How we measure AI visibility, following Rand Fishkin
SparkToro had the same questions put thousands of times to ChatGPT, Claude and Google AI. Less than 1 in 100 times did the same list of brands come back, and less than 1 in 1,000 times in the same order (SparkToro, January 2026).
Rand Fishkin concluded that a position in an AI answer says nothing, but that the percentage of times a name appears, over enough questions and repetitions, is a reasonable measure. Our measurement is built on that.
Please, dear god, ask ten times before you trust the list.
Six rules for our measurement
- No position weighting. Mentioned or not mentioned is what counts. Here we deliberately depart from benchmarks that do weight by place in the answer.
- Many repetitions. Proposal per edition: 40 questions, each asked 60 times, in four AI tools. That makes 9,600 answers. One answer names many businesses, so that set counts for all participants together.
- Real questions included. People phrase the same question very differently (SparkToro, February 2026). We have some of the questions written by business owners.
- Per AI tool separately. Only then do we average. The ChatGPT app and ChatGPT via the API do not always name the same businesses (Gold Lemon, September 2026).
- Show the size of the choice set. For each question, the number of different names. 30% visibility among eighty names is stronger than 30% among eight.
- Show the calculations. Questions, repetitions, raw data and scoring rules are public. Names that AI mentions but that no longer exist (or never did) are filtered out.
The formula
Visibility V of specialist b on AI tool e = number of answers in which b is mentioned, divided by (number of questions × number of repetitions). Total visibility is the average across the AI tools.
For each AI tool we calculate a 95% Wilson interval. The points for pillar 5 are 8 × min(1; V / P90), where P90 is the 90th percentile of all participants. That way a single outlier does not drag the rest down.
What visibility does and does not prove
The visibility of a specialist themselves says little about what they do for clients. In the Gold Lemon benchmark of October 2026, the most frequently mentioned agency had the lowest technical score in the top 20 (Gold Lemon). That is why this pillar counts for only 10 points.
For client case studies we look beyond the mention: more branded searches and direct traffic, as Fishkin recommends. Read how branded searches reveal AI influence or measure it yourself with the free AI visibility tally sheet.
Frequently asked questions
Why do you not measure a position?
Because the order in AI answers is random. In the SparkToro research, the same list in the same order came back less than 1 in 1,000 times.
What do you measure instead?
The percentage of answers in which a name appears, per AI tool, over many repetitions of the same questions. See also: GEO agency: why most of them cannot really help you.
How often do you ask each question?
The proposal is 60 times per question per AI tool, in line with the 60 to 100 repetitions that Rand Fishkin mentions. See also: Best GEO agency: how to recognise a fair ranking.
Which AI tools do you measure?
The proposal is four: ChatGPT, Gemini, Claude and Google AI. We report each tool separately and only average them afterwards. See also: AI visibility tool: do you really need one?.
Where do the questions come from?
Partly from the editorial team, partly written by business owners themselves. Real people phrase the same question very differently, and that belongs in the measurement.
Is this not very expensive?
Less than it seems. One answer names many businesses at once, so the same set of answers counts for all participants together.
What is a choice set?
The number of different names an AI tool mentions for a question. Being visible among eighty names is harder than among eight, so we show that size alongside.
What is a Wilson interval?
A margin of error for a percentage. It shows between which values the true percentage probably lies.
Can I measure this myself?
Yes, on a small scale with our tally sheet. Ask the same question at least ten times per AI tool and count how often you are mentioned.
Do you publish the raw data?
Yes. Questions, numbers of repetitions and outcomes are published per edition as open data.