AI visibility is how often, and how prominently, an entity is mentioned unaided in AI model answers. 'Unaided' is the important word: the question never names the entity. If you ask a model 'who are the best footballers of all time?' and it answers with a list, every name on that list just demonstrated AI visibility — and every plausible name that didn't appear demonstrated its absence.
This definition is deliberately narrow. It says nothing about whether the mentions are accurate, flattering, or deserved. It is an occurrence measurement, like counting citations — not a judgment of merit. The narrowness is what makes it measurable.
Why AI visibility suddenly matters
A growing share of questions that used to go to search engines now go to AI assistants, and an assistant's answer usually contains a handful of names rather than ten blue links. For a public figure, a brand, or an organization, being inside that handful is the new equivalent of ranking on the first page.
The practices growing around this shift go by several names — generative engine optimization (GEO), answer engine optimization (AEO), or simply AI search optimization — but they all start from the same question: when the model answers unaided, do we appear?
You cannot manage what you have not measured, and most claims about 'what AI says' are anecdotes: one person, one prompt, one day, one model. Model outputs vary between runs, between model versions, and between phrasings of the same question, so a single screenshot is closer to a coin flip than a finding.
What a real measurement requires
To turn the anecdote into a number you can compare over time, a measurement needs the same properties as any repeatable experiment:
- A fixed, versioned prompt set. The questions must be declared before the test and never name the target. Fameproof uses a 24-prompt pack spanning recognition, domain prominence, cultural association, and bounded recommendation lists.
- A declared model configuration. Which API models, what temperature, how many repetitions — all recorded, because a score only describes that configuration on that date.
- Deterministic extraction. Whether a response 'mentions' someone is decided by explicit name-and-alias matching rules, not by asking another model for its opinion.
- Honest failure handling. Failed API samples are excluded and reported, and a result with insufficient coverage is marked insufficient instead of being padded.
What AI visibility is not
A high score does not mean someone is more famous, more important, or better liked — it means they surfaced more often and earlier inside a defined test. It also does not describe consumer chat applications, which layer retrieval, personalization, and product logic on top of the underlying models.
And it is not static: model updates and retraining can move visibility without the subject doing anything at all, which is why every serious measurement carries a date.