๐Ÿš€ OrcaScope private beta โ€” first 10 design partners get 3 months free Join now!

Product2026-07-24~2 min read้˜…่ฏปไธญๆ–‡็‰ˆ โ†’

How we score AI visibility: the full story behind the 0โ€“100 number

The six steps from monitoring prompts to the 0โ€“100 score: mention detection, citation parsing, small-sample down-weighting and versioning โ€” in plain language.

Where does the big 0โ€“100 number on the dashboard come from? This is the plain-language version. The normative spec lives in full on the methodology page โ€” if the two ever disagree, the methodology page wins.

From a question to a score, in six steps

  1. Monitoring prompts: every brand gets a set of questions buyers actually ask AI โ€” "which vendors are worth shortlisting in X".
  2. Continuous sampling: our worker regularly sends that set to every live engine (8 as of this update: ChatGPT, Kimi, DeepSeek, Gemini, Doubao, Qwen, Perplexity, Grok) and records the full answers.
  3. Mention detection: we check whether the answer body names the brand, including common aliases. The current version deliberately does no sentiment analysis โ€” a mention is a mention; tone scoring is on the roadmap, not in the number.
  4. Citation parsing: some engines append reference links to their answers (Kimi's references, for example); we parse them and count citations per source domain โ€” "who the AI considers authoritative" is among the most valuable inputs to a content strategy.
  5. Normalization: mention performance per engine becomes a 0โ€“100 engine sub-score.
  6. The overall score: a weighted average of the live engines' sub-scores; when an engine's batch has too few samples, its weight is dynamically reduced so small-sample noise cannot distort the total.

Three honesty-first design choices

  • Insufficient data says so. When the sample is too small, the UI shows "not enough data" instead of fake precision.
  • Versioned scoring. Methodology iterations bump the scoring version (score_version); historical batches keep their version tag, so trends always compare like with like.
  • The methodology is public, in full. Digiday (May 2026) quoted a buyer's complaint: โ€œrun the same prompt through three tools and you get three different answers.โ€ When measurement is a black box across the industry, scores are guesswork. Publishing our full methodology is a deliberate, white-hat choice.

What we refuse to do

No volume-gaming (re-sending the same prompt to inflate mention rates), no prompt injection, no content poisoning, no fabricated reviews. Scores exist to guide decisions โ€” a manipulated score is worthless.

For how to act on a score, start with What is GEO; for the full rules, read the methodology.

Theory read โ€” now measure reality: what do AI answers say about your brand today?

Run my free checkup

How is the score computed? Full methodology โ†’

How we score AI visibility: the full story behind the 0โ€“100 number ยท OrcaScope