How we measure whether AI assistants recommend you

The short version. Every number we report comes from this, and the same rules apply every round.

Version 1.5 · 8 November 2026 · Satayu Isariyaphorn · the full rulebook

1. 40 buyer questionsWritten with you for your market, then frozen
2. 4 assistantsChatGPT, Claude, Perplexity and Google's AI
3. 3 days, 480 answersEach question asked fresh, on separate days; every answer kept
4. CountedWho is named, who is named first, which sites are cited
5. FixedPages that answer the question, listings, indexing
6. Re-measuredDay 30, 60 and 90

Same questions and rules every round, so a change in the count is a change in the answers, not in our wording.

Five rules, and why

1. Three separate days. The same question asked on two different days named the same set of companies only 55% of the time, across 1,800 of our own answers. One run is a guess, so each question is asked on three days and counted together. The data.
2. Frozen questions. The questions are written with you before the first measurement and never reworded, down to the punctuation, and each one names the buyer’s country. When the count moves, the answers moved, not our wording.
3. A stranger’s view. Each question goes bare, in its own logged-out session with no memory and nothing wrapped around it, and the search setting is recorded beside every answer. That is what a buyer who has never heard of you is told.
4. What isn’t counted. Questions that ask the assistant to compare two named rivals are reported on their own line, and an assistant run without web search is reported beside the number, never inside it. The headline is always named answers out of all answers.
5. The limits. We count naming, not sales. We can’t see what an assistant tells someone with months of chat history. And providers change their models mid-month, so the model behind every answer is recorded and a jump can be checked.

Check us, or do it yourself

Every rule in detail is in the full rulebook, and the method, our cost notes and our own frozen questions are on GitHub under CC BY 4.0. A real result measured this way: Vilasinee Oil, 1 to 9 of 10 ChatGPT answers. To see where you stand, ask for the free check.