Dissecting winners into components
Yesterday you crowned winners with a clean read. Today you take them apart — because the prize was never the ad. It's the element inside it that you can carry into the next hundred ads.
A winning ad is a bundle: a few load-bearing elements plus passengers along for the ride. Dissection means finding which element did the winning — so you stop cloning ads and start reusing elements.
Day 16 gave you a fair race and a definition of "winner" — the deepest money event (CPA / ROAS), with thumbstop, hold rate and CTR as leading indicators, gated by a minimum-data threshold. It also drew the line between the two loops: Meta's inner loop picks among the creatives you already gave it; your outer loop learns what to make next. Crowning a winner is where the inner loop ends. Dissection is where the outer loop begins. This is the step almost everyone skips — and it's why almost everyone's "creative testing" never compounds.
1The winner is a bundle. Unbundle it.
When an ad wins, it didn't win for "being good." It won because some subset of its parts did the heavy lifting, while the rest came along for the ride. The whole job of dissection is to separate the load-bearing elements from the passengers.
You can only do this because of Day 4 — the Creative Genome. Every asset was tagged at birth on the same nine axes: Concept, Persona, Message Angle, Hook type, Visual treatment (hi-fi / lo-fi), Production source, Format, CTA, and Funnel stage. You cannot dissect what you didn't label. If you launched a batch of untagged near-duplicates (the Day 10 mistake), today is the day that negligence sends you the bill — there's nothing to pivot by, and the learning is already lost.
So dissection isn't a creative act. It's a read. Take one winner and lay it out axis by axis, scoring each element by how much it appears to be driving the result. Here's a worked example from a 12-asset batch for a fictional sleep-supplement brand, "Drift". The winner: a vertical UGC clip aimed at the exhausted-new-parent persona.
Notice what the verdict column refuses to say. It doesn't say "this ad is great." It says which axes look load-bearing. The CTA was Shop Now — but so was the CTA on three losers, so it earns "neutral," not credit. The production source was an AI UGC avatar — but a hi-fi human-shot loser used the same angle and also did well on hook rate, so "AI" isn't the lever either; the lo-fi treatment is. You're not scoring the ad. You're scoring the elements, one at a time.
2One winner is an anecdote. Read the batch.
A single dissected ad gives you a hypothesis, not a finding. The real signal lives in the contrast across the whole batch: which elements recur among the winners, and which recur among the losers? An element that shows up in three winners and three losers is telling you nothing. An element that's in four of your top five and none of your bottom five is shouting.
Stay with Drift's 12-asset batch. Pivot it by genome tag and you get a far stronger picture than any one autopsy can:
- Lo-fi UGC treatment: appears in 4 of the 4 winners, 0 of the 5 clear losers. Average CPA €19.10 vs €34.60 for hi-fi — a ≈45% CPA gap. Loud signal.
- Pain-agitate-solve angle: won across two personas (new parent and shift-worker), n=5 ads. Travels — that's worth more than a within-persona fluke.
- 9:16 vertical: every winner; but every asset this batch was also mostly 9:16, so the contrast is weak. Confounded — hold it, don't bank it.
- Hi-fi studio static: lost everywhere — bottom 3 of 5 on CPA. A Graveyard candidate — tomorrow's Creative Memory keeps a section for proven dead ends — for this brand-and-persona, pending one confirmation.
This element-level reading is exactly the chain Day 3 taught you to diagnose: a strong hook rate that didn't convert points at angle or offer, not at the first frame. And it's exactly what Day 18 will feed into the brief. But before you bank any of it, you owe yourself the uncomfortable caveat.
3From reads to candidate insights
Correlation is not isolated cause. Your winner changed several axes at once versus the losers — different hook and angle and treatment — so you genuinely don't know which one carried the day, or whether it was the interaction of two of them. Be honest about this. You firm a candidate insight up three ways:
- Repetition across batches — if lo-fi UGC wins again in batch 4 and batch 5, the confound washes out.
- The occasional clean element-level A/B — same ad, only the treatment swapped. Change one thing, then give it enough conversions to clear noise (Meta's learning phase wants roughly 50 events per cell; the same logic applies here).
