A local services account, eight ideas in one launch: energy savings, health, financing, a free inspection, the buyback offer, a before-and-after, one AI image and one photo from a real job. Different angle every time, and because the files came from three people, different formats too. Two weeks and €2,000 later, the buyback ad has the best cost per lead.

Nobody in the room can say why.

I have launched that exact test. It feels like eight shots on goal. It produces one result you cannot explain, which means you cannot repeat it either.

Short answer: A Facebook creative test only teaches you something if one layer changes and the others hold still. Vary the angle first, because it moves results the most, then format, then the hook, then the edit. Your ad set's weekly conversions decide how many of those comparisons you can afford.

The takeaways

The angle carries the most variance. A different promise can halve or double cost per result. A different crop almost never does. Spend your first reads on the layer with the most room to move.
Every clean comparison costs a full read. Meta's "About learning limited" documentation puts a stable read at roughly 50 optimization events per week, so a strict one-variable-at-a-time plan across four layers is a multi-month project on an ordinary account.
Read the gap before the ranking. Five creatives finishing within a few percent of each other on 20 conversions apiece have produced an order and no finding.

Why did your last creative test teach you nothing?

Because the winner differed from everything else in more than one way, so its win has more than one explanation. In the launch above, the buyback ad was a new offer and a 9:16 video, while three of the losers were 1:1 statics from a different designer. Offer or format? The test has no way to answer.

This is rarely carelessness. It is the default state of a creative round. Ideas arrive as finished objects from a designer, an editor and a founder with opinions, and each one differs from the rest on every axis at once: promise, aspect ratio, pacing, the person on camera, the first line of text.

A creative brief and a test design are two different documents. Most accounts write only the first one, then read the results as though the second one existed.

What should you vary first?

Vary the layer you have the least evidence about, which on a new offer is almost always the angle. Angles are cheap to think of and expensive to get wrong: pick the wrong promise and no amount of editing rescues the ad. After that my order runs format, hook, edit.

Read three testing frameworks and you get two answers. One puts format at the top, because format and placement drive the biggest performance swings. Another says to throw dramatically different angles at the wall first, to find which direction is worth pursuing at all.

That disagreement is the useful part. No hierarchy was ever discovered in Meta's auction. The order depends on which layer your account has already settled.

Can you afford to test one variable at a time?

Often no, and nobody handing out a testing framework prices it for you. A comparison is readable only when each side produces enough conversions to separate from the other, and Meta's own learning-limited threshold sits at roughly 50 optimization events per week.

Run that arithmetic. An ad set doing 40 conversions a week funds about one honest comparison. Four layers with 2 options each, tested cleanly and in sequence, is 8 weeks of calendar time, by which point the offer has changed and the season has moved. That plan is fiction with a rigorous vocabulary.

So confound on purpose. Launch bundles: angle plus the format that suits it, three packages instead of eight ads. Find the bundle that wins by a distance, then unbundle that one alone. You are trading attribution for speed, which on a small account is the right trade, provided you say out loud which questions you left open. The count has its own ceiling, which I worked through in how many ads per ad set is too many.

Why is a cross-format comparison unfair?

Because most format tests compare a file composed for its placement against one cropped to fit. Take a 1:1 master, cut a 9:16 out of the middle, and the product sits off-centre while the headline collides with the caption. What you measured there was production effort.

Formats also start from different baselines. A 9:16 video and a 4:5 static do not convert at the same rate even when both are good, so a raw side-by-side ranking quietly punishes whichever format has the lower normal in your account.

That is the correction I built into Adscalr's scoring: through Bayesian shrinkage with format-specific priors, a new ad's wild early number gets pulled toward what its format usually does, so a lucky first week on an unusual format cannot take the top slot alone. Upstream of that, each of the four native formats gets composed for its placement instead of cropped out of one master.

What should you read first in the results?

The distance between the creatives, before their order. 5 ads landing between €38 and €44 cost per lead on 20 conversions each have told you nothing you would act on. The same 5 landing between €22 and €61 have told you something real, even on thin volume.

The number that decides whether you learned anything is the size of the gap, not the position of the rows. A ranking always exists. Somebody always finishes first, including in a test where every creative is identical.

Which is also why a wide test needs more evidence than a narrow one: the more creatives you rank at once, the better the odds that one leads on luck alone. I wrote up how to tell a winner from a lucky streak separately.

Turning this result into the next test

The useful output of a creative test is the next test. If the buyback angle won by a wide margin, the next round holds that promise still and moves the format. If everything finished in a cluster, the next round needs bigger differences, and more spend here will not buy them.

Adscalr's testing engine scores creatives on a blend of six metrics whose weights you set per project and funnel stage, then recommends the next test worth running, reproducible per calendar week. The ad-intelligence page covers how that scoring holds up on small samples.

None of which you need today. Pick the layer, hold the rest still, and count how many honest reads your weekly conversions will buy before you launch anything.

How to test Facebook ad creative