Methodology
How the score is calculated
Published before the assessment begins, so you do not have to take our word for it. Four components, fixed weightings, and an open list of what the method does not capture.
The formula
The formula
A weighted sum across four measures
Each component is calculated separately over the same set of answers, then multiplied by its own weighting.
- Presence and position× 45%
- Share of voice× 20%
- Citation as a source× 20%
- Accuracy and tone× 15%
Components
Four measures, every one of them checkable
The names and the weightings are fixed. If they change, the change is made public.
Presence and position
Weight45%How often your business appears, and how high it sits when it does.
- How it is assessed
- The share of answers that mention you, multiplied by where you rank inside them. Rank is counted only across the answers you appear in, so an absence is never charged twice.
- What it does not capture
- It does not separate a passing mention from a settled recommendation. Tone measures that.
Share of voice
Weight20%How much room you occupy against your competitors in the same answers.
- How it is assessed
- The share of the text devoted to your business compared with the other businesses mentioned.
- What it does not capture
- Longer text is not always more favourable text.
Citation as a source
Weight20%Whether your site is cited as a source where models give references.
- How it is assessed
- The presence of your address among the cited sources, counted only across models that cite.
- What it does not capture
- Two of the four models rarely cite sources, so the base for this component is smaller.
Accuracy and tone
Weight15%Whether what is said about you is correct, and how you are described.
- How it is assessed
- Coding of the statements about your business: correct, incorrect, or unverifiable. And how you are described: recommended, mentioned in passing, or described unfavourably.
- What it does not capture
- Tone is judged once per answer and stored with its reason, so next month's reading stays comparable. This is the most subjective component and we write it down when it moves.
Reading it
Three bands, fixed
The same threshold applies to the overall score and to each component.
| Range | Name | What it means |
|---|---|---|
| 67 to 100 | Strong | The presence holds across more than one model. |
| 34 to 66 | Gaps | You appear, but not consistently or not in the positions that carry weight. |
| 0 to 33 | Critical | Effectively invisible in the questions that carry commercial value. |
Uncertainty
What we do not know, written down
There is no accepted standard
Assessing visibility in AI is a new field. This method is ours, it is reasoned and it is public, but it is not an industry standard. We do not claim otherwise.
The variance is reported, not hidden
Every question is repeated within the same assessment, and the number of repetitions depends on the service chosen. When the repetitions disagree, we report the full range.
We cannot see inside the models
We assess only the answer that comes out. Why a model mentions a business sits inside it and cannot be reached from outside.
Coding tone involves judgement
Accuracy is assessed objectively, tone not entirely. That is precisely why this component carries the smallest weighting.
The base for citations is uneven
Only some of the models give sources. The citation component is calculated over that portion, and this is noted in the result.
The method will change
As the field matures, the questions and the weightings may be revised. Every revision carries a version and a date.
Version
The method carries a version and a date
So that an older result can always be read against the method that produced it.
- Last change
- The first published version. No preceding changes.
- Policy
- The weightings do not change inside an engagement. If they change, existing clients are notified and the earlier assessment stays readable.
Next step
The method is open. Try it on your own business.
We start from your customers' questions.