuracy. Wrong or stale facts do more damage than thin descriptions.
Sources
- Shopify: agentic-ready product data
- Lengow: ChatGPT product feed specification
- Gladly: preparing your product feed for AI agents
- Alhena: prepare your store for AI shopping agents
Last reviewed October 16, 2026. Figures from vendors and roundups are marked as such and change often. Check current prices and specifications before you rely on them.