Method
How AI visibility is measured.
Answers from AI engines change from one run to the next. This is how Monitor turns many answers into a share you can trust, and what it will not claim.
01 · Questions
Questions in your buyers' words.
A question set starts from what you sell, where and to whom. Suggestions are written the way a buyer asks an assistant ("Which accountants in Bristol work with small charities?"), and you edit, add or remove them before anything runs.
Your name is never in the question. Asking "Is Smith & Co a good accountant?" will almost always name Smith & Co, so it tells you nothing about whether a buyer who does not know you yet would hear of you.
Monitor tracks 15 questions and up to 3 competitors. The same questions are asked every week, so one week can be compared with the last.
02 · Runs
Every question, every engine, more than once.
Each question goes to ChatGPT, Perplexity and Gemini with web search turned on, 3 times a week on Monitor. Asking once is not enough: the same question can name different businesses an hour apart.
That gives 45 answers per engine each week (15 questions × 3 runs). Every answer is kept, with its sources, so you can read exactly what the engine said.
03 · Mentions and citations
Named, cited, or both.
A mention is your business named in the answer text. It is matched against the names you confirm: your trading name, common short forms and your domain.
A citation is a link to your website among the sources the engine shows. An answer can cite your page and still recommend someone else, so mentions and citations are counted separately.
Competitors are counted the same way, from the names you give, so the comparison is like for like.
04 · Shares and ranges
Every share comes with its range.
A share is the number of answers that name you, out of the answers asked. On its own it overstates how much is known, so it is always shown with a range: the values the true share could reasonably take, given how many answers there were. The range is a 95% Wilson interval, which stays honest with small samples and with shares near zero.
In plain words, a week reads like this:
Perplexity: named in 7 of 45 answers; likely 8–29%.
The best single guess is 16%, but anything from 8% to 29% fits what was seen. More answers make the range narrower; that is what higher plans buy.
05 · Change
A change is claimed only when the ranges part.
If last week's range and this week's range overlap, the difference could be chance, and the report says "no clear change". Only when they no longer overlap is it reported as a rise or a fall.
Last week named in 7 of 45 answers; likely 8–29%. This week named in 16 of 45 answers; likely 23–50%.
Last week named in 7 of 45 answers; likely 8–29%. This week named in 25 of 45 answers; likely 41–69%.
06 · Limits
What cannot be known.
- How many people ask a question. The engines do not publish it, so there is no search volume for AI answers.
- What one particular person sees. Answers depend on whether they are signed in, what they asked before and where they are.
- Why an engine chose a business. The sources it shows are a clue, not an explanation.
- Anything about engines that are not sampled, including Google AI Overviews and Claude.
07 · Wording
Why a report never says "ChatGPT doesn't recommend you".
A sample can show that you were named. It cannot prove that you never are. Even when none of this week's 45 answers name you, the range still runs up to 8%. And a buyer who asks differently, or is signed in, may hear your name.
So a report says what was seen ("not named in any of 45 answers this week") and never turns it into a verdict about the engine.
08 · Readiness
Readiness is not visibility.
Readiness is whether AI tools can read your site: crawler rules, an llms.txt file, structured data, page speed. Visibility is whether the engines name you. A site can be perfectly readable and never named, or named often from what others write about it.
The two are reported separately, and neither is reduced to a third-party score to chase. A readiness fix is worth making when it removes a real barrier, not to move a number.
FAQ
Common questions.
Why does my share move when I have changed nothing?
Because the engines do not give the same answer twice. With a few dozen answers a week, a share can move by ten points by chance alone. That is why each share comes with a range, and why a change is only reported when the ranges no longer overlap.
Is this what my customers see?
It is a fair sample, not a copy. The engines are asked directly, with web search on. A customer who is signed in, has chatted before or is somewhere else may get a different answer.
Why are Google AI Overviews not included?
They are not sampled. Once Google reports AI Overview appearances through Search Console, that will be a better source than sampling Google's results pages.
Can I trust a share from a single week?
Read the range, not the single figure. A share is the best estimate from that week's answers, and the range shows how far the true figure could reasonably sit from it.