Test ideas, not pixels
A button colour or a slightly different crop rarely moves results enough to measure, and a platform will happily spend budget proving it. Build tests around hypotheses that would change what you make next: a problem-led opening against a product-led one, a price message against a quality message, a person on screen against product only, a demonstration against a testimonial. Write the hypothesis before production, including the result that would make you drop it. Then produce several executions of each idea, so you are testing the idea rather than one lucky or unlucky execution of it.
Budgets and sample sizes
The smaller the difference you want to detect, the more impressions and conversions you need; halving the detectable lift roughly quadruples the required sample. That arithmetic is why small tests on purchase rates so often end inconclusive. Fix in advance the minimum difference worth acting on, the metric and the test length, then run a sample-size calculation with your real baseline rates. If the budget cannot reach a clear answer on purchases, test on an earlier signal such as hold rate or click-through, and confirm the winner on purchases later. Stopping a test the moment one variant pulls ahead is the most common way to crown a false winner.
Working with platform automation
Ad platforms do not split budget evenly between creatives in the same ad set. Delivery shifts towards whatever looks like an early winner, often on thin data, so a creative can lose simply because it was starved of impressions. For a clean read, use the platform's split-testing tool, which divides the audience so each group sees one version, or run separate campaigns with matched budgets. Features that automatically mix headlines, images and formats are good for performance and poor for learning, because nobody can see which combination did the work. Use them to scale ideas that have already been proven, not to discover new ones.
A library of learnings
A test is wasted if its result lives in a slide nobody reopens. Keep a learning library in which every entry records the hypothesis, the variants, market, audience and dates, the result with its confidence, and what it means for the next brief. Tag the assets with the attributes that were tested, so later analysis can find patterns across campaigns rather than within one. Re-test important findings from time to time, because audiences tire of what once worked. The studio produces the next round of variants from one brief with the product locked and the brand rules applied, which keeps the cost of acting on a learning low.
Updated 25 September 2026 · General information, not legal advice. Rules change, so check the current text with your legal team before relying on it.