How We Measure AI Visibility
Famewise measures AI brand visibility by asking 3 AI models — OpenAI GPT, Anthropic Claude, Google Gemini — the same brand-free consumer questions across 40 food and beverage categories, each question with web search on and off, for about 3,690 answers per weekly run. Visibility is the share of a category's answers naming a brand at least once; share of voice is its slice of all brand mentions in those answers.
| Categories tracked | 40 |
|---|---|
| Brands tracked | 796 |
| Questions per category | 15 |
| AI models | OpenAI GPT, Anthropic Claude, Google Gemini |
| Search modes | with and without web search |
| Answers per run | ~3,690 |
| Refresh frequency | weekly |
1. What we ask
For each of the 40 food and beverage categories we track, Famewise generates 15 synthetic consumer questions — the "best", "top" and "which" questions where an assistant actually names brands. Question generation runs on a high-reasoning model with web search enabled, and is given the category's brand roster with a single instruction: never name any of them. A post-filter re-checks every generated question against that roster, so a question that leaks a brand name never reaches the models.
Questions are shared across every brand in a category. That is deliberate: one question set per category means every brand is scored against the same prompts, so a leaderboard compares like with like.
A small number of categories are sampled at double depth — fruit snacks (30 questions) — where we publish and get cited most and want the larger denominator.
2. Who we ask
Every question goes to 3 current production models — OpenAI GPT, Anthropic Claude, Google Gemini — and each one is asked twice: once with the model's web-search tool enabled, once without. The two modes answer different questions. Search off shows what the model believes from training; search on shows what live retrieval surfaces. The gap between them is the web-search lift.
Every stored answer records the exact model version that produced it, so a shift in the numbers can always be traced to a model change rather than guessed at.
3. How a mention is counted
A separate extraction model reads each answer and returns every brand it names, the 1-based position of each brand's first mention, and the sentiment toward it. Brand names are then canonicalised in code — case, punctuation, ampersands, possessives and a reviewed alias map all fold to one key — so "Ben and Jerrys" and "Ben & Jerry's" are one brand, and a brand cannot inflate its score by being spelled several ways.
A brand counts once per answer, however many times that answer repeats it. Mentions are not weighted by how enthusiastic the answer is; sentiment is reported as its own separate metric.
4. The formulas
| Visibility | answers naming the brand at least once ÷ all answers in scope |
|---|---|
| Share of voice (share of mentions) | the brand's mentions ÷ all brand mentions in the same answers |
| Average position | mean 1-based rank of the brand's first mention, across the answers that name it (lower is better) |
| Sentiment | the majority label across the brand's mentions, with a mean score from −1 to +1 |
| Web-search lift | visibility with search on − visibility with search off |
The two percentage metrics have different denominators, which is the single most common misreading of an AI-visibility board. Visibility divides by answers; share of voice divides by mentions. Since a typical answer names several brands, a brand can appear in most of a category's answers and still hold a small slice of its mentions.
5. Scope and denominators
Each metric is computed inside the scope of the page showing it. A category leaderboard divides by that category's answers only; an audience leaderboard pools the audience-framed questions across every category; a brand report card counts every answer in the run that names the brand. Every page prints the denominator it used, so no figure has to be taken on trust.
6. Refresh frequency
The full matrix re-runs on a schedule — currently weekly — and each run is stored as an immutable point-in-time snapshot rather than overwriting the last one. Public pages always render the latest completed run and carry an Updated date showing when that run was measured. Because snapshots are kept, visibility is comparable run over run and the dashboard can show movement rather than only a current state.
7. Publishing thresholds
A page is published only where the data supports it. A category leaderboard needs at least 30 answers and 5 brands named; an audience leaderboard at least 24 answers; a head-to-head comparison at least 4 mentions for each brand. Below those floors there is no page at all rather than a thin one. When a page that previously qualified stops clearing its threshold, it keeps serving its last qualifying snapshot with a visible notice saying which run that was — we would rather show honestly-dated older data than present it as current.
8. What this does not measure
Famewise measures what AI assistants say, not what shoppers do. High visibility means an assistant names a brand when asked; it does not measure traffic, purchase intent or sales. Answers vary between runs, so any single run is a sample. And extraction is itself model-driven: it can occasionally miss an unusual brand name or merge two similar ones, which is why every page reports the answer count behind its numbers.
Frequently asked questions
How does Famewise measure AI brand visibility?
Famewise generates brand-free consumer questions for each of 40 food and beverage categories, asks every question to 3 current AI models — OpenAI GPT, Anthropic Claude and Google Gemini — twice, once with web search and once without, then runs an extraction model over every answer to record each brand named, the position of its first mention and the sentiment toward it. Visibility is the share of a category's answers that name a brand at least once.
What is the sample size?
About 3,690 AI answers per run: 40 categories x 15 questions x 3 models x 2 search modes, plus extra depth on the categories we sample twice as heavily. Each category page states its own denominator, and no page is published below a minimum answer and brand count.
How is AI share of voice calculated?
Share of voice is a brand's mentions divided by all brand mentions in the same set of answers. Every brand named in an answer counts once for that answer. Because AI answers usually name several brands, the shares across a category add up to well under 100% per answer — which is why share of voice is always smaller than visibility for the same brand.
How is that different from visibility?
They divide by different things. Visibility counts answers: the share of a category's answers that name the brand at least once. Share of voice counts mentions: the brand's slice of every brand mention in those answers. A brand named in most answers, but always alongside six rivals, has high visibility and modest share of voice.
How often is the data updated?
Every scheduled run — currently weekly — re-asks the full question set across every model and search mode and stores a fresh point-in-time snapshot. Public pages show the latest completed run and carry the date it was measured.
Are the questions written to favour any brand?
No. Questions are generated per category with an explicit instruction never to name any tracked brand, and a post-filter drops any question that slips one through. The models are asked what a shopper would ask, so they are never led toward a brand.
Which brands appear — only the ones you track?
No. The extraction step records every brand an answer names, so leaderboards include organically surfaced competitors, not just a curated roster. That is why a category can show thousands of distinct brands named.
What are the limitations of this method?
Three worth stating. AI answers vary run to run, so a single run is a sample, not a verdict. Extraction is done by a model and can occasionally miss or merge a brand name, which is why each page reports an answer count you can weigh the numbers against. And visibility measures what assistants say, not what shoppers then buy — it is a presence metric, not a sales metric.
See it applied
Is your brand visible in AI answers? Check it free — then track your share of voice run over run.
Check your brand free Sign up