Two things that cut the insanity on our side.
First, kill the calendar-based reruns. Rerun on change events instead. New creative, new offer, or a budget step change north of 30 percent are what invalidate your last read. Quarterly is arbitrary, and monthly will burn out the team.
Second, stop treating each test like a binary verdict on the channel. Use it as a calibration factor. Take incremental ROAS divided by attributed ROAS per channel, then apply that factor to your attributed numbers between tests. You get a continuous rough-right read instead of a quarterly cliff. The tests become about updating the factor, not relitigating the whole channel.
If you can swing it, keep a small permanent holdout running. Even 5 to 10 percent of geos held out continuously gives you a trend line, and it kills the "is this just noise" arguments fast.
On when to cut: most geo holdouts at normal budgets can't resolve a difference smaller than 20 to 30 percent. Anything inside that range shouldn't move spend. Retargeting reading far below attributed is expected, not a finding. Only cut when incremental ROAS sits below your real payback threshold, not when it merely disagrees with the platform.
One check before trusting any of it: look at the pre-period trends for test and control geos. if they weren't tracking each other before the holdout started, you're measuring the geos, not the campaign. That's the first thing that'll save your sanity.
