Methodology

Everything kwerve displays is recomputed from the raw answers we keep. This page explains how, so that every number in a report can be reproduced by hand.

What we measure

For each brand, a list of prompts is sent to each engine, several times over the period. Each send is one check: one prompt, one engine, one answer. The answer is then read to find the tracked brand and the declared competitors, and the position each of them appears at. Raw answers are kept: no number is estimated, all of them are recomputed from those answers.

The three metrics

Visibility rate: The number of checks whose answer names the brand, divided by the number of successful checks. Failed calls are excluded from the denominator and counted separately.
Average rank: The brand's mean position, counted from 1, across only the answers that name it. Lower is better.
Share of voice: The brand's mentions divided by the total mentions of every tracked entity (the brand and the competitors you declared) across all answers in the period.

The AI visibility score (0 to 100)

The score is a deterministic weighted mean of three components, each normalised between 0 and 1, then multiplied by 100 and rounded once, at the end.

Visibility (50%): the visibility rate as it stands.
Rank (30%): 1 at rank 1, falling linearly to 0 at rank 5, as 1 − (average rank − 1) ÷ 4. A brand never named scores 0.
Share of voice (20%): the brand's mentions divided by those of the most-mentioned entity. Level with the leader, the component is 1.

Share of voice is measured against the leader rather than as a raw percentage for a reason: a raw share falls mechanically as a category gains competitors. That would penalise a brand for its market rather than its performance.

A worked example

A brand named in 41 answers out of 106, at an average rank of 3.46, with 41 mentions against the leader's 71:

Visibility: 41 ÷ 106 = 0.387 → 0.387 × 50 = 19.34 points

Rank: 1 − (3.46 − 1) ÷ 4 = 0.385 → 0.385 × 30 = 11.55 points

Share of voice: 41 ÷ 71 = 0.577 → 0.577 × 20 = 11.55 points

Total: 42.44 → a score of 42 out of 100.

Grounded engines and text-only engines

Perplexity answers by searching the web: its answers cite sources, and that is where the domains under “Cited sources” come from. ChatGPT and Claude are queried as text completions, without browsing: their answers reflect what the model learned. This is a deliberate choice: it keeps the cost per check low, which is what makes a flat price possible. Every report states which engines were grounded.

What the score does not tell you

AI answers are not deterministic: two runs of the same question can differ. That is exactly why every prompt is repeated. A score compares only to itself, week on week, on the same prompt list and the same market, never across brands or categories. And nobody, kwerve included, can guarantee that an action will move a score up.

The weights above are the ones running in production today. If they change, this page changes with them.