In my ErgoSleeve creative test DCT 38, the opening I marked in the brief as my primary bet was shown 49 times. The third opening, in the same ad set, was shown 70,849 times and, with its copies in other ad sets, went on to spend €54,016 at a 2.08 ROAS in Triple Whale. A creative testing framework has to turn a result like that into the next brief without claiming more than the data shows.

Mine is a log. Before launch I write down the question, what changes between the ads, what stays the same, and what I will do if it works. After launch I record what happened and what the result cannot tell me. Below is the full entry for DCT 38: the brief, the delivery data, the results, the limits and the decisions. The same log, with the DCT 38 entry filled in, sits near the end so you can copy it.

Start with the question, not the ad count

A test answers one question. "Which of these twelve ads is best?" is a question about a batch. "Does selling who she wants to be work better than selling what the sleeve does?" is a question about the argument, and its answer tells you what to brief next.

I split tests into two kinds. A concept test changes what the buyer is asked to believe. An execution test keeps the belief and changes the opening, the footage, the format or the voice. Both are useful, but they answer different questions, so they get separate lines in the log. How many of each a budget can carry is its own topic, covered in how many ads to test in a Meta CBO. This article picks up after launch.

The brief: one concept, three openings

My hypothesis for DCT 38 was about the argument. The account had been selling the feature stack: sleeve, stand and wrist rest in one product. A feature stack reads as a list. I wanted to sell an identity instead, a woman who cares less about her posture than about being the person with the beautiful setup at the café.

The format came from a personal-recommendation ad for a tote bag: one woman telling a friend why she loves the thing she carries. Three openings lead into one shared body. The full brief and the finished ad are on my work page.

What happened, with dates and limits

The test ran as one ad set containing the three ads, inside the brand's ABO testing campaign, on a budget of about €100 a day. The ad set is dated August 15, 2026.

On August 17, the third opening was copied into the account's scaling and cost-cap ad sets for the US, Canada, the UK, Australia, Singapore, Germany and Austria, and the Benelux countries with Denmark.

The brand judges purchases and ROAS in Triple Whale, using its Triple Attribution model, and uses Ads Manager only for delivery and engagement numbers. On that basis, from July 1 to September 1, 2026, DCT 38 spent €54,021 across all its ads at a 2.08 ROAS, with 1,218 purchases at €44 each. The third opening accounted for €54,016 of that spend. Over the same window, all of the account's Meta spend returned 1.90 at €46.90 per purchase on the same model, and DCT 38 made up about 21% of that spend.

Read those numbers for what they are. They are attributed results from a click-based model. They are not an incrementality test, so they cannot say how many of those sales would have happened without the ad. The last days of the window also ran while the account was cutting spend because stock was running low. What they do show: one ad took the budget and held a return above the account average while it did. Whether the extra budget behind a winner pays for itself is a separate calculation, covered in the marginal ROAS calculator.

Why uneven delivery is not a randomized experiment

Inside the test ad set, Meta showed the third opening about 70,000 times and each of the other two fewer than 70 times. Most of that split was already in place by the end of the first reporting week.

I act on that allocation. Openings 1 and 2 are finished. I do not move them into their own ad set with forced budget to give them a second hearing. When delivery piles onto one ad at a return the business can live with, Meta has picked the ad to build on, and the useful response is new creative, in the same campaign, built from the ad it picked.

What I do not write in the log is "opening 1 converts worse than opening 3." Nothing here supports that sentence. A randomized test splits comparable people between versions and gives each version comparable delivery before comparing results. Here, Meta decided who saw what from its own predictions and stopped showing two of the ads almost immediately. Forty-nine impressions cannot tell you how an ad converts.

So the log line reads: opening 3 took the ad set's delivery and made money at scale; openings 1 and 2 were not chosen, and how they would convert is unknown. The next brief does not depend on that unknown, so I leave it unknown.

When a decision does depend on it, use a setup built for comparison. Meta's creative testing option in Ads Manager splits a set test budget evenly across a small number of ads for a fixed number of days, instead of letting delivery favor one. That buys evidence on ads Meta would not have chosen, which costs money. It is worth paying for when the losing version matters, for example before rebuilding a landing page around one angle.

The iteration that looked decisive

After DCT 38 won, the next brief was an iteration, DCT 38.2. It kept the voiceover and cut one beat, the line where she recommends the sleeve to her friends because it replaces her case, stand and wrist rest. It also opened differently. DCT 38 opens on a woman's face in a café with the red sleeve already set up; DCT 38.2 opened on hands placing a sleeve on a desk.

On the same Triple Whale basis, DCT 38.2 spent €738 at a 1.37 ROAS with 12 purchases. Its hook rate, from Ads Manager data, was 21%, against 46% for DCT 38.

That comparison became a rule in my next briefs, and it is also the weakest evidence in the log. DCT 38.2 changed two things at once, the opening shot and the missing beat, so it cannot separate them. Its ad set is dated August 29, which gave it a few days and 12 purchases before the window closed. About half of its spend landed in the first days of September, while the account was cutting spend for stock and DCT 38 itself fell to 1.68 in Ads Manager's weekly numbers.

What it earned is a working rule, not a proven one: when you iterate on a winner, keep its first frame and change the footage inside the beats. The next entry tests exactly that, the DCT 38.2 rebuild with the original first frame and new footage. It has not run yet, because the brand paused new briefs in September while it restocks.

The creative test log

Use one row per brief. Write the hypothesis, the variation, the budget and your planned next decision before launch. Fill in the rest, and revise the next decision, at your weekly review.

Running the log every week

Before you read results, check that each ad was eligible to deliver. A rejected ad, a paused ad set or a spending limit is a setup fact, not a creative result, and why a Meta CBO ad is not spending walks through the checks.

Set the loss limit before launch and apply it when the window closes. When to pause Meta ads covers the decision criteria.

Put at least one new argument into every batch. Delivery can only rank the ads you enter. An argument you have not tested has no data at all, and another hook on the current winner will not produce any.

Then write next week's briefs from the "Next decision" column. That is the point of keeping the log.

This is the weekly work in my creative strategy for DTC brands: research, briefs, the test log and the next batch built from it. If you want it run on your account, Book a call.

Meta's references: About creative testing and About A/B testing.