How we check whether AI assistants name you

Buyers ask AI assistants which product to choose. We ask the assistants the same questions and count who they name. This page shows every step and both formulas, so you can recount any score yourself.

Six assistants and how we ask them

Each question goes to each assistant as a new conversation — no earlier questions, no chat history from us. We ask in English. When an assistant does not answer, we ask again. An answer that never comes does not go into the score.

Which questions we ask

  1. We read your site: what you sell, to whom, and who you are compared with.
  2. We gather real demand in your category: Google's "People also ask" questions, related searches, keyword volumes and Reddit threads.
  3. A language model writes 20 buyer questions from that data.
  4. No question names your brand or anyone else's. A question that still names a brand is dropped.
  5. We ask ChatGPT each of the 20 questions once. We keep the 10 where ChatGPT names the most vendors: in those, the assistant chooses who to name.
  6. A question ChatGPT did not answer goes to the end of the list. On a tie, we prefer a different type of question, then the one with more searches.

The diagnostic uses 30–50 questions.

How many times we ask

What counts as a finding

What we mark in each answer

A language model marks the answers. Every "named" has to rest on a word-for-word quote from the answer. When the quote is not in the answer, our own match on your name and domain decides.

Mention score

Mention score = 100 × answers that name your brand ÷ answers received. An answer we could not get is left out of both numbers.

We round the result to the nearest whole number; a half goes up.

Next to the score: the same count for each assistant, for example 7/10 — named in 7 of the 10 answers received.

Example: you are named in 32 of 60 answers received. 100 × 32 ÷ 60 = 53.3, so the score is 53.

Site score

Site score = access (60) + what the bot sees (40).

Access: 5 bot families — OpenAI, Anthropic, Perplexity, Googlebot, Bingbot — up to 12 points each. 12 if robots.txt lets in the family's search and user-facing bots and the page, fetched with their own User-Agent, returns at least 80% of the text a browser gets, with no challenge page; 6 if only robots.txt lets them in; 0 otherwise.

The bots in each family:

Bots that only collect training data (GPTBot, ClaudeBot) are shown but not scored: blocking them does not take you out of the answers.

What the bot sees, in the raw HTML without JavaScript: 10 — main text and an H1; 15 — the H1 and the first ~150 characters say what the product is and who it is for (7 if only one of the two); 10 — a price on the homepage or the pricing page; 5 — the page can be indexed.

Each item in detail:

What we do not score

Schema markup and llms.txt are not scored: we found no measured effect.

The range in the diagnostic

In the diagnostic, the mention score comes with a range. We get it by resampling questions: we draw random sets of the same questions many times and recount the score each time. The unit is the question, not the answer.

Why not answers: ten runs of one question look alike. Counted as separate answers, they would give a range that looks far tighter than it is — about ±2 points at 1,800–3,000 answers.

Limits

Check it yourself

Back to the homepage