What the ± on a score means

AI answers vary

Ask an engine the same question twice and the answer can name a different set of brands. A single run is an observation, not a rate. So every score AEOTrack shows carries a range — "62 ±8", the range to expect between two checks — that says how much the number depends on the sample it was built from. The sample is stated beside it: "from 25 AI responses", and on the Dashboard "5 engines × 5 questions × 1 answer each".

How it is computed

The range is a 95% interval on the score itself (the method since scoring version 6, 27 September 2026, unchanged in versions 7 to 10). Three estimates are computed and the widest is shown:

1. A question-level bootstrap. The tracked questions are resampled with replacement 2,000 times, each question carrying its engine answers intact, and the score is recomputed each time. This answers "how much does the score depend on which questions happen to be in the set?" and needs at least two tracked questions.
2. A within-answer bootstrap. The individual engine answers inside each question and engine cell are resampled. Each check takes one answer per engine, so for every check since that policy this estimate is zero wide; it still applies to older checks that took several samples.
3. A floor from the mention rate. A Wilson interval on how often you were named at all, rescaled onto the score. This keeps the range honest at the small sample sizes a real account has — ten questions across five engines is fifty answers — and it is why a score of 0 is never shown as "0 ±0": with no mentions the floor uses what a first mention would earn.

The range is re-centred on the score and clamped to 0–100. Fewer usable answers means a wider range; a check where two engines were dark is less certain than a full one, and the number says so. Answers that were cut off before naming anyone are excluded rather than counted as misses, and a failed answer is never a miss: the engine's last good answer stands in for it.

What the range does not do is predict next week. The engines re-search the live web between checks and can rank a different list; that week-to-week movement shows on the trend, where a move inside the range is labelled "within noise". The full formula, with a worked example and the version history, is on the public methodology page.

Older points on a trend keep the range their scoring version produced, so ranges are only compared across points with the same version.

How to read a change

The ± is the range to expect between two checks of the same questions: a move inside it is normal variation, not a change in your visibility. Compare ranges, not points. A move from 40 ±8 to 46 ±8 is inside the noise; a move from 40 ±8 to 58 ±7 is real. The trend chart shades the range so this is visible at a glance, and the "What changed" card only lists movements that cleared it.

Why alerts wait for it

Score-change alerts fire only when the move is at least your threshold and the new range no longer overlaps the old one; a move measured on a check where engines were dark never fires at all, and no drop is reported in the first 48 hours after your first check. That is deliberate: an alert on every fluctuation trains you to ignore alerts. If you have not had one in a while, it is because nothing significant happened, which is information too. The Automation guide has the full rules.