Detecting silent shave needs a cohort, not a hunch

Detecting silent shave needs a cohort, not a hunch
Shave/scrub = network suppressing legitimate conversions to lower paid totals.
—
Expected conversion-rate noise on stable traffic, 7-day window: +/-8% ▏▏▏
Drop that survives 3 consecutive windows on flat traffic: signal, not noise
—
Networks flagged for shave in samples showed approval rates falling 12–22% with no source change (n=19 documented cases)
—
Control: split 15% of identical traffic to a second network as a baseline mirror.
Read: only a parallel control offer turns a suspicion into a measurable shave figure.