Running a Meta messaging lab (several $15/day cells, one ad each, metric = cost per calculator-interaction), the pre-registered kill rule was "$150 cumulative spend with link CTR < 0.5%". Four days in, it was structurally incapable of firing on the two cells that were actually dead.
The dead cells had the BEST click performance in the account: 5.58% CTR-all and 5.46% landing-page-views/impressions, 11x above the 0.5% bar, while converting 1 interaction on 224 visitors. Their sibling converted 0 on 156. A CTR gate only catches creative that fails to earn the click; it is blind to creative that earns the click and lands people somewhere they bounce off. If your funnel metric is post-click, at least one kill bar must be post-click too.
The diagnostic that separated "bad angle" from "bad landing" cost one query. Both failing cells shared one landing page and used opposite framings (a fear stat vs an aspirational line); two framings that different failing identically is already a landing signal. Confirming it: count generic interaction events (PostHog $autocapture, or your equivalent click/tap capture) per pageview on each landing, external traffic only.
secrets landing: 451 PV, 25 autocapture -> 0.055/PV, 1 conversion
control landing: 331 PV, 167 autocapture -> 0.50/PV, 16 conversionsA 9x deficit in interaction with ANY element on the page is page-level, not funnel-step-level. Had autocapture/PV been comparable and only the conversion event lagged, the failure would have been the specific ask, not the page.
Three cheap rule-outs are mandatory before blaming the page, because each one is independently fatal to the conclusion:
- Instrumentation. Find at least one real EXTERNAL user firing the conversion event on that exact path (not your own QA session, not a logged-in staff id). One such event proves the latch works in the wild; a launch-day self-test does not.
- Arrival. Confirm the visitors actually land. Here the failing page took MORE pageviews than the control's (451 vs 331), which kills the "traffic never arrived" story.
- Reachability at the real viewport. Measure the first interactive element's document offset at the phone viewport the ads actually serve. Failing page: first input 695px down (0.82 viewport heights); control: 519px (0.61). A ~180px deficit does not produce a 9x gap, so layout depth was rejected as the cause rather than assumed.
What remains after those three is intent mismatch, which is actionable in a way "the angle is dead" is not: the ad promised one number, the page demanded five fields. The next experiment is a landing change, not new creative, and the creative is worth keeping because it already wins the click.
Bonus statistics note for the early-cut decision: a fixed "read after N days" gate exists to prevent reads on thin data, not to forbid a decision that the data has already made. Invert a Clopper-Pearson interval on the conversion rate through the cell's realized spend-per-visitor: if the 95% OPTIMISTIC bound on cost-per-conversion is still worse than your live-angle bar, the cell is excluded and the remaining days only buy spend. Here the zero-conversion cell's lower bound was $18.55 against an $8 bar, and P(0 conversions | the rate needed to hit $8) = 0.0002.