Glossary
The terms in the dashboard
Every figure in the dashboard follows a fixed rule. This page says how each is worked out: what counts, what does not, and how wide the margin of error is.
Last updated: 1 November 2026
Valid answer
An answer an AI engine actually gave. Only valid answers count towards the shares in the dashboard.
On every measuring day each question goes to each AI engine in your plan. We store what comes back, also when it goes wrong. An answer has one of four states: valid, error (the engine or our data provider did not answer), refusal (the engine would not answer) or no AI answer (Google showed no AI Overview for the search, or an empty one).
Visibility, share of voice, “named first” and stance count valid answers only, in the numerator and in the denominator. Errors and refusals show as the failure rate, searches without an AI Overview as the panel shown rate. We never delete a failed answer and never fill in a day: a day without valid answers stays a gap in the trend.
Only questions that are active today count, over the whole period. Archive a question and it leaves every figure, also backwards; its answers stay stored.
Visibility
The share of valid answers that name your brand. The dashboard calls it “Named”.
Visibility = valid answers that name the brand, divided by all valid answers. An answer that names the brand three times counts once.
A brand counts as named when its name or one of its spellings appears as a whole word in the text. We compare in lower case and without accents, and read “&” and “+” as “and”. You set the name and the spellings under Settings, Brands. A link to your website alone is not a mention: it counts among the cited sources.
For one question on one engine the margin of error is the Wilson interval; for several questions together it is worked out over the questions (see margin of error).
Named first
The share of valid answers that name your brand before every other tracked brand.
For each answer we order the tracked brands by the place in the text where each first appears. If your brand comes first, it is named first. When two names overlap, the longer one counts.
The denominator is every valid answer, not only those that name you. An answer without your brand therefore counts as “not named first”. We show no average position: second in a list of two and second in a list of ten are not the same.
Margin of error (Wilson interval)
The range that holds the true share with 95% confidence. The dashboard shows it as ± beside the figure.
AI engines do not answer the same question the same way every time. A measured share is therefore an estimate, and the margin of error says how precise it is. For one question on one engine we work out the Wilson interval at 95% confidence (z = 1.96). It stays between 0 and 100% even with few answers and with shares close to 0 or 100%.
Why 14 answers per question and engine
Unless you choose otherwise, the dashboard shows the last 14 days, today included. Each question goes to each engine once on every measuring day. One question on one engine therefore has at most 14 answers in this view, fewer after errors or on days without an AI Overview.
Once a day rather than several times, because repeats on the same day add little: in a test in October 2026 they mostly agreed with the first answer. So the 14 answers to a question come from 14 different days, and the money goes into more questions instead of repeats.
This is how wide the Wilson interval is with 14 answers:
| Named | Share | Margin of error (95%) |
|---|---|---|
| 0 of 14 | 0% | 0 to 21.5% |
| 1 of 14 | 7% | 1.3 to 31.5% |
| 3 of 14 | 21% | 7.6 to 47.6% |
| 7 of 14 | 50% | 26.8 to 73.2%, about ±23 points |
| 10 of 14 | 71% | 45.4 to 88.3% |
| 14 of 14 | 100% | 78.5 to 100% |
One question alone therefore says little. The figures gain meaning over many questions: for all questions together we work out the margin with a bootstrap over the questions (1,000 draws), because answers to the same question resemble each other and are not independent. That range is never narrower than the Wilson interval over all answers together.
With fewer than 20 answers the dashboard draws a point hollow and adds “small sample”; in the question grid, cells with fewer than 5 answers are hatched. No figure is hidden: a small sample simply has a wide range.
Sources cited and read
Cited is a page an answer links as evidence. Read is a page the engine opened while answering, cited or not.
For each answer we store the sources the engine returns, with address, domain and title, kept apart as cited and read. A domain’s “cited” share is the share of valid answers that cite at least one page of that domain; each answer counts once per domain.
The dashboard shows “read” only as a number of answers, not as a share, because not every engine says what it reads: ChatGPT and Claude do (Perplexity too, if you add it), Google AI Overview and AI Mode never. For a citation, Claude and Gemini also say which passage of the page they used.
We give each domain a kind, for example your own website, a competitor’s website, a directory, a news site, a review site or a forum. Subdomains count with their main domain. You can correct a domain’s kind for your project.
Fan-out searches
The web searches an AI engine runs before it answers. In the dashboard under “Searches”.
Many AI engines search the web before they answer, often with several search terms of their own. These searches are also called fan-outs. They show the words the engine uses to look for providers like you, which makes them good ideas for new questions and content.
ChatGPT and Claude reveal their searches, and so does Perplexity. Gemini, Google AI Overview and AI Mode do not; their answers are left out of the searches figures rather than counted as “no search”. “Answers that searched” is the share of these engines’ valid answers with at least one search, “searches per answer” the number of searches divided by those answers. We group identical search terms regardless of upper and lower case.
Stance
How an answer names your brand: recommended, listed, only as a source or with a caution.
Every valid answer that names a tracked brand is classified for each brand it names by a language model from Anthropic (Claude Haiku). It sees the question and short excerpts of about 250 characters around each mention, at most six per brand.
- Recommended: presented as a good choice, or as someone to contact or use.
- Listed: a neutral entry among options.
- Only as a source: only used as an information source.
- With a caution: named with a warning or a downside.
When several fit, this order decides: with a caution, recommended, listed, only as a source. The “recommended” share is the share of all valid answers that recommend the brand. An answer not classified yet counts as not recommending it; the dashboard says how many there are.
Panel shown rate
For Google AI Overview, the share of searches for which Google shows an AI Overview at all.
Google does not show an AI Overview for every search. A search without an AI Overview, or with an empty one, is not an answer that leaves you out but no answer at all. So it does not count towards visibility but towards the panel shown rate: valid answers divided by valid answers plus searches without an AI Overview.
Change (▲ ▼)
An arrow appears only when the change against the period before is bigger than the noise of the measurement.
We compare the chosen period with the same number of days just before it; unless you choose otherwise, 14 days against the 14 before. The margin of the difference comes from a paired bootstrap over the questions and is never narrower than Newcombe’s interval. If that range excludes zero, the dashboard shows ▲ or ▼, otherwise no arrow.
This applies to visibility and to the “recommended” share. One measurement on one day is a snapshot and never a result.
Data version
An internal counter per project that goes up with every change to the data. It keeps the dashboard from showing an outdated view as current.
A project’s data version goes up by one as soon as something is stored that a view shows: new answers, sources, searches, mentions and stances, costs, runs, changed questions, brands or settings. We keep finished views with their version and show them only while the version matches.
You never see the number itself. While a measurement is running and you have not changed the setup, the dashboard may show a view up to 10 minutes old; it then says “As of” and the time above it.