Share of Model
How often AI assistants name your brand, as a share of the whole field, measured repeatedly, reported with a confidence interval, and never dressed up as more certain than it is.
A share, not a rate
Share of Model is your brand's weighted presence divided by the total weighted presence of every brand in the field, across the same questions, the same engines and the same cadence.
That denominator matters. "We were mentioned 30% of the time" is close to meaningless without knowing who else was in the answer and how often. A field share tells you whether you are the default recommendation, one of several, or absent.
Product screenshotWhat goes into the score
Presence is not a yes/no. For each brand, in each answer, we score five things, and the weights sum to 100.
| Signal | Weight | What it captures |
|---|---|---|
| Mention | 30% | Was the brand named at all |
| Recommendation | 30% | Was it actually recommended, or just listed |
| Position | 20% | Where in the answer, first is not the same as fifth |
| Sentiment | 15% | How it was characterised |
| Entity accuracy | 5% | Did the model describe you correctly |
The prompt portfolio is the instrument
You define a set of real buying questions, the things your prospects actually ask. That set is then frozen and versioned. A measurement instrument that changes underneath you produces a trend line that means nothing, so a running cycle cannot have its own questions edited mid-flight.
Each question is asked repeatedly, not once. AI answers vary run to run; a single sample is an anecdote. The full sampling standard is on our measurement methodology page.
Honest statistics, by construction
Product screenshotEvery figure has an interval
We report the range, not just the point. A 12% share from twenty samples and a 12% share from two thousand are different claims, and the interval is what says so.
The significance gate
If the interval on a change spans zero, the interface says "no significant change" and draws no arrow. We would rather show you nothing than a movement that is sampling noise.
Thin cells say so
Cut the data by engine and by question type and the sample per cell shrinks fast. Below the minimum needed to carry an interval, a cell reports "not enough data" rather than a number.
Failed runs are excluded, not counted as zero
If a measurement fails to complete, it leaves the denominator entirely. Counting a failure as an absence is a systematic downward bias, and it is a common one.
Raw answers are kept
We store every AI response verbatim, not just the numbers we extracted from it. When our extraction improves, we can re-analyse your entire history against the better method instead of starting the trend line over.
Every stored score carries the version of the rubric that produced it, so figures from different methods are never silently compared.
Coverage by plan
| Plan | Prompts | Cadence | Engines |
|---|---|---|---|
| Diagnostic | 25 | Monthly | ChatGPT, Google AI Overviews |
| Visibility | 100 | Weekly | + Google AI Mode, Copilot |
| Program | 250 | Monthly | + Perplexity, Claude, Grok |
Diagnostic and Program measure monthly on purpose. Ten samples in one cycle carries a real interval; the same ten spread across four weekly cycles gives you four under-powered ones. Full pricing →
The method in depth: Share of Model, measured properly and the versioned measurement methodology.
Measure your share of the answer
Start with a free audit, then add measurement on a paid plan.