AI ad creative testing: how many ad versions do you need?
How to test ad creative with AI: what to vary first, how many versions to make, how long to run them, and how to read results without fooling yourself.
Short answer: test big differences before small ones. Start with a handful of truly different ideas (different hooks, angles or formats), not one ad with ten colour changes. Give each enough budget and time to get real conversions, keep the winners, then make new versions of the winners. AI makes this cheap, