How the visibility score is calculated.
Your visibility score is a weighted average across the seven AI engines Discoverable measures. The weights come from each engine's published reach, dampened so no single platform swamps the score, and every change to how we measure is dated and published below.
A check is a sample
Each tracked question is sent to the AI services your plan measures, for the market on your account. Ten questions across seven services produce up to 70 answers.
The results describe those captured answers. They are not all customer searches, and they are not every possible AI response.
Three different observations
Every figure in your report is one of these three. They are counted apart because they mean different things, and an answer can be more than one of them at once.
- Named in an answer
- The answer mentions your business. It does not need to link to your website.
- Cited in an answer
- The captured answer includes a source link to your website. An answer can both name and cite you.
- Found on a checked page
- Your business appears in the readable content of a source page. A page that could not be read is unavailable, never absent.
What one check does
A check runs your buyer-style prompts through each engine, brand-blind: the engine never sees which brand we are measuring while it answers. Each answer is recorded in full, along with whether your brand was named, how prominently, and which pages the answer actually cited. Naming and citation roll up into a 0 to 100 score per engine, and the per-engine scores combine into one overall score using the weights below.
Where an engine offers no public API (Copilot, Google AI Overviews and Google AI Mode), we read the public, logged-out answer surface itself. Every reading in your report is labelled with the surface it was collected from, because an API and a consumer app can answer the same question differently.
The engine weights
The visibility score is a weighted average, not a plain one. Each engine counts in proportion to its published reach, and when a check cannot measure an engine the weights are re-balanced across the engines it did measure.
Weights are derived from published reach figures, not editorial taste. We take each engine's most recent published user figure, dampen it by square root so the largest platforms count more without drowning out the rest, and normalise the whole set to sum to 100.
| Engine | Weight | Published reach | Source |
|---|---|---|---|
| Google AI Overviews | 28 | 2.5 billion monthly users | Sundar Pichai, Google I/O, May 2026Official figure |
| Gemini | 18 | 1 billion monthly users (app) | Google, August 2026Official figure |
| Google AI Mode | 18 | 1 billion monthly users | Google I/O, 2026Official figure |
| ChatGPT | 17 | 900 million weekly users | OpenAI, February 2026Official figure (weekly, so a floor for monthly) |
| Copilot | 7 | ~145 million consumer users | Industry analyses, 2026. Microsoft's 420M figure counts every touchpoint and is not consumer-comparable.Estimate |
| Claude | 6 | ~100 million monthly users | Midpoint of third-party estimates; no official figure published.Estimate (lowest confidence input) |
| Perplexity | 6 | ~100 million monthly users | CEO figure via Sacra, April 2026Company statement |
| Total | 100 |
When a run cannot measure an engine, the weights are re-balanced across the engines it did measure, and the report says which engines those were. We review the reach figures and this table quarterly, and any change to the weights is treated as a change of methodology: scores are never compared across it. The exact vector in force is stamped on every check as chatgpt:17,.
What one answer is worth
Within each engine, every answer scores on one ladder that rewards being recommended first and decays with position. The per-engine score is the average across your answers, and the overall score is the weighted average of those. Only sources the answer actually attributed count as citations. Sources an engine merely searched are shown in your report, labelled, but never scored.
| What the answer did | Points |
|---|---|
| Named first among the brands in the answer | 100 |
| Named second | 80 |
| Named third | 65 |
| Named fourth | 50 |
| Named fifth | 42 |
| Named sixth or lower | 38 |
| Named, in an answer that puts the brands in no order | 45 |
| Cited as a source, not named in the answer | 50 |
| Not named and not cited | 0 |
An answer that failed to collect and a surface that rendered nothing at all are not zeros. Neither is scored: they are reported as coverage, and your report says which engines carried the score.
Share of voice
Share of voice is not a straight count of mentions. Each mention is weighted by where the brand sits in the answer, so a brand named first counts for more than the same brand named last. Shares are apportioned across every brand any answer named, and add to 100%.
A mention counts 1 divided by the square root of its position in the answer, so a first place is worth about 1.4 second places, and the totals are apportioned so the column adds to exactly 100%. The comparison set is not a list you pick: it is every brand any answer named across the questions checked, competitors you have not thought of included, which is why your share falls when AI starts naming a new entrant.
Two other columns stay plain counts, and mean what they say. Named in counts the answers that named you. Where AI lists them is the position you were first given among the brands in an answer. Neither is weighted, and neither is a ranking of how strongly AI recommends a brand.
Changes to how we measure
When the definition of the score changes, we do not splice the old and new numbers into one line. Every check is stamped with the methodology it was measured under, charts mark the break, and no change figure is shown across one. The dated history:
v5 · 31 August 2026 · all seven engines in one weighted score
The score now includes all seven engines we measure, weighted by each engine's published reach, where before it averaged four engines equally and reported the other three alongside. The change was made once, deliberately, and the weights and their sources are published on the methodology page. A score either side of this point is a different quantity rather than a better or worse one.
v4 · 9 August 2026 · collected from the consumer surface
Some engines are now read from the logged-out consumer surface rather than through their own API, which changes what a cited source IS on two of them. On Perplexity the API returned one list under two names, so what counted as cited was really everything it searched; the consumer surface distinguishes them and roughly a quarter as many sources now count. On Gemini the API returned redirect links that never named the destination page, and the consumer surface returns the real page, so the same source is a different row before and after. Points either side of this are different quantities rather than better or worse ones.
v3 · 8 August 2026 · cited, not retrieved
Sources and the score moved from everything the engine searched to what the answer actually attributed, and engines were given a higher search ceiling at the same time. Source lists shrink and the citation gap grows for definitional reasons, and the score is a different quantity either side of this point rather than a better or worse one.
v2 · 5 August 2026 · brand-blind answering
Engines stopped seeing the brand name while answering. Points before and after are different measurements.
Work and later results
Marking a step done, recording a pitch you sent or adding a published URL records work. A citation found afterwards is a separate observation with its own date and its own evidence, and the two are kept apart in your work history.
A score change after an action does not prove the action caused it. Citations already observed stay in the history even when a later check does not find them again.
Evidence availability
How long a captured answer's full text is kept depends on the service it came from, and on the terms of the route it was captured through. The shortest window any route is granted is 30 days; answers read through a service's own API are kept for as long as the account is open.
When an answer's text expires, the measurement taken from it does not. The reference, the date, the services and the counts stay in your report, and the answer itself is marked unavailable rather than quietly disappearing.
The free check
The free, no-login check runs on Claude alone, so its number is a Claude snapshot rather than the weighted score, and it is labelled as one. A free account also gets a monthly check of its core prompts across all seven engines; paid plans measure all seven engines every week.
See where you stand in AI search, and fix it.
Run a free check in under a minute. No card, no login to see your score.