Creative testing dies two ways: no data, or constant thrash. Creative DNA is how you remember what worked without lying about sample size.
Two DNAs
Performance DNA - what angles and formats earn attention and efficient outcomes in your account.
Brand Visual DNA - palette, lighting, composition, camera personality so generation stays on-brand.
Lab vs live
Regenerate-with-Ralph and Creative Lab exist to explore, not to surprise your account at 3am. PAUSED push is the respectful default.
Also: a winning UGC that features a sold-out size is still a stock problem. See dead size creatives.
Minimum evidence
Learning periods and conversion counts exist for a reason. Promoting a tester on 12 impressions is pretending. The rotation helpers encode thresholds (learning days/conversions, Bayesian promote/retire) so humans have something better than guesswork.
Questions people actually ask
Does Creative Lab publish live ads without me?
Lab push is designed to land PAUSED / reviewable. Approval culture still applies.
Is DNA built from opinions or from ads?
Performance DNA is built from Meta ad performance with decay weighting. Brand Visual DNA is a separate visual system from store imagery and edits.
When does a tester become champion?
Rotation uses Bayesian promote/retire style thresholds and anti-thrash cooldowns, not a fixed “swap every week” rule.
Why did my library look pixelated before?
Meta thumbnail URLs are tiny. Sync prefers high-res creative sources and post images for catalog ads when static assets are missing.
How to run a Meta creative test that tells you something
Most creative testing produces confident conclusions from data that cannot support them. The fix is unglamorous: one variable, enough budget, and a decision rule written down before launch.
- Write the question down before you build anything. Not "which ad is best" but something answerable: does the problem-first hook beat the product-first hook for cold traffic. One question, one variable. A test comparing two ads that differ in hook, format, music and offer teaches you nothing, because you cannot attribute the difference to anything.
- Change exactly one element. Hook, format, opening frame, offer framing, or talent. Hold everything else identical, including the landing page. The single most common reason creative tests are inconclusive is that two or three things moved at once and the result cannot be read.
- Size the budget from your cost per purchase. You need enough conversions per variant to distinguish signal from noise, and that means roughly 50 conversions each before you should feel confident. Multiply your current cost per purchase by 50, then by the number of variants: that is the honest cost of the test. If the number is unaffordable, test on a cheaper upper-funnel event and accept a weaker conclusion rather than reading a underpowered purchase test.
- Use Meta's A/B test tool rather than eyeballing two ad sets. In Ads Manager, use A/B test so Meta splits the audience properly. Two ad sets running side by side share and cannibalise audience, which biases the result toward whichever one got cheap delivery first. That bias is invisible and it is large.
- Leave it alone through the learning phase. No edits, no budget changes, no pausing the one that looks bad on day two. Every edit restarts learning and invalidates the comparison. Plan for at least seven days, longer if your purchase cycle is long.
- Write the decision rule before you look at results. Decide in advance what result changes what you do: for example, if the winner beats the loser on cost per purchase by more than 20 percent, it becomes the new control. Deciding after you see the numbers is how people talk themselves into keeping the ad they liked.
- Record the finding as a pattern, not as a winning file. The asset will fatigue. The insight, that problem-first hooks beat product-first for cold traffic on this category, survives and briefs the next twenty ads. Keep a running list of what has won and lost, or you will re-run the same test next quarter.
- Rotate on evidence, not on the calendar. Refresh when frequency climbs and cost per result degrades against its own baseline, not because it is Monday. Swapping a working ad on a schedule throws away the learning you just paid for.
Doing this properly costs real money per test, which is why the pattern library matters more than any individual winner.
If Meta is the problem: Ralph reads Meta next to your stock every night: dead sizes still in ads, hours that lose money, sets gone stale. Each fix is staged for your yes. See Ralph · How Ralph handles out-of-stock sizes · Docs