Measuring a brand’s visibility in AI models is different from measuring a search position: answers vary, sources change and the same request can give different results. Even so, it is possible to track mentions, citations and sentiment with a simple, repeatable method that is honest about its limits.
The essentials in a definition
Brand visibility in AI models is the frequency and quality with which a brand is mentioned, cited and described in generative assistants’ answers. It is measured through repeated samples of real questions, based on helpful content, guidance on generative AI and crawler access to content.
What to measure
Three indicators are enough to start: whether the brand is mentioned, whether its site is cited as a source and how it is described (correctly or not, in a positive, neutral or negative tone). Add comparison with two or three direct competitors, for context.
Define the set of questions beforehand: those real customers ask before buying, in several wordings and in the languages of your markets.
Repeatable method
Use the same list of questions, at the same time of the month, in more than one assistant, and record the answers (date, assistant, question, mention, cited source, description). Repeat each question a few times, because the answer may vary. Keep the text to compare over time.
Avoid drawing conclusions from a single answer. Look for repeated patterns in reasonable samples.
A monthly AI visibility dashboard
- Define questions and competitors
- Record answers repeatably
- Analyse mentions, sources and description
- Fix content and data at the source
Interpret and act
When the brand does not appear, understand why: lack of clear content, outdated information, inconsistent data or stronger external sources. When it is described incorrectly, find the origin of the error (an old page, a directory, an article) and correct it. Guidance on helpful content and on generative AI remain the basis.
Also check that the relevant crawlers have access, according to what you want to allow; consult the documentation on OpenAI’s crawlers as an example of how these controls are described.
A well-recorded sample is worth more than an isolated answer.
Limits and ethics
Avoid trying to manipulate answers with hidden text or mass questions. Respect services’ terms of use and privacy. Results are samples, not absolute truths, and do not replace measurement of traffic, enquiries and sales.
Review the method every six months, because assistants and sources change fast.
Decision table
| Indicator | How to record | Limit |
|---|---|---|
| Mention | Yes or no, per question and assistant | Varies between repetitions |
| Citation | Brand site among the sources | Not all assistants cite |
| Description | Correct or incorrect, and tone | Subjective assessment |
| Comparison | Competitors on the same questions | Depends on the sample |
Common mistakes to avoid
- Concluding from one answer. Answers vary.
- Changing the questions every month. It prevents comparison.
- Measuring without acting. The value lies in fixing the source of errors.
How to measure the result
Compare, month by month, the percentage of answers with a mention, with a site citation and with a correct description.
Limits and exceptions
Assistants differ from each other and over time, and the same question can give different answers. The sample is small and the interpretation of tone is subjective. These measurements are indicative, not definitive, and should be read alongside traffic, enquiries and sales.
Illustrative example
Illustrative example: an insurance company defines thirty real questions, repeats each three times in two assistants and records mentions, sources and descriptions in a sheet. It notices a product is described with old conditions, because an outdated page is cited. It updates the page and measures again the next month.
Frequently asked questions
Is there an official metric?
There is no single official metric; use your own simple, repeatable method.
How many questions should I use?
A set of twenty to fifty real questions is usually a good starting point.
Can I automate?
You can, respecting services’ terms of use and keeping human review.
How often should I measure?
Monthly or quarterly, with the same method.
What if the description is wrong?
Find the source of the error, correct or update it, and track in the following measurements.
Measure mentions, citations and description with real questions, a repeatable method and samples, and use the results to fix content and data, without trying to manipulate answers.
Sources and recommended reading
- Google Search Central: generative AI content guidance
- Google Search Central: helpful, reliable, people-first content
- OpenAI: crawlers and user agents
Last editorial review: October 2026. Informational content; always confirm current conditions in the official sources before deciding.