AI citations swing 40-60% month to month. A single point-in-time percentage isn’t a measurement - it’s a coin flip dressed up as a fact. So every number we show ships with four things competitors leave off: a confidence range, the sample it came from, when it was measured, and whether a change is real signal or just noise.
58%95% CI 41%–73%
We sample each prompt across engines multiple times and report the Wilson score interval - the honest band the true rate sits in 95% of the time. At small samples it’s wide on purpose; a wide band is the truth, a bare number is a guess.
58%95% CI 41%–73%
Every figure carries its n - the number of runs behind it. Below 3 runs we don’t report a rate at all; we say so. A paid cycle samples 3 runs per prompt per engine, so the merged number rests on many observations.
Search indexes refresh on their own clock. Stale data isn’t wrong, but pretending it’s current is. Every number wears a date, and goes visibly stale past its refresh window.
Competitors show a naive rank delta. We only call a difference real when the two confidence intervals don’t overlap - otherwise it’s a tie within the noise floor, and we say that too.
For a presence rate we use the Wilson score interval rather than the textbook normal (Wald) approximation. Wald collapses to a false zero-width band at 0% or 100% - dishonestly claiming certainty exactly where we know least (“0 mentions in 5 runs” is not “never mentioned”). Wilson stays inside 0-100% and keeps near-nominal coverage at the small samples AI measurement forces on us.
A number without a range is a claim you can’t check.
→Every CitedOS metric - in the dashboard, the export, and the MCP API - inherits the same honesty gate at the source. The range travels with the number everywhere it goes.