How we measure AI visibility
Read the full technical whitepaper (the mathematics) →
Most tools hand you a single number and ask you to trust it. We treat AI visibility as a measurement problem, and we show our work. Here is how the numbers in your report are produced, and where our edge stays proprietary.
This page documents the Recommendation visibility measurement. The Source authority measurement (whether AI cites your site as a source) runs its own protocol: 24 questions asked 4 times each on 3 assistants across 4 to 6 subjects, versioned as Measurement protocol · Source authority v1 and documented inside its own report. See that sample →
AI assistants answer your customers' "who is best for this?" questions with a short list of recommended businesses. Measuring whether you are on that list sounds simple, but AI answers are noisy, they vary by engine, and they change over time. A number that ignores all of that is a guess with a decimal point. These are the principles that make our numbers trustworthy instead.
We ask the questions your customers actually ask
We generate the real prompts a buyer would type to find a business like yours, not one flattering query we picked. If the questions are wrong, every number after them is wrong, so this is where we start.
In August 2026 we measured this directly: three Toronto buying situations, each phrased two ways, six runs per phrasing, one AI model, 36 answers in total. Rewording the question changed which companies were named, and at the extreme it reversed the result: one firm went from named in 6 of 6 answers to 0 of 6, and another from 0 of 6 to 6 of 6. That is one small test on one model in one market, so we treat it as a demonstrated risk, not a general law. It is why your questions are locked once measurement begins.
Live in every reportWe measure many times, and show the uncertainty
Ask an AI the same question twice and you can get two different answers. We ask every question 8 times per engine and report your mention rate with a 95% confidence interval, so you get an honest margin of error instead of a falsely precise score.
Live in every reportWe prove a gap is real before we report it
If a competitor looks like it is beating you, we run a statistical significance test, corrected for the many comparisons in a report, before we say so. We do not call a coin biased after a single flip, and we do not report a lead that is inside the margin of error.
Live in every reportWe check every major AI, and how much they agree
ChatGPT, Claude, Gemini, Perplexity, and Google AI Overviews do not always agree. We measure across engines and report their level of agreement, so you can see where you are consistently recommended and where one engine is carrying, or missing, you.
Live in every reportEvery number links back to the evidence
This is the one that matters most. Nothing in your report is a black-box score. Every finding traces back to the actual AI answer that produced it, so you, or a skeptical colleague, can read the exact response and verify it yourself.
Live in every reportWhat we refuse to do
- Hand you a vanity score with no way to check it.
- Report a competitor gap that is really just sampling noise.
- Judge your visibility from a single AI answer.
- Recommend keyword-stuffing tricks that AI engines ignore or penalize.
Where our edge stays proprietary
A raw mention count can mislead: an easy question flatters you, a competitive one punishes you, and a model update can move everyone at once. Our research is working to extend the measurement further, toward a rating that is comparable across businesses and stable through model changes rather than fooled by them. That model is a research prototype today; it is not part of the reports you buy.
We are transparent about our principles and every finding is verifiable, but the specific models behind this work are our own. It is the part we keep to ourselves.
In active developmentSee it applied to your business
Run a free preview and see the evidence behind every number.
Check my AI visibility →