The creative testing framework we run

This is the entire method behind our bench, the one that produces the 83% kill rate, the six-day median funeral, and the winners in the client stories. We publish it because the method isn't the moat; running it every single week is. The market context that makes it necessary: top advertisers now ship 2,400 to 2,600 creatives a quarter, 78% of top-quartile campaigns refresh at least weekly, and video plus playables together took two-thirds of gaming ad impressions in 2025. The arms race is real; the framework is how a small team stays in it.

Step 1: every concept is a written hypothesis

No concept enters the bench as "let's try this." The unit of work is a hypothesis card: the hook family, the audience feeling it targets, the metric it should move, and the number that would kill it. A card for a fail-state hook might read: players will stop to correct a visibly wrong run; expect IPM ≥ control +30%; kill if below control at 4,000 impressions or if cohort D1 lags control by 3+ points.Writing the kill condition first is the whole trick; it makes the funeral a clerical act instead of an argument.

Step 2: hook families, not one-offs

Concepts belong to named families: fail-state, choice, progress-bar, transformation, urgency, UGC-testimonial, and whatever this quarter invents. Families matter because learning compounds within them; a dead variant still teaches its family something, while a dead one-off teaches nothing. Portfolio rule: every monthly batch spans at least three families and two formats, and a batch is 20 to 40 concepts depending on how much traffic the account can feed honest cells.

Step 3: cells that don't lie

One variable per cell. Test hooks against hooks with endings held constant; test endings only when a hook is settled. Keep a standing control, the current champion, in every cell, and never let a test cell share budget with a scaling campaign, because the platform's optimizer will starve your test to feed its favorite. Statistical floor before any verdict: 3,000 to 5,000 impressions per variant, more when install rates run low. Verdicts before the floor are astrology.

Step 4: the three-gate read

A concept must clear three gates in order. Gate one, attention: hook rate, the share of viewers still present at second three. Gate two, action: installs per mille against control. Gate three, quality: the cohort's day-1 retention, because an ad that recruits tourists fails after the click, and that failure is the expensive kind. Plenty of our graveyard residents won two gates; the timer-pressure hook in theGraveyard won the first two by a street and died at the third by seven points.

Step 5: funerals, on a clock

The bench's median from launch to kill decision is six days. Losers are retired the day their kill condition trips, logged with their numbers, and the instructive ones get public epitaphs. Winners earn iteration, not retirement: shorten the opening, move the fail state earlier, re-cut the end card. Our median winner peaks on its second revision. Then, inevitably, fatigue arrives, the champion's IPM sags, and it too gets a dignified burial while its validated successor, already on the bench, takes the throne. No panic, because the pipeline never stopped.

What breaks this framework

Three things, all human. Approval layers that take longer than the six-day funeral clock. Falling in love with a concept, usually the expensive cinematic one, and granting it mercy extensions. And skipping the retention gate under scale pressure, which converts your ad account into a machine for buying churn. The framework survives every market shift we've seen; it does not survive being optional. Want it run for you? That's thecreative testing bench, priced flat.