AI citations are volatile — the same prompt can name you today and skip you tomorrow. CitedOS's answer is statistical honesty: sample many times, report a range, and only flag a change when it clears the noise.
Multi-run sampling
Each prompt × engine is run M times (M≥5, configurable) per collection cycle, and every raw run is stored. CitedOS aggregates those runs into a distribution — a mean with a spread — rather than trusting any single answer.
Wilson confidence intervals
Rates like presence and share of voice are proportions, so CitedOS uses the Wilson interval — the right tool for proportions at small sample sizes, where the naive interval misbehaves. The interval is the band you see on every metric and chart; the true value lives inside it with ~95% confidence.
Freshness windows
The engines' underlying indexes (ChatGPT/Bing) refresh roughly every 24–72 hours. Re-running faster than the data can change just burns cost and adds no information, so CitedOS caches within the freshness window and stamps every metric with when it was last measured.
Change detection
Alerts fire against the confidence band, not the raw line. A move that stays inside the interval is noise and stays quiet; only a shift that clears the band is surfaced as a real change. For the why behind all of this, see the GEO guide.