How Ecommerce Brands Can Track How LLMs Describe Their Products

Why track how AI describes your products?

Shoppers increasingly ask assistants about products before visiting a store: "Is the X200 waterproof?", "What sizes does it come in?", "Is it worth the price compared with the Y300?" If the answer states the wrong price, a discontinued colour or a feature the product lacks, you lose the sale or get a return. Assistants also summarise reviews into a reputation, which can lag behind product improvements. Tracking tells you what shoppers hear before they reach you.

What do you need to set up tracking?

A truth table from your catalogue

Export the facts engines are likely to state, one row per product or variant:

Your product information system or feed is the source. Refresh the table whenever prices or products change.

A product question set

For each important product, write six to ten questions in shoppers' words:

• "How much is [product]?"
• "Is [product] [key feature, such as waterproof]?"
• "What sizes or colours does [product] come in?"
• "Is [product] good for [use case]?"
• "[Product] vs [competitor product]?"
• "What do reviews say about [product]?"
• "Where can I buy [product]?"

Include a few category questions, such as "best waterproof trail shoes under £100", to see whether and how the product is described when it is recommended.

How can I audit differences between my live store data and what LLMs say about it?

1. Run each question three times in the assistants your customers use, such as ChatGPT, Google's AI Mode, Gemini and Perplexity, with search on.
2. Extract each factual claim from the answers into rows: product, attribute, value stated, engine, date and cited URL. An AI model can help extract claims at scale, but check a sample by hand.
3. Match each claim to the truth table and mark it correct, wrong, outdated or unverifiable.
4. Record the source cited for each wrong claim.
5. Summarise by attribute and product: price accuracy, specification accuracy, availability accuracy.

The result is a discrepancy report such as: "Price stated correctly in 14 of 30 answers. 11 of the errors quoted a marketplace reseller's price, 3 quoted last year's price from an old review, 2 had no source."

What kinds of discrepancies turn up, and where do they come from?

How poor product content leads to wrong answers is covered in how poor product content can lead AI to give shoppers incorrect information.

How often should the audit run?

Run the full audit monthly for your priority products, and a short check within a week of any price change, launch, discontinuation or policy change. Those are the moments when AI answers are most likely to be wrong, because old pages and listings still say the previous thing. Keep the question set fixed between runs so results are comparable, and add new questions only at the start of a quarter. A quarterly summary for the wider team should show accuracy by product group, the three most common error types, and which fixes worked.

How do you track descriptions, not just facts?

Accuracy is only half the picture. Also record how each answer characterises the product:

• Positioning words: "premium", "budget", "best for beginners", "durable".
• Recurring praise and complaints attributed to reviews.
• Recommendation role: recommended, listed as an option, or advised against.
• Comparisons: which competing products it is compared with, and who wins.

Track these monthly. A shift from "comfortable but runs small" to "comfortable, true to size" after a sizing change shows engines picking up your update. Persistent negative themes point to product or review issues rather than content.

Which metrics should you report?

Report by product group, and prioritise best sellers and high-margin lines. Tracking visibility at SKU level is covered in how to track product visibility by SKU in LLM answers.

How do you fix what engines get wrong?

1. Fix your own pages first: clear, current prices, specifications and availability in the HTML, with matching feed data.
2. Send the same data to the shopping feeds engines read: Google Merchant Center for AI Mode and Gemini, and OpenAI's merchant programme at chatgpt.com/merchants for ChatGPT shopping, with prices and availability updated as often as your store changes.
3. Correct retailer and marketplace listings that engines cite, through your channel partners.
4. Ask publishers of outdated reviews or comparisons to update, with the new facts.
5. Mark changes clearly, for example "Updated for 2026: now true to size", on product pages.
6. Collect recent reviews that reflect the current product.
7. Re-test after four to eight weeks and record which fixes changed answers.

Keeping information consistent across sources is covered in how to make a brand's information consistent across AI-visible sources.

What do people ask about tracking AI product descriptions?

How many products should I track?

Start with your top 20–50 by revenue or margin, plus any with frequent returns or complaints. Expand once the process works.

Can AI tools extract claims from answers automatically?

Yes, with care. Use them to speed up extraction, then check a sample, since extraction can misread numbers and product names.

How often do answers change after I fix a listing?

Search-based answers can change within weeks of the corrected pages being recrawled; answers from model memory take longer.

Should I track marketplaces as well as my own store?

Yes. Engines often cite marketplace listings, so their accuracy affects what shoppers hear about your products.