How we measure

Platforms

We ask the real assistants through their official APIs, with web search enabled: ChatGPT (OpenAI), Perplexity and Claude (Anthropic). Free scans run on 2 platforms; paid plans on 3. Model identifiers and per-call costs are configuration, not hardcoded — when providers ship better models, we switch without changing the method.

Sampling

A monitored brand runs up to 25 questions × 3 platforms × 2 samples per question — the double sample smooths out the natural variance of AI answers. The free scan is a snapshot: 10 questions × 2 platforms × 1 sample. Questions are generated in the language of YOUR market, anchored in the arena you confirm at scan time — never in a generic umbrella category.

The four identification states

A literal name match is not visibility. Every answer that names you is classified into one of four states — and each state is treated differently in the score and in the recommendations:

• Identified

  The answer demonstrates the assistant knows WHAT you are — it names your actual activity and engages with it. Full weight in the score.

• Discussed, but unverified

  The right business is discussed, even recommended — but the assistant flags missing public proof: independent reviews, documentation, company information. Half weight, and its own recommendation: publish the proof.

• Confused

  The answer is about different entities that happen to share your name. Zero weight — and the entities it mentions are never counted as your competitors.

• Absent

  Your name does not appear in the answer at all. Zero weight. Absence is not confusion — it is the position problem the category recommendations attack.

The score, in short

Each valid answer contributes a weight: recommended in first position 1.0 · recommended in positions 2–3 0.8 · recommended lower or unranked 0.6 · mentioned without recommendation 0.4 · not mentioned 0. Discussed-but-unverified answers count at half weight (×0.5); confused answers count 0; negative sentiment multiplies by 0.25. The visibility score is the average over all valid answers, on a 0–100 scale. Share of voice applies the same weights to every competitor detected in the same answers.

Honesty about nondeterminism

AI assistants do not answer the same question identically twice. A single scan is a photograph, not a verdict: scores can move a few points between runs with no change on your side. That is why the product's real value is the FILM — weekly scans, trends and alerts — and why we double-sample paid scans. We never smooth, backfill or hand-adjust results: what the models said is what you see, word for word, in the transcripts.