Set the test budget from your CPA
Take your target cost per acquisition and multiply by three. That is the minimum spend before an ad has had a fair chance. At a £50 CPA, £150 per creative. Below that you are reading noise and killing ads that would have worked.
Cap the test at four days. Longer tests leak budget into losers and shorter tests judge an ad before delivery stabilises. Run 4 to 6 creatives per test cell so the algorithm has something real to choose between.
Keep a permanent testing cell at 10% to 20% of spend
Ring-fence a fixed share of monthly budget for new concepts. At £20,000 a month that is £2,000 to £4,000. It never gets raided to prop up a bad week, which is the mistake that leaves an account with one tired winner and no pipeline.
Run the cell as its own campaign with its own budget so scaling decisions elsewhere cannot starve it. Measure it on learning output, counted as validated concepts per month, alongside its own ROAS floor.
Test concepts first, variants second
A concept is a different angle, format or claim. A variant is a new hook, thumbnail or caption on the same idea. Concepts move performance in steps. Variants typically add 5% to 15% and extend the life of a winner.
Always test a concept against the incumbent winner. Running five new ads against each other with no control teaches you which of five is least bad, and you still will not know whether any of them beat what you already run.