SEO, GEO and AEO consultants: we're looking for partners. See the program →
Metrisque
ARTICLE17 Nov 2026 · 7 min read

What does ChatGPT read on a website product page?

A orange product page split between a stacked archive on the left and a search panel reading it on the right

ChatGPT knows a product in two ways. From training, it holds a memory of pages it saw before its cutoff, which changes only when the provider updates the model. With web search on, it fetches live pages and reads their visible text, and it may draw on product data from feeds and listings. A store can change what web search reads in days, while memory changes slowly.

Product knowledge comes from memory and web search

When a shopper asks ChatGPT, Gemini or Claude for a product recommendation, the answer draws on one or both of two sources. One is the model’s memory: patterns learned during training from a large slice of the public web, frozen at a cutoff date, and the other is web search, where the assistant runs a search when the shopper asks, opens some of the results and reads them before answering.

The two sources differ. They see different things and update on different clocks. Memory holds how the wider web described your brand and products months ago, weighted toward what pages said often and consistently. Web search shows what your pages, retailer listings and reviews say today, filtered by whatever the search returns for that question. A product can be strong in one and absent from the other.

Model memory is broad and old

A model’s memory is broad, blurry and months old. It tends to know well-established brands, popular products and the categories they belong to, because those appear on thousands of pages. It knows smaller brands less well and may confuse them with similarly named companies. Metrisque’s Brand Recall measure asks the models directly whether they know a store or brand name, and scopes the verdict to the exact name asked, because a model can know a full name and confuse a short one.

A store cannot edit memory. What it can do is make sure the pages the next model will learn from describe it consistently: the same brand name, the same product types and the same category on its own site, on marketplaces and in coverage. That work pays back at the next model update, which the store does not schedule and cannot predict.

Web search reads visible page text

Web search is different. Here a store has direct control. When an assistant opens a product page, it reads the text the page shows: the title, the description, specification tables written as text, visible reviews and breadcrumbs. Text inside images, PDFs or elements that load only after a click may go unread. Structured data, such as Product schema, gives a machine-readable statement of type, price and availability, though Metrisque has not measured how much assistants rely on it.

The description is the page element with the most measured evidence behind it. In a synthetic test on stored products, models classified products into the right category 150 of 150 times from clear descriptions and 116 of 150 times from vague ones. A page that reads plainly through web search gives the assistant something firm to repeat. A vague one leaves it to guess, and a guess can file the product under the wrong category.

Feeds and listings are a third path

Feeds matter too. Product feeds, such as the ones a store sends to shopping channels, and marketplace listings give assistants structured product data outside the store’s own pages. Metrisque has not measured how directly any one assistant uses a store’s feed. For a store, the value of a clean feed is that it states the product type, category, price and availability plainly and consistently with the product page.

The rule for feeds matches the rule for pages: say what the product is, in the category a shopper would name, and make sure every source agrees. A feed that files a product under a different category from the page gives the assistant two stories.

Source of knowledge When the assistant uses it What a store can change
Model memory (training data) Answers without web search, and as background when web search is on Consistent brand and product descriptions on its own site, marketplaces and coverage, ahead of future model updates
Web search: product page text Answers with web search on, when the search returns the page and the assistant opens it Title, first 100 words, specifications as text, visible reviews, breadcrumbs
Web search: third-party pages Answers with web search on, when the search returns reviews or comparisons Accurate listings and reviews on the sources the assistant already cites
Structured data and feeds When the assistant or its search layer reads machine-readable product data Product schema and feed fields that match the visible page

Table 1. Stores control their own pages and feeds, while model memory and third-party coverage sit outside their direct control.

The memory-search gap

Metrisque’s Question Finder runs each shopper question on the model’s memory and again with web search, and records who gets recommended and in what position. Comparing the two shows what we call the memory-search gap: the difference between how a product fares when the model answers from training and when it reads live pages. The gap tells a store where to put its effort.

A stored case shows how this works. One store in Metrisque’s stored runs, which we will not name, received no first-place mentions in 11 saved reads that covered both memory and web search. The models recalled the store by name, so memory was not blank. Eleven reads cannot prove a cause, but a miss that holds in both modes points away from a passing search result and toward category filing, descriptions or third-party coverage as the first things to check.

