Most accounts never read the results. This is James Dunn's approach to creative testing in the Andromeda era: one hypothesis per concept, five gates per test, and one principle underneath everything — direction before speed.
By James Dunn · First edition July 2026 · Diagram values marked illustrative where invented
Fifty near-identical variants a month is motion, not progress. A vector has magnitude and direction, and so does a good testing programme. A smaller number of deliberate tests, run properly and compounded over quarters, beats a high-volume spray of lookalike assets. Most programmes fail not because of bad creative but because of bad process: no hypothesis, no power calculation, no logged learning, and decisions called from gut feel three days in.
Meta's Andromeda retrieval architecture reads the visuals, audio, on-screen text and landing page of an ad to decide who sees it. Manual interest targeting is largely vestigial: if the ad shows a home, the system finds home-interested people on its own. What remains in the advertiser's hands is the input the machine learns from.
Creative diversity is a structural requirement. An account feeding the system a handful of creatives a month gives it little to match against a fragmented audience.
Ten crops of the same idea don't expand the matching space. The unit of testing is the concept: message, emotional territory, implied audience.
And this is not only a Meta story. Performance Max asset groups, TikTok Smart+, DV360's optimised rotation and Demand Gen all reward breadth of high-quality input and punish thin, repetitive feeds. Feed the machine deliberately varied concepts, measure which territories work, then reinvest in territories, not just assets.
Every creative that enters an account travels the same loop. Each gate protects the one after it: weak assets never reach the auction, unpowered tests never reach a read, noisy reads never reach a verdict, and no verdict escapes the learning library.
Before a pound is spent, score every asset against platform best practice: text coverage, legibility, visual hierarchy, scroll-stop potential, brand consistency and format fit. Weak assets don't just underperform; they waste learning budget and pollute the read for everything tested alongside them.
Every test gets a documented hypothesis, a pre-registered primary metric, a sample-size and minimum-detectable-effect calculation, and a read date. Tests that can't reach significance at the available budget shouldn't run, or should be redesigned as directional reads and labelled honestly as such. Keep a backlog of six or more prioritised hypotheses, reviewed quarterly.
If any field is blank, the test isn't designed yet. It's a guess with spend attached.
Calling winners early is a classification problem, so validate it rather than feel it. Check 6–12 months of history: did early signals (hook rate, thumbstop, CTR) actually predict the final winners? If the early signal doesn't classify past winners, it doesn't earn the right to kill future ones. Direct-response accounts should bias toward killing fast; brand work toward protecting potential heroes.
Measurement is a ladder, and confidence climbs it. Climb as high as budget and addressability allow, and respect the full attribution window rather than the dashboard on day 14. Never let the bottom rung set targets on its own.
Log the delta first, the interpretation second and the action third. Then cite the library in every new brief, so each quarter's brief is smarter than the last.
On a quick-service restaurant conversion campaign, the softer sell drove cheaper orders. It didn't become a footnote; it became a deliberately designed follow-up test.
When one insurance product line outperformed its sibling, the diagnosis traced to incentive-led messaging, so the same logic was briefed into the other line: transferring the territory, not cloning the asset.
Learning that lives in people's heads leaves with them. Without the library, the same mistakes repeat each quarter.
Two reads from James's own accounts, anonymised; staged the way learnings should be logged — delta first, interpretation second, action third.
Speed matters: these platforms reward fresh, plentiful input. But speed without direction is expensive noise. The job is velocity — high creative throughput, pointed by disciplined hypotheses and honest measurement.