We publish how our score works — including where it is weak.

Whether a named business appears in the answer to a specific question, put to a specific AI system, at a specific time — and if it appears, whether it is recommended, merely mentioned, or described inaccurately. We do not combine those into a single weighted score with a hidden formula. Where we show a composite, the components are shown beside it.

The prompt panel we configured for that business, and nothing wider. It is a description of a test panel, not an estimate of what every buyer in a market encounters. We do not present it as general AI visibility, because it is not.

They are not, and we say so. The phrases come from the business's own trade and specialism, written as the questions a buyer would plausibly ask, and they are reviewed by a person. What we do NOT have is observed first-party data on what real users actually type into ChatGPT about that trade — nobody outside the model providers does. Profound comes closest — they license real conversations from consumer panels, which is better raw material than we have. But this question asks how a prompt set was SHOWN to represent the population, and we could not find that published either: the panels are not named, the scaling to population is described only as statistical modelling, and no figure we found carries a margin of error. On this question they have the better data and we have the same missing argument.

Fewer than the raw prompt count, and we do not currently report an effective sample size adjusted for the fact that paraphrases of one intent are correlated. This is a real gap. Until it is closed, treat the prompt count as a count of tests, not of independent evidence.

They are not weighted. Every phrase counts once. That is a crude choice and it has a cost — a rare phrasing counts as much as a commercially central one — but the alternative is a weighting model we could not substantiate, which would move the arbitrariness somewhere less visible.

The phrases are recorded against the site on the first run and LOCKED on every repeat, so the second score is measured against the same questions as the first. A client cannot quietly add favourable phrases or drop hard ones between runs. If the phrase set is deliberately changed, the run starts a new series rather than continuing the old line, and the report says so.

Only precision. Running more variations of an unrepresentative panel produces a more stable estimate of an unrepresentative panel. We do not sell prompt volume as a proxy for rigour.

We do not know. We currently run each phrase once per audit and we have not measured the run-to-run variance that would tell us whether once is enough. This is the single largest hole in our own method and it is first on the list to close.

None is calculated, because question 8 is unanswered — without a measured run-to-run variance there is no honest interval to publish. Until there is, no movement in our score should be read as meaningful on its own.

All of it. There is no statistical model. We would rather show a raw count with its limits stated than a confidence interval that quietly excludes prompt selection, panel composition and model drift — a narrow interval around a biased panel is a more confident-looking wrong answer.

No, and we do not treat them as if they were. Phrases sharing an origin, and every phrase run against the same model on the same day, move together when the model changes. We do not multiply them up into an independent evidence count.

It is not, so we do not run daily. The audit is a point-in-time reading and the retainer reports monthly. A daily line would mostly plot answer variance, and a chart that moves every day is commercially useful and methodologically empty.

It has not been validated against an external benchmark. We have no study comparing our score to referral traffic, conversions or survey recall. We are a young company with no such dataset, and we would rather say that than cite someone else's validation of a different system.

There is no validation, so no. When there is, it will use data we did not generate, select, weight or score — and we will publish the negative result if that is what it is.

The only assumption we make is that a logged-out, one-shot query is a defensible standardised probe — comparable across runs and across businesses, and NOT the same as what a logged-in user with history sees. See question 16.

We do not, and we could not. Every probe is logged-out and single-turn. That means our reading is a controlled measurement, not a simulation of a real session, and personalisation is an excluded source of error rather than a modelled one. This limitation applies to every vendor in this market, including the ones that do not mention it.

Nothing we would rely on, which is why we do not use synthetic personas. Typing 'I am a CFO in London' into a prompt does not reproduce account history, location or memory; it produces a differently-worded prompt. Selling that as a persona segment would be inventing a variable.

The inputs are: the phrase set is locked and recorded, so a client can rerun exactly what we ran, against the same models, outside our platform. What we have not published is a measured test-retest agreement rate — see question 8.

We have not run the same panel through a competitor's platform and compared. It would be a genuinely useful experiment and we would publish it either way.

Per model, side by side. We do not average ChatGPT, Gemini, Perplexity and Google AI Overviews into one number, because we have no defensible basis for the weights that would require. A business invisible in one and strong in another has a specific problem, and averaging hides exactly the thing worth acting on.

A first-place recommendation and a passing mention are recorded differently and shown separately. A negative mention does not count as visibility. Aliases, trading names and parent companies are resolved by hand at setup and a client can correct them. What we do not yet do is recompute history when a resolution rule changes — the report is annotated instead.

One response, one observation. A business named five times in a single answer counts once. Share-of-voice denominators use every brand named in the answer, not just the competitors a client chose — otherwise the client sets their own denominator.

A citation is a source the engine itself surfaced — a linked source or an entry in a sources panel — attributed to the domain. We do not infer citations, we do not count a plain-text URL in prose as one, and we do not reproduce verbatim AI Overview text as evidence, on a supplier restriction we chose to build around rather than argue with.

A refusal, timeout or malformed response is recorded as a FAILURE, not as an absence of the brand, and it is excluded from the denominator. The failure count appears on the report. Counting a timeout as 'you were not recommended' would manufacture a decline out of an outage.

Clients get the exact phrase set, the model and the run date, which is enough to rerun it themselves. Full unedited outputs and timestamps are available on request but are not yet exported by default. That should be a button, not a request, and it will be.

A change to the phrase set or the scoring starts a new series rather than continuing the old line. We do not silently restate history, and we do not present readings taken with two different instruments as one continuous trend.

None. We do not claim a content change caused a movement. We have no control group and no counterfactual, so temporal sequence is all we have, and temporal sequence is not causation. What we report is what we did and what the reading was afterwards, labelled as exactly that.

We do not run autonomous agents that edit content or create pages off the back of a score. Every change is written by a person or reviewed by one before it ships. If we ever automate an action from a measurement, the repetition threshold, the minimum effect size and the human-review requirement will ship at the same time and be visible to the client — not added afterwards.

We do not know, and we will not claim otherwise. We have not demonstrated that a higher reading predicts referral traffic, pipeline or revenue while controlling for brand size. Improving the score currently predicts one thing reliably: a later improvement in the score.

It should not, and we do not describe it as one. It is a directional monitoring tool for a configured panel. That limitation is printed at the same size as the number, on the report and on this page.

Any of these, and we would say so publicly: a measured run-to-run variance large enough that ordinary movements are indistinguishable from noise; a test-retest agreement rate below what we would accept from a supplier; a cross-vendor comparison showing our readings disagree materially with two others and we cannot show ours is closer to what a user sees; or a first outcome study finding no relationship with anything a client cares about. If none of those would change our mind, the method has not been tested.