Memory only: fix the pageBoth: protect and watchNeither: filing, then coverageSearch only: build coverageFound by web search →not foundfoundKnown in model memory →not knownknown
Figure 1. A store’s quadrant decides its first move: page fixes pay back fastest for products found by web search, and coverage work is the only route for products absent from both.Source: Qualitative framework. Metrisque Question Finder runs each question on memory and on web search; quadrant actions are reasoned, not measured.

Each quadrant has its own first move

A product that is strong in both memory and web search needs little work. The risk there is a page change that breaks what is working, so record a baseline before any redesign and compare after.

A product known in memory but weak in web search usually has a page problem: the model knows the product from older coverage, but when it reads the current page it finds something vague or contradictory. Run the description check (does the first 100 words name the type, the buyer and the specifications that set it apart?) and look at whether Category Fit marks the product Outside. In Metrisque’s latest read for 53 stores, 13 of 923 products were Outside, meaning their Gemini similarity to the expected category fell under 0.40.

The reverse case is common too. A product that does well through web search but is weak in memory has pages that work and a reputation that has not caught up. Keep the pages strong and invest in accurate coverage on the review sites, marketplaces and comparison articles the assistant already cites, so the next model update has more to learn from. A product absent from both needs the full sequence, and that sequence starts with category filing and descriptions.

A quick test you can run today

You can see the split yourself in a few minutes. Pick one of your best-selling products and one unbranded buying question it should win, such as “best 12-inch cast iron skillet for induction”. Ask ChatGPT the question in a fresh chat with web search off, note whether your product appears and where, then ask again with web search on and note the sources it shows. Repeat both three times. A product that appears only with web search has pages doing the work, a product that appears only from memory has a weak current page, and a product that never appears needs a category filing check first.

A caution on similarity scores

Scores can mislead. Category Fit uses similarity to flag products that look misfiled, and that flag is a useful pointer. In Metrisque’s synthetic tests, though, similarity scores did not predict what a model recommends. A product can sit well inside its category and still lose the recommendation to a rival with stronger coverage. Treat any page check as a way to find causes, and judge the result by a recommendation read on the same questions, in both modes, after the page has been recrawled.

Our guide to generative engine optimization for online stores explains how this split fits with titles, descriptions, breadcrumbs and structured data in a whole catalog.

Questions people ask

What does ChatGPT read on a website? With web search on, ChatGPT reads the visible text of the pages it opens, such as titles, descriptions, specification tables and reviews. Text inside images or content that loads only on click may go unread.

Does ChatGPT use my product feed? Metrisque has not measured direct feed use by any one assistant, so treat the feed as a second statement of the facts on the page. Keep feed fields consistent with the product page so every source names the same type and category.

What is product feed optimization for AI? It is making sure each product’s type, category, price and availability appear plainly and consistently in the feed. A feed that disagrees with the page gives an assistant two different descriptions of the same product.

Sources

  • Metrisque stored runs: Brand Recall and Question Finder case (one unnamed store, 11 reads, memory and web search). description clarity study (300 classifications). Category Fit latest read (13 of 923 Outside, 53 stores). synthetic similarity tests. Read on 16 November 2026.
  • Semrush and DataForSEO keyword estimates, US. Read on 16 November 2026.

A free Metrisque diagnostic shows which of the four quadrants your best-selling products sit in, on the same buying questions, with memory and web search side by side.

FREE DIAGNOSTIC

See whether ChatGPT, Gemini and Claude recommend your products, and which fix moves them first.

Start free diagnostic

Keep reading

Orange product shelf with a search panel above it, linked by thin lines to three products on the shelf
GUIDE
Generative engine optimization for online stores
Two overlapping orange circles, a ranked list and a product shelf, with three products in the overlap
ARTICLE
GEO vs SEO for a product catalog
Five orange steps rising to a product box on the top step, with a search panel shape above it
ARTICLE
How to rank in ChatGPT as an online store
A orange product page outline with its opening text lines shaded above a small spec table
ARTICLE
LLM SEO for product pages: what to write and where