Measure your AI visibility without fooling yourself.
A visibility number is the share of answers that name you, measured on a sample. This chapter covers which questions to ask, how many times to ask them, what to record, and how to read the result honestly.
Chapter 2 of 6 of the AI visibility guide
Which buyer questions should you track?
Track the questions a buyer asks before they know your name: “what is the best X for Y”, “X alternatives”, “X vs Y”, “which X has a free trial”. Those are where shortlists are made.
- Start small. Five or six topics with up to three questions each is enough to see the pattern. More questions are worth adding once you act on the first set.
- Keep brand questions separate. A question that names you will mention you whether or not the engine knows who you are, so it says little about visibility. Use it to check accuracy instead.
- Write them the way buyers do, in full sentences, with the constraints they care about: price, team size, integrations.
Mixing in questions that name a brand moves the number a lot. In our own benchmark of 220 answers, the category leader was named in 59% of them overall and in 40% of the answers to the 16 questions that named no brand at all (the benchmark).
To build the list, see how to find the buyer questions AI recommends you for; to word it, how to test the ways buyers phrase a question; and to read the results, prompt coverage.
Does the wording of a question change the answer?
More than asking it again does. Across 13,994 answers for 96 brands, a brand named on one phrasing of a buying decision was named on a sibling phrasing the same day 55% of the time. The same wording asked again later kept the brand 78% of the time (the phrasing study).
The public Tesla example shows the spread (scans of Oct 4, 8 and 15, 2026). Of five phrasings of one grid storage decision, two named Tesla in 12 of 12 answers and two named it in none. Those two asked about cost per MWh and buying criteria, and the engines answered with prices and checklists, naming none of the suppliers the other phrasings named. Report each phrasing separately before you average them; a topic score that blends 12 of 12 with 0 of 12 describes neither.
The 95% interval for one check that named you: consistent with an engine that names you one time in five.
The arithmetic, Sep 20, 2026Daily changes in a single answer that reversed the next day, across 40 buyer questions on four engines.
Oct 5 to 13, 2026Times two companies with the same 44% mention rate were named first in the same 160 answers.
Oct 4 to 9, 2026Why is one check not enough?
Ask the same engine the same question twice and you can get two different answers. One answer is a sample, not a fact, and a handful of samples can mislead. A single answer that names you has a 95% interval of 20.7% to 100%; a single miss, 0% to 79.3% (why one answer says so little).
The honest way to report a rate from a small sample is with an interval. In the Tesla example, ChatGPT named Tesla in 16 of 25 answers, which is 64%, but the 95% Wilson interval runs from about 45% to 80%. Perplexity's 10 of 25 runs from about 23% to 59%. The two ranges overlap, so even that gap is not yet proof that one engine favors Tesla. More answers narrow the range: 60 of 200 runs from about 24.1% to 36.7%. A change only means something once it is bigger than the noise; chapter 6 covers how to test that.
The full method is on the methodology page, and the terms are in the glossary.
What should you record besides whether you were named?
- Where you were named. Answers to buying questions usually name several companies; in our sample the median named five. Two companies each named in 70 of the same 160 answers were named first 38 and 15 times (position, not just presence).
- Whether the engine searched. An answer that never looked anything up cannot reflect a change to any page (chapter 1).
- Which pages it cited, and which competitors it named instead. That is the raw material for chapter 3.
- Each engine on its own. A blended score hides the engine you need to work on. The IAB's August 2026 measurement framework lists position beside mention rate as a core brand metric and treats fewer than 50 queries as exploratory.
The AI share of voice and AI visibility score pages show how these roll up into one number, and what each one hides.
How often should you measure?
Single answers churn every day; a fixed set of questions moves far less. Over 40 buyer questions asked daily on four engines, about one name in five that was in an answer was gone the next day, and 47% of those changes reversed a day later. Yet a single day's rate landed within a median 3.6 points of that week's average (the cadence study).
- Reporting a trend on a large, fixed set of questions: weekly is usually enough.
- Tracking 20 questions or fewer: daily runs averaged over the week give a much narrower interval.
- Testing a change you just shipped: daily shows which day the answers moved and whether the move held.
More on the trade-off in daily AI visibility monitoring.
Keep reading
Every chapter
Related reading
- IAB: Measuring Visibility in the AI Era (August 2026) www.iab.com/wp-content/uploads/2026/07/IAB_Measuring_Visibility_in_the_AI_Era_August_2026-1.pdf
- Brown, Cai and DasGupta: Interval Estimation for a Binomial Proportion, Statistical Science 16(2), 2001 projecteuclid.org/journals/statistical-science/volume-16/issue-2/Interval-Estimation-for-a-Binomial-Proportion/10.1214/ss/1009213286.full
- Wilson score interval en.wikipedia.org/wiki/Binomial_proportion_confidence_interval#Wilson_score_interval
Proofsource figures link to the post or study that measured them, with its dates and method. Tell us if anything here is out of date.
Find where AI leaves you out. And fix it.
See which buyer questions name your brand, who gets named instead, and what to fix. 200 answers. 25 Prompts/Question. Top AI engines. 2 days. Free. No card.
Google AI OverviewsGoogle AI Mode
