Ad Creative Testing With an Incrementality Backstop

What the platform will not tell you: Meta will name a winning creative. It is comparing creatives it chose to deliver, to audiences it chose to deliver them to, optimised toward an outcome it predicts, which means the reported gap between two ads is partly a measurement of the auction rather than of the creative. This framework separates the two: in-platform tests structured so budget, audience, placement and optimisation event are identical across cells, and a periodic geo holdout that prices what share of the reported conversions would have arrived anyway. Teams running the second for the first time usually find that share lands somewhere between 40 and 70 percent, and the entire creative programme has to be re-read in that light. Test structure: One ad set per cell, dynamic creative off, and a spend floor set so each cell reaches at least 50 optimisation events before anything is read. Below that you are reading delivery noise through a confident interface. Cost cap rather than lowest cost when the comparison matters more than the volume. Learning phase exits respected rather than reset by mid-flight edits, which is the most common way a creative test inside a busy account quietly destroys itself. The metric ladder: Creative effects appear early in the funnel and dilute later, so the ladder is read in order: thumbstop rate at three seconds, hold rate to fifteen, outbound click rate, then cost per acquisition, then cost per incremental acquisition from the holdout. A creative that wins on hold rate and loses on cost per acquisition is an offer or targeting problem wearing a creative costume. The framework exists to make that visible, not to hand you a winner. Taxonomy: Concept, hook, format, iteration. Four fields, enforced in the naming convention and exported to your warehouse. Without them you cannot answer the only question that compounds: which concepts keep producing winners after their first iteration. Most accounts I open have a naming scheme that describes the file rather than the idea, which makes eighteen months of creative learning unrecoverable no matter how much was spent producing it. Incrementality: Geo holdouts using GeoLift or a matched-market design, typically 10 to 15 percent of markets held for three to four weeks. In-platform conversion lift studies where spend supports them, with their limitations stated in the readout rather than in a footnote. On marketing mix models: below roughly three million euros of annual media a model will not identify creative-level effects, and I will not sell you one that pretends otherwise. Cadence: A new tested concept every two weeks, iterations only on survivors, and a written kill rule agreed in advance so a losing creative is not defended into a second month by whoever briefed it. The kill rule is the part clients push back on and the part that produces the gain. Not included: No creative production. I do not write scripts, direct shoots, edit video or design statics. I decide what gets tested, against what, and what the result means. No media buying and no account management. No attribution model rebuild, no multi-touch implementation, no pixel forensics beyond confirming the events used in a test are trustworthy. Who this is not for: Accounts spending under about 50,000 euros a month, where cells will not reach 50 events inside a readable window and the sensible advice is to buy the obvious and iterate on instinct. Also not for teams who need the conclusion to be that their current agency's creative is working. The holdout does not negotiate, and I present its number unedited to whoever is in the room.