Readable is not recommended

Making your site readable for AI and getting your products recommended by AI are two different jobs.

Most advice on AI search starts with the website. Let the AI crawlers in, put the important content in plain HTML, add structured data. All of that is worth doing, and none of it answers the question a shopper asks. When someone asks which running shoe to buy, AI picks a product. How easy your site is to read does not decide which one.

What readable gets you

A readable site keeps your product eligible and gets the facts right. For that, AI assistants must be allowed to fetch your pages. ChatGPT’s crawlers do not run JavaScript, so a price, a stock status or a review score that only appears after a script runs is invisible to it (Vercel and MERJ, December 2024). When AI states the wrong price or specification for your product, this is usually where to look.

This is a requirement. Much of the advice on AI search stops here, but it only settles whether you can be found. Being chosen depends on other things.

For a “what should I buy” question, AI assistants lean on what others say about the product. A University of Toronto preprint (September 2025) looked at ChatGPT’s answers to “best” product queries in three categories (cars, consumer electronics, software; US and Canada). Earned media such as reviews, tests and editorial made up 69 to 92% of the sources cited. In Google’s top ten results for the same queries the share was 32 to 54%.

The brand’s own pages play a smaller part. They supply facts like specifications and price, while tests, best-of lists, reviews, YouTube and Reddit make the case for the product.

Our own data points the same way. On 27 October 2026 we asked ChatGPT and Google AI Overviews Dutch shopping questions about 50 fashion categories. ChatGPT placed brands among the first five named in its answers 1,948 times, and in 41% of those a page from the brand’s own site was among the sources it linked or searched. Google AI Overviews placed brands 1,617 times, with the brand’s own site among the sources in 9%. In most of these cases the brand was named among the first five without its own site among the sources.

The offer counts as well. In randomised tests with AI shopping agents in mock shops, a higher rating, more reviews and a lower price changed which product the agent chose (the ACES study, August 2025). Across three models, a rating 0.1 star higher lifted a product’s chance of being picked from 10% to between 15 and 20%. Price, availability and ratings are product and commercial questions, and no change to the website fixes them.

Being known matters too. The model decides which brands to look for partly from what it learned in training. An academic audit of 37,000 runs (May 2026, not specific to the Netherlands) of AI answers that use web search found that about half of the specialist and regional brands never appeared at all, while category leaders almost always did.

Why a page on your own site is not enough

An article on your own site saying your product is the best is rarely what AI assistants retrieve for a buying question. They look for what third parties say.

Much of the work sits outside the website, in the tests, best-of lists, reviews and videos that answers already draw on, and in the product and the offer themselves. They are harder to win than your own site, and the recommendation is made there.

No one controls what AI recommends, and most published studies show correlation, not proof. We measure, for each product, how the answers treat it, and improve the odds on the levers you own.

Where this leaves the website

Keep it readable: a product page that AI assistants can fetch, with the right price, stock and specifications in the HTML, is the base. When the question is why a competitor’s product is picked and yours is not, the answer is often not on your site.

How we measured it

The figures about brand sites come from the study described in “Who owns the AI shelf”: 300 Dutch product questions about 50 fashion categories, put to ChatGPT, Gemini and Google AI Overviews on 27 October 2026, each asked three times. Gemini is left out here. We counted the 900 ChatGPT answers and the 438 Google AI Overviews to those questions, 1,338 of the 2,238 answers that study uses. Like the study, it covers one country, the Netherlands, and one sector, fashion. A brand counts as having its own site among the sources when a source domain matches the brand’s name. Retailers and store brands are not counted as brands. A brand placement is one brand among the first five named in one answer, so one answer can hold up to five.

Want to know how AI is recommending your products?

The free AI Shopping Audit puts real shopper questions to ChatGPT, Gemini and Google AI Overviews, then walks you through the answers.

Free audit