Sensitivity (σ / sigma): a statistical measure of how unusual a number is versus the recent average. Higher σ = the rule needs a bigger departure before speaking up = quieter. Lower σ = more sensitive = more alerts. Around 2σ is the sweet spot - it catches genuine drops while ignoring day-to-day wobble.
Baseline window: how many recent complete days define "normal". Shorter reacts faster; longer is steadier. Most performance rules default to about two weeks.
Minimum impressions / conversions: a reliability floor. With very little data, CTR and CPA swing wildly for no real reason - requiring a minimum amount of traffic keeps every alert trustworthy.
Multiple (e.g. 3×): "this many times the normal level". Because it's relative to each account's own normal, one setting works across big and small accounts.
Materiality (±%): how big a change has to be before it's worth telling you about. A ±30% budget materiality ignores small tweaks and flags the moves that matter.
Must hold for X days: requires a problem to persist before alerting, so one odd afternoon never fires a rule.
Conversion lag: the most recent days are ignored so conversions can settle before judging - today's incomplete data can't look like a fake CPA spike.
MAD multiplier (k): how many median absolute deviations above the median counts as abnormal - a robust version of σ that one freak day can't distort. Higher = less sensitive.
Cooldown: after a rule fires, how long it holds off before it can fire again for the same issue.
Guardrails: extra conditions that make a rule smarter and quieter - e.g. the disapproval rule's policy-topic allowlist, or the network-leak whitelist for deliberate opt-ins.
Would have fired N times: a live preview on some rules showing how often the current settings would have triggered over the last 30 days, so you can feel the noise before saving.
