How we measure the AI shelf

1. Shelves and questions

A shelf is a product category, such as LED face masks or menopause telehealth clinics. For each shelf we write a small set of questions in the words real buyers use, like "What's the best LED face mask for wrinkles?" A person writes every question. None are generated.

The index uses generic questions, because that's what most buyers type. A smaller set of questions from women over 40 ("I'm 52 and my skin is getting thinner...") is measured separately and reported in paid reports as the 40+ gap. It never changes the public index.

2. Sampling

AI answers change from one ask to the next. Identical questions rarely return the same list of brands, so one answer tells you very little. We ask each shelf's questions about a hundred times on every assistant, every edition, and show the number of answers next to every figure.

3. Assistants

We sample 6 assistants: ChatGPT, Gemini, Google AI Overviews, Google AI Mode, Perplexity and Copilot. Each is sampled through its consumer app, as shoppers see it, using a data-collection provider.

4. Reading the answers

A language model reads each answer and lists the brands it names, in order, and whether it recommends them, mentions them or warns against them. Stores, publications and ingredients are left out. Brand spellings are merged through a reviewed alias list, so "Brand Co" and "brandco.com" count once. A person checks a sample of every edition.

5. The numbers

• Brand-carrying rate: the share of an assistant's answers that recommend at least one brand.
• Visibility: of the answers on a shelf that recommend any brand, the share that recommend this one. "Visibility 60% on ChatGPT" means 60 of every 100 brand-carrying ChatGPT answers on that shelf named the brand.
• Rank: brands ordered by visibility. Ties go to the brand named earlier in the answer on average.
• Combined view: visibility averaged across assistants, counting zero where a brand doesn't appear.
• Warnings: brands an answer tells people to avoid are counted apart from recommendations and never raise a rank.

6. Why your own test looks different

One chat is one draw. Your account history, location and the exact wording you use all shift the answer. The index shows what an assistant recommends most of the time, which a single chat can't.

7. What we never do

• No pay-for-placement. The index is computed the same way for every brand, customer or not.
• No vendor-supplied numbers. Everything comes from our own sampling.
• No product verdicts. We measure what AI assistants recommend, not which product is better.

8. Change log

• Sample edition: placeholder data used to build the site. Not a measurement.