Why this cannot be looked up
Search measurement works because the engines expose enough information to build tooling on: a position exists, it can be checked, and the click it produced shows up in your analytics. Neither condition holds for AI answers. No platform publishes how often it names your business, and a mention that resolves somebody’s question without a visit leaves almost no trace in your reporting.
Worse, the thing being measured moves. Generative systems are non-deterministic — the same prompt can produce different wording and sometimes a different set of businesses — and answers also vary by platform, model version, location, personalisation and phrasing. Any method that ignores that variation produces numbers that feel precise and mean nothing.
So the honest approach borrows from survey research rather than rank tracking: sample a defined population of questions, repeatedly, and report a distribution with a stated confidence.
Step one: build a prompt set that represents real demand
Your prompt set is the instrument, and everything downstream depends on it. Write thirty to sixty questions a genuine prospect might ask, in their words, covering the range of intent rather than only the flattering end.
- Category questions with no brand named: "who does commercial electrical work in this city?"
- Problem-first questions, describing a situation rather than a service.
- Comparison and shortlist questions: "what are the three best options for a small business?"
- Qualification questions: "what should I check before hiring someone for this?"
- Brand questions: "what do you know about this business?" — the accuracy test, not the visibility test.
- Awkward questions: cheapest, fastest, most specialised, and anything a sceptical buyer would ask.
Step two: sample properly
Ask each question more than once, on more than one platform, and ideally from more than one location if your market is geographic. A single run per question tells you what happened once; three or five runs begin to show a rate. Keep the wording identical between runs, because rephrasing changes the question being measured.
Two practical cautions. Use clean sessions rather than a logged-in account that has discussed your business before, or you will measure the platform’s memory of you rather than its knowledge of you. And record the date, platform and model version with every observation, because an unexplained shift next quarter is usually a model update, not your marketing.
Decide the sampling plan before you start, and write it down. Choosing how many runs to do after seeing the first few results is how a measurement exercise quietly turns into a search for reassuring answers.
Step three: record six things, not one
A single "visibility" figure hides the distinctions that tell you what to fix. For each answer, capture the following.
- Presence: were you named at all?
- Prominence: named first, in the middle, or as an afterthought?
- Nature of the mention: actively recommended, neutrally listed, or mentioned with a caveat?
- Accuracy: is every factual claim about you correct and complete? Log each error separately.
- Citations: which sources did the answer attribute, and was your site among them?
- Competitive set: which other businesses were named, and how consistently do they appear across your prompt set?
Step four: turn observations into something you can track
From those records a few honest metrics emerge. An inclusion rate: the share of sampled answers in which you were named, stated with the number of samples it came from. A recommendation rate: of the answers that named you, how many actively put you forward. An accuracy rate: the share of answers containing no factual error. A share-of-answers figure against the competitors your own sampling surfaced.
Report them with their limits attached. As an illustration of the difference in phrasing: "named in 11 of 40 sampled answers across two platforms in August" is defensible, because the reader can see what produced it. A bare "22% AI visibility" is not, unless the prompt set, platforms, dates and sample size are shown alongside it. This is why AI Says reports a data confidence level beside every score and distinguishes observed results from estimates derived from website signals.
Proxy signals when direct observation is thin
Direct sampling is expensive, and for some niches the question volume is simply low. Measurable signals on your own side can fill part of the gap: whether your pages are answer-shaped, whether structured data is present and accurate, whether AI crawlers are permitted and are actually requesting pages in your server logs, and whether your core facts are consistent across third-party listings.
These are readiness indicators, not outcomes. They tell you whether you have given the systems something usable; they do not tell you whether you were named. Keep the two clearly separated in any report, or you will end up celebrating a tidy website while remaining invisible.
Mistakes that invalidate the measurement
The commonest is measuring only brand-name prompts. Asking an assistant about your business by name almost always produces an answer, which feels reassuring and tells you nothing about whether you are found by somebody who does not know you exist. That prompt type measures accuracy, not visibility.
The others are familiar from any research discipline: too few samples, changing the prompt set between periods so the comparison breaks, running everything on one platform and generalising, reporting the best run rather than the distribution, and over-reacting to a single absence. Fix the method before you read anything into the trend.
If the underlying mechanism is still unclear — why the same question yields different shortlists — read how ChatGPT chooses which businesses to recommend. For what to do once you have a baseline, see how to improve your AI visibility.
Last reviewed 19 August 2026. AI Says is not affiliated with or endorsed by OpenAI, Google, Anthropic, Perplexity or other AI and search providers; platform names are used descriptively.