- Sample size — n=14 across 3 batches beats n=2 in one.
So the output of dissection is never "facts." It's candidate insights, each scored by confidence and sample size. Pull batches B1 and B2 from the tracker alongside B3 and the lo-fi + pain-agitate set totals n=14 — the cross-batch CPA edge settles at −28%, smaller than this batch's −45% but far more trustworthy:
- HIGH · n=14 · 3 batches Lo-fi UGC + pain-agitate open → −28% CPA for the exhausted-parent persona.
- MED · n=5 · 1 batch Pain-agitate may generalise across fatigue-led personas — re-test on shift-workers.
- LOW · n=2 · 1 batch 9:16 beats 4:5 — confounded; design a clean A/B before trusting.
That confidence tag is not bureaucracy. It's the dial that decides, in Day 18, how much budget the next brief bets on each element — and it's the Day 16 minimum-data threshold doing its job at the element level rather than the ad level.
A novice drinks a great wine and says "nice wine." A sommelier says "cherry on the nose, oak on the finish — the tannin is doing the work." Same glass, different read: one stops at the verdict, the other names the elements. "This ad won" is the novice's sip. "The lo-fi UGC pain-agitate open won; the CTA was a passenger" — that's the read you can pour into the next bottle.
Dissection is a pivot table before it's anything fancy. Take the creative tracker from Day 4 — one row per asset, columns for the nine genome axes plus CPA, hook rate, hold rate, CTR — and pivot it: rows = a genome axis (e.g. Visual treatment), values = avg CPA, win-rate, and ad count. The contrast jumps out. Here's the Drift batch pivoted by treatment:
Past a handful of assets you graduate to creative-intelligence tooling that does the tagging and the pivot automatically — the category is the durable part; the names below are a snapshot of who leads it as of mid-2026:
- Motion — auto-tags every running creative and ships a weekly leaderboard; the standard for a review ritual.
- Rule1 — ~20 dimensions when Motion's buckets feel too coarse and you want granular element-level attribution.
- Atria — grades each creative A–D across ROAS / CTR / hook rate / retention and gets prescriptive: it tells you which element to change next.
- Foreplay — works the input side, mining competitor ad libraries into tagged swipe files so your explore hypotheses aren't pulled from thin air.
These read your ads the way you just did by hand — the manual pivot is still the mental model you're automating, and you should run it by hand once so you can tell when a tool is wrong.
Now you (15 min): open your Day 4 tracker, take your most recent finished batch, and pivot it by Visual treatment and by Message angle. Write exactly three candidate insights in the format above — CONFIDENCE · n · batches · element → effect. If you can't produce them because tags are missing, that's your real homework: backfill the genome tags before tomorrow.
"This ad won — make more like IT." So the team clones the whole winner: same actor, same script beat, same B-roll, same caption — and bottles the noise right alongside the signal. You re-pay for the passengers (the neutral CTA, the incidental music, the specific face) as if they were the win, and you carry forward whatever was a lucky confound. Worse, near-identical clones fatigue together — Day 9's lo-fi lesson plus the fatigue clock means five copies of one ad all die the same week, and you've learned nothing transferable when they do. The discipline that separates a real creative engine from a copy-machine is isolating the winning element, not duplicating the winning artefact. Make more lo-fi UGC pain-agitate opens — across new actors, scripts and concepts — and you've grown an asset you can spend against for a year. Clone the ad, and you've bought one more dying twin. That distinction is your edge.
Today's recap — 30 seconds
- A winner is a bundle; dissection separates the load-bearing elements from the passengers — and it's only possible because every asset was genome-tagged at birth (Day 4).
- Score elements, not ads: lay one winner out across the nine axes and mark each as drives-up / neutral / loses.
- One winner is an anecdote — read the whole batch: which elements recur in winners and not in losers. Contrast is the signal.
- Correlation ≠ isolated cause. Firm candidates up with repetition across batches, clean element-level A/Bs, and sample size — watch for confounds.
- Output = candidate insights scored by confidence + sample size (HIGH / MED / LOW · n · batches), which set how much the next brief bets.
- The expensive error: "make more like IT" — cloning the artefact (and its noise) instead of reusing the element.