Cold prospecting on tCPA ran 60β90% above CPA target for five straight weeks. I almost killed the campaign on day 30. I didn't, and by day 70 it was delivering 40% more volume at a lower CPA than the manual CPC holdout arm.
The 50-conversion learning threshold Google publishes is a floor, not a graduation certificate. Campaigns I operate across β©30Mββ©80M/month accounts routinely exit "learning" status at 50β60 conversions while still thrashing β CPCs swinging 40β80% day over day, CPA spiking 2x on Tuesdays and collapsing on Thursdays. The behavior that actually looks stable starts closer to 80β120 conversions within the attribution window. At $2Kβ$8K daily spend, that gap can mean 4β6 weeks of genuine instability that most operators misread as algorithm failure.
The most expensive mistake I see is adjusting the tCPA target mid-flight the moment CPA spikes for two consecutive days. Every target change resets the learning window. I've watched accounts stay in permanent learning-adjacent status for 6+ weeks because someone moved the target 15% every time Tuesday looked bad. The actual fix is to set your launch target 20β30% above your real goal, leave it alone through day 35, and let signal density accumulate. The algorithm is over-bidding on branded and retargeting-adjacent queries in the first 7β10 days because that's all the prior signal it has β it will find mid-funnel non-brand volume, but only if you don't starve it by resetting every week.
One thing that cut the thrash window noticeably: broad match during the first 21 days. Counterintuitive when you're watching spend, but broad match feeds more auction-level signal per day than phrase or exact. I shift to tighter match types after day 21 once the algorithm has enough data to bid those queries without burning the target.
The full 90-day holdout covered three audience temperatures β cold prospecting, warm retargeting, and lapsed purchasers β and the results split in ways I didn't expect. The retargeting arm stabilized by day 21. The lapsed purchaser arm never fully stabilized at all.
I wrote up the full breakdown β including the attribution window trap that kept one account flying blind on 40% of its actual buyers, and where this approach falls apart entirely for thin-SKU catalogs β over on themedilog.
