We sell a measurement tool.
So we publish our own ruler.
Plenty of tools now put a number on "visibility in AI search". Almost none publish how accurate that number is. Everything below is a value we actually measured — and we have not cherry-picked the good ones.
Per engine: 97.1 / 100 / 93.3 / 97.3%. Zero false negatives (no missed mentions).
99.6–100% per engine (0.4pt spread). Fixed up from 87.6% after real-answer testing exposed the same outlet being counted as different sources depending on the engine.
Exact match 93.8%. Every remaining mismatch is "could not determine = null" — zero misattributed positions. We would rather report nothing than guess a number.
How often a number in the report does not exist in the input data. Even with 0% mention rate and 0% citation rate — the strongest incentive to invent — nothing was invented.
Report layer measured 2026-07-24 (one report through the production path)
This is the reason this page exists. Publishing accuracy without the weaknesses is just advertising.
F1 76.8–96.5% — below our internal bar (85%) on two engines. The weakness is extraction from long answers, and we know the cause is the analysis model. Upgrading every task would fix it but costs roughly 10× more, so conditional routing is still on the shelf.
In one real measurement, one sentence that summarised several already-satisfied items was scored as claiming they were missing. This is a single-sample observation, not yet enough to conclude from.
The figures above come from 80 real answers across 4 engines (Japanese, English) plus a 16-item seed set. We do not yet have enough real answers in Korean, Chinese or Cantonese — so accuracy in those three is assumed to match JA/EN, not measured.
ChatGPT, Gemini, Claude, Perplexity. Answers differ by engine, so a number from one AI is not "visibility in AI search". Paid plans measure all four at once; the free plan measures the one you pick.
One question = one prompt × one language. Engine count does not multiply it — the same question goes to every AI on your plan. Measuring one prompt in five languages is five questions.
Mentions, position and cited domains are pulled from the actual answer text. Anything we cannot determine stays null rather than being filled with an estimate. Filling it looks better on screen — but it is not measurement.
We auto-score detection, position and citations against a human-labelled answer set (a golden set). The report layer has its own scoring (fabrication rate, grounding rate), and both fail the build in CI if they regress. "It feels better" is not evidence.
What we cannot estimate stays null. Demo figures on this site are always labelled as samples. We do not publish numbers about real companies that we have not measured. If we ever stop keeping those three, there is no reason to sell this product.