How the measurement works

A scan asks buyer questions — the kind your customers would type — and those questions never contain the brand being tracked. If we asked “is Ritmi a good driving school?” the answer would be about Ritmi, and the measurement would be worthless. We ask “which driving schools in Tbilisi would you recommend for a first licence?” and then look at what came back.

You write the questions or generate them from a description of the business, and you see and can edit every one of them before anything runs.

These models do not answer the same question the same way twice. So every question is asked repeatedly: three times on a free scan, five times on a subscription, and on a subscription each question also goes to Gemini and Perplexity as well as ChatGPT.

A free scan is therefore between 6 and 15 answers; a tracked scan between 30 and 75. The headline is computed from all of them.

An answer counts as naming you only when your name is actually in it, after normalising case, accents and spacing. A domain also matches its host and its first label, so linear.app matches “Linear”. A plain name gets no invented nicknames: Atlas does not match Atlassian.

Each answer gets exactly one of three verdicts:

Recommended — you are there and actively put forward as an option. Mentioned — you are named but not suggested. Absent — your name is not in the text at all.

Only the first counts toward the number. A language model is never allowed to overrule the absence of your name: if the literal is not there, no amount of interpretation puts it back.

We compute the share of runs marked recommended for each question-and-engine pair, then average those shares. That stops one chatty question, or one engine that answered more times than the others, from dominating the result.

A percentage from a sample is an estimate, so we report how much to trust it. The 95% interval comes from resampling the collected answers 1,200 times and recomputing the headline each time. The result is deterministic: the same scan always produces the same interval.

A wide interval means the answers disagreed and you should not act on the number alone. That is drawn on the scale under every result rather than hidden in a footnote.

A provider that errors reduces coverage. It is never counted as “not recommended” — a failed request says nothing about your brand. If fewer than 90% of the planned answers came back, or an engine the tier requires produced nothing at all, the result is marked incomplete instead of being published as a headline.

Week to week, a number moves on its own. We only call a change meaningful when it is at least 5 percentage points and the resampled interval around the difference excludes zero. Anything smaller is shown and labelled as within the noise, in grey, because saying otherwise would be telling you something the statistics do not support.

Editing a question starts a new baseline, and results from different question sets are never compared — otherwise rewording would look like progress.

We do not resolve entities. Two spellings of one competitor are merged only when they differ by a bracketed note; deciding that “Auto Way” and “Auto Way Driving School” are one company would be a guess, and a wrong guess would quietly corrupt the ranking.

Brand names read out of answers are observations of what the text said, not verified facts about real companies. We do not train models on your data. And the stored answers are deleted after twelve months — the numbers they produced stay, as described in the privacy policy.