How we measure Model Mindshare

Model Mindshare is what the model recalls about you from memory — not what it fetches live. So we measure it closed-book: we switch the web off and ask.

The web, switched off

Live visibility measures what AI says when it searches. Model Mindshare measures what it knows without searching — so we run the probes with retrieval off, and the answer comes from what the model learned, not what it just read. One signal you only get this way: if, with the web off, the model can't place you at all, that blank is itself the measurement — the clearest sign you're not yet in its memory. Closed-book only runs where a model API can genuinely switch retrieval off — the models behind ChatGPT (OpenAI), Claude (Anthropic) and Gemini (Google); search-native surfaces like Perplexity and Google AI Overview are measured live, in the retrieval layer, not here.

Six dimensions, from memory

Live visibility has seven dimensions. Switch the web off and one — the citation footprint — has nothing to read: there are no live sources in memory. So it moves to an input (below), and six remain.

Memory recall: Does the AI name you without looking you up? From its own knowledge, no search — do you come to mind at all.
Top-of-mind primacy: Are you the first name it reaches for? Volunteered early, or only after a nudge.
Recall accuracy: Does it get you right? The facts it recalls — what you do, where, for whom — correct, and attached to your name, not a competitor's or a guess.
Recalled framing: How does it talk about you from memory? Favourably, neutrally, or poorly.
Answer stability: Does it say the same thing every time? Across repeats and rewordings, steady or drifting — reported as a range with a confidence interval.
Knowledge depth: How much does it actually know? Just your name, or the full picture — what you do, where you operate, what sets you apart.

The six Model Mindshare dimensions are reported separately, each with its own confidence range. We do not combine them into a single signal: a calibrated weighting would be required, and we will only publish one derived from real measurement.

Where it learns about you

The citation footprint doesn't vanish — it changes job. With the web off it can't be read from the answer, so we measure it directly as an input: your presence in the sources models train on — Wikidata, Wikipedia, trusted third parties. It predicts the six, so we show it beside them, as the input, not the score.

What we don't claim

This is a behavioural probe, not a readout of the model's weights — we infer memory from how it answers across many phrasings, not from inside the model. Closed-book is something we enforce and disclose, not a perfect lab condition. It's a sampled signal with a range, not a census or a promise. And we don't claim your record is permanent: that would mean measuring across model generations, which we don't run — so we don't say it. We report what we measure, and name what we don't.