How we measure AI search visibility

We ask ChatGPT the questions your buyers ask, with the location set to your city, three times each.

Then we count the answers that name you. The result reads “named in x of y answers”, with the date and the city. On the same day we check Google for each question's short search. A person also asks the questions in the ChatGPT app and records what the app said.

For our Toronto client on 2026-09-25, the result read: The Kay Media named first in 15 of 15 answers, her site cited in all 15. We asked five questions her clients ask, each three times in a new temporary chat in the ChatGPT app, with web search on. The case

1. The questions

   Each line of work gets questions a buyer would ask: a problem in their own words, the best firms for a size or situation, or a comparison or a cost. None names a firm. Each carries the short search the same buyer would type into Google, three to six words. For operations consulting in Toronto, one question reads: “Which Toronto strategy consultants take on companies with fewer than 200 employees, and what does a strategic plan usually cost?” Its Google search is “strategic planning consultant toronto”.

2. The assistant

   We send each question to OpenAI's API, the Responses API, with its web search tool on. The model's name is printed on every capture; in September 2026 it was gpt-5.5. We use the official API and never scrape the ChatGPT app. In our September captures, every answer searched the web before replying and linked the pages it used.

3. The city

   OpenAI's search tool takes an approximate user location. We set it to the buyer's city, for example Toronto, Ontario, Canada, which tells the search where the buyer is. Each capture names that setting: “location set to Toronto”, never “from Toronto”, because the request itself does not leave from there.

4. Three runs, each firm counted once per answer

   Answers change between runs. In our September captures for Toronto, 53% of the firms named for a question appeared in only one of its three answers. So we ask every question three times and count a firm once in each answer that names it. SparkToro, an audience-research company, had 600 volunteers run 12 prompts through ChatGPT, Claude and Google's AI 2,961 times, and the same list came back less than once in 100 runs (published 2026-01-27). One answer tells you little. Three runs is our floor.

5. Who counts as named

   A second, smaller model (gpt-5.4-mini in September 2026) reads each answer and lists the firms and people it puts forward. A name registry, keyed by each firm's website, keeps one spelling per firm across every capture, so a practice the answers name on its own, such as “Deloitte Private”, counts as its firm. A person reviews the names and counts before an edition is published.

6. The sources

   Every page an answer cites is stored with the answer. We count how many answers cited each site, and whether your own site was one of them. In Toronto in September, the named firms' own websites were 84% to 93% of the sites cited, line by line. OpenAI's terms ask that citations be clearly visible and clickable, and every source in a capture is a link.

7. Google, the same day

   For each question's short search we record Google's organic results, location set to the same city, through Serper, a search API. The capture shows your best position in the top 20; page one means the top 10. Ads, the map pack and AI Overviews are outside that data, so the capture marks them “not measured”.

8. The hand check

   The API and the app can differ. Before the first capture for a line of work goes out, a person asks its questions in the ChatGPT app, in a browser with its location set to the city, and notes whether the app named each firm and how many of the API's firms it named. The capture prints both: “On [date] the ChatGPT app named [your firm] in [k] of 3 answers, and [n] of the [m] firms the API named.”

9. Failed runs and fresh dates

   A run that fails is logged and left out, so y counts only the answers that came back. A line with too few answers stays out of the public index until it is run again. Every capture carries its date, and a line older than 14 days is run again the day you ask for your capture.

ChatGPT, through OpenAI's API, and Google's organic results.

The capture leaves out:

• Claude and Perplexity, until we have measured them the same way;
• Gemini, whose API terms forbid anyone to “cache, frame, syndicate, resell, analyze, train on, or otherwise learn from Grounded Results”, which a capture has to do;
• Google's ads, map pack and AI Overviews, marked “not measured”;
• a buyer's own ChatGPT account, with the memory and history the API doesn't have;
• what happens after the answer: a capture shows who was named, not who got the call.

Why “named in x of y”, and no single number.

People ask what a good AI visibility score is. We don't publish one. A score folds the questions, the city, the date and the number of runs into one figure, and those are what you need to check it. “Named in 3 of the 15 ChatGPT answers on operations and management consulting, location set to Toronto, 2026-09-24” can be checked by anyone. A single number can't.

Search Console now counts Google's AI answers too.

Search Console has a Generative AI performance report. It counts impressions in AI Overviews and AI Mode, and has been open to all sites since 2026-08-31. For clients we read it beside the capture: one is Google's count of its own AI features, the other is ours of ChatGPT.

Your capture is private. We send it to you and to nobody else.

Who AI names, our public index, publishes each month who ChatGPT named in each line and city, from 27 October 2026: the firms and experts named, in alphabetical order, with their counts and the pages cited. It never shows a firm's absence, and never what an assistant said about a firm. Nobody pays to appear. Anyone listed can ask to be removed, and is, within 48 hours.

We measure ourselves the same way, every month.

On 2026-09-25 we asked ChatGPT for the best SEO agency in Morocco, three times in English and three in French. It recommended Maroc SEO, the agency whose founder and team run grizaille, in 4 of 6 answers and named it first in 3.

From October, grizaille's own name goes through this method each month: the questions a buyer would ask about studios like ours, three runs each, location set to Toronto. The count goes on Results.

Every capture ends with this line, filled in:

Method: OpenAI API, [model], web search on, approximate location [city], [province], [country]. Three questions a buyer in this line would ask (a problem, the best firms for a size or situation, a comparison or cost; none names a firm), each asked 3 times on [date]; a firm counts once per answer. Google: organic results for the short search shown under each question, location set to [city], via Serper, the same day. Not measured: Claude, Perplexity, Gemini (its API terms do not allow stored answers), and Google's ads, map pack and AI Overviews. Firms extracted by [model].