How we measure AI visibility

What we measure

For every tracked prompt (“best coffee subscription for offices”), we ask each AI engine the question a real buyer would ask, capture the full answer, and extract three facts per tracked brand: was it mentioned, at what rank against competitors, and with what sentiment. Every stored answer is available verbatim in the product — the score is never the only evidence.

The sampling model — and its limits

A scheduled sweep queries each engine once per prompt. AI assistants are non-deterministic: the same question can get a different answer an hour later. One sweep is therefore a sample, not a census — single-day score moves are directional at best, and we say so in the product. The signal is in the trend: repeated sampling across sweeps, the same way a poll tracker treats individual polls.

Under engine cadence (below), a scheduled sweep queries the engines due that day, and the blended score covers exactly the engines swept — the trend chart shows the engine count per point so a two-engine day is never mistaken for a five-engine day.

APIs vs the apps consumers use

We query provider APIs under our own consumer-style system prompt. That is close to, but not identical to, what a user sees in the ChatGPT or Claude apps — the apps add their own system prompts, memory, and search grounding we can't reproduce. Gemini is queried with search grounding deliberately off, so results reflect the model rather than a transient search index. Google AI Overviews is the exception in the other direction: it is a real capture of the live Google results page, where an absent AI Overview is tracked as a presence rate and never scored as a zero.

Engines & cadence

Cadence is tunable per client. The default mix keeps the two conversation engines fresh daily and samples the rest — measurement frequency is a cost dial, and we'd rather publish that than pretend otherwise. Deployments without engine API keys run deterministic simulations, always labelled “Demo” in the product — demo data never blends into a paying workspace's scores.

The Visibility Score

The 0–100 score is a weighted blend, published here in full:

• Position (50%): rank 1 in an answer is worth 1.0, declining to 0 past rank 5; a prose mention without a rank counts 0.5; not mentioned counts 0.
• Share of voice (30%): the brand's mentions as a share of all tracked-entity mentions in the sweep.
• Sentiment (20%): averaged over the mentions where sentiment is readable; when none are, the term defaults to neutral (0.5) rather than rewarding or punishing silence.

The blended score recomputes the formula over all engines' answers together — position averages engines per prompt, while share of voice and sentiment pool mentions across engines. Weights are published constants in the codebase; if they ever change, this page changes with them — by construction, since it renders the same constants the scorer uses.

Model-version stamping

Every stored answer is stamped with the model version the provider reported — shown on answer receipts, in the prompt explorer, and in CSV exports. When an engine's recorded version changes between sweeps, the trend chart marks it, so a score move can be attributed to their model changed rather than your content changed.

The honest fine print: we record what providers report. Some providers resolve model names to dated snapshots (a change is detectable); others return the requested name verbatim, so a silent same-name update is notdetectable, and we don't claim it is. Google AI Overviews carries our own capture-pipeline version rather than a Google model version, and its markers say so.

How we read what AI says about a brand

Beyond whether a brand is named, we record how it is characterised — fast, expensive, enterprise-ready. Those characteristics come from a fixed, published vocabulary of 46 attributes across 10 facets (v1), listed in full below.

A fixed list rather than free text, for one reason: free text does not add up. Fast, quick and speedybecome three rows that can never be summed, and “you are described as fast in 40% of answers” becomes unstateable. A closed vocabulary is also what makes two brands comparable at all.

The cost, stated rather than buried: a characteristic outside this list is invisible to us.If an engine calls a brand something we have no term for, it is not counted — the number is a floor, not a census. Attribution is also per entity: a characteristic is only recorded when it appears in the answer's description of that brand, not merely somewhere in the same answer. And counts start when the feature shipped; historical answers are not re-analysed, because doing so would mix two vocabularies inside one chart.