A day 3/7/14 decision tree for a running ad test

Why the cadence matters

Reading conversion data too early produces false reads, because the sample is too small. Reading click-through rate too late wastes budget on an ad that was never going to work. The fix is a fixed read cadence, with a different question asked and a different rule applied at each point, agreed on before the test launches so no one is tempted to move the goalposts once real numbers come in.

Day 3: hook signal only

At day 3, only click-through rate is read; conversion volume is still too small to be meaningful. A CTR below 0.5% on 2 or more days kills the ad outright, no exceptions for budget adjustments. CTR between 0.5% and 1.5% continues as-is, no action. CTR above 1.5% gets flagged as a potential hook winner to watch. Separately, if cost-per-thousand-impressions runs more than 3 times the pool's median, that's treated as an audience or bid problem, not a creative problem, and gets investigated before any creative gets blamed.

Day 7: iterate on the winner

By day 7, enough link clicks should have accumulated (the target is 100+ per ad) to read CTR directionally rather than as pure noise. The bottom third of the pool by CTR, with no sign of recovery, gets killed. A clear top performer (CTR more than 1.5 times the nearest competitor in the pool) gets its hook angle promoted into a new ad set at double budget. If the whole pool is flat (less than 0.5x variance across it), the conclusion isn't "try more creative": it's that the offer itself is the problem, and the fix is back at the corpus-mining and frame-falsification stage, not in new ad copy. Audience and creative are never changed in the same iteration; it breaks the ability to tell which change caused which result.

Day 14: promote or pivot

At day 14, a winner needs 200+ link clicks with a stable CTR and cost-per-click within target before it's cloned at double budget (with losing ad sets archived, not deleted, so the data stays available). If every variant is still flat after 200+ clicks per ad, the test stops entirely: the fix is a new corpus-mining pass and a hard look at whether the offer structure (price, commitment type) is the actual blocker. A different pattern, high CTR but low conversion, means the ad is doing its job and the landing page is the problem, so the next test is a landing-page split test, not new ad variants. And if one specific audience-scene variant clearly outperforms the others, that scene becomes the primary audience to build around, with 5 or more new variants generated inside that one scene specifically.