What is AI visibility monitoring? Definition, metrics, and how to judge a tool
· by boneLau
Definition. AI visibility monitoring is the practice of repeatedly asking AI assistants the questions your buyers ask, then recording four things about each answer: whether your brand appears, what position it holds, which source the model cited, and whether the description of your brand is accurate. It is measurement, not promotion — the output is a record of what AI said, on a fixed question set, over time.
The category exists because buying behaviour moved. Buyers now ask ChatGPT, Perplexity, Gemini, Claude or Google's AI Overview before they ever open a results page. When those answers name three tools, the fourth one is invisible in a way that does not show up in analytics, rankings, or bounce rate.
Is "LLM visibility" the same thing?
Mostly yes. People search for LLM visibility, AI visibility, AI search visibility and GEO to describe the same practice: finding out what large language models say about a brand. The terms differ by audience more than by method — "LLM visibility" tends to come from SEO teams, "GEO" from content teams, and "AI visibility monitoring" from the measurement side.
One distinction is worth keeping. "Visibility" often means appeared somewhere once, while monitoring means repeated measurement on a fixed question set. A screenshot of ChatGPT naming your brand is visibility; knowing whether it still names you next week is monitoring. Many tools sell the first and call it the second.
We use "AI visibility monitoring" because it names the part that is hard: repeating the same questions, recording the raw answers, and reporting the denominator honestly.
The five metrics that actually matter
1. Mention rate — and the denominator problem
Most tools report one number: the percentage of answers that mention you. That number is meaningless without its denominator. A brand can score 90% on questions that already contain its name ("is BrandScope any good?") and 0% on questions from buyers who have never heard of it ("best AI visibility tool for agencies?"). Blending the two produces a healthy-looking number that hides a zero.
We report them separately: unbranded discovery (the buyer does not know you) and branded recall (the buyer already named you). The first predicts growth; the second measures reputation. On our own brand, the same weekly run produced 17/18 branded and 1/26 unbranded — the blended figure would have looked fine.
2. Citation rate and citation ownership
A mention is not the same as a citation. An answer can name you from memory with no source at all, cite your homepage, or cite a third-party page about you. Each implies a different fix: no source means you are not in the retrieval set; own-domain means you are the only thing feeding the answer; third-party means someone else is doing your explaining.
There is a fourth case that most tools miss: the cited page belongs to a different company with the same name. Counting that as your mention is a false positive. We classify every URL as own / third-party / same-name, and on our own brand we have caught a competitor's pricing page being scored as ours.
3. Position and share of voice
Where you sit inside the answer (first recommendation vs footnote), and which competitors appear in the same answers. Share of voice answers "when AI lists alternatives, what slice is mine" — the closest thing the category has to competitive ranking.
4. Accuracy — does AI describe you correctly?
Being mentioned and being described correctly are different failures. An AI that calls you an "SEO tool" when you are a monitoring platform is worse than silence in some ways: silence is a gap, a wrong category is a wrong first impression that persists.
Accuracy has to be judged per answer, against your own positioning — not inferred from sentiment. Sentiment tells you whether the tone was positive; accuracy tells you whether the facts were right.
5. Uncertainty
Every AI answer is a sample from a non-deterministic process. The same prompt, asked twice, can name different tools. That means every percentage should carry a sample size and an interval — and any week-over-week move inside the noise band should be reported as noise, not as news.
A tool that shows "42%" without saying "3 of 7 answers, plus or minus 26 points" is selling precision it does not have.
How a measurement is actually built
- Fix the question set. Discovery questions that never name the brand, comparison questions, and objection questions. If the set changes every week, week-over-week comparison is meaningless.
- Pick surfaces deliberately. Different assistants retrieve differently: Perplexity returns citation URLs, Gemini and ChatGPT often answer from memory, Google's AI Overview only appears for some queries. Report per surface, then aggregate — never the reverse.
- Sample more than once. One answer per question is an anecdote. Two or three samples per cell gives you a variance estimate.
- Record the raw answer. Not a dashboard verdict — the text. Verdicts can be recomputed; lost context cannot.
- Separate infrastructure failures from genuine silence. A timeout is not "AI did not mention you". Mixing them corrupts the baseline.
Four questions to ask any vendor (including us)
- Do you separate unbranded discovery from branded recall? If they show one mention rate, ask what the denominator is. If the answer is "all questions", the number is inflated by their own brand-name prompts.
- Do you verify the cited URL belongs to me? A citation to a same-name competitor is not your win.
- Do you print sample size and interval? If not, the precise-looking percentage is decoration.
- Do you judge accuracy separately from sentiment? Positive tone about the wrong category is still wrong.
What AI visibility monitoring is not
- Not SEO. SEO measures where your pages rank. This measures what AI says when it does not necessarily cite any page at all.
- Not a content generator. Measurement tells you the gap; it does not write the fix. Any tool claiming both should be judged on which half it does well.
- Not a guarantee. No tool can promise that changing a page will change an AI answer within a week. The honest framing is a measured baseline plus repeated observation.
- Not a single number. A vendor with one headline score is hiding the denominator, the sample, or both.
Where to start if you are new to this
Run the check before buying anything: ask ChatGPT and Perplexity five questions a real buyer would ask, without naming your brand. Count how many answers mention you, and read what they say. That is the whole method, done manually. If the number surprises you, a monitoring tool is worth paying for; if it does not, you just saved a subscription.
Last updated 2026-10-27. Method numbers come from our own weekly runs (run 29708e8e, 2026-10-26). Full method and limits: [how it works](/how-it-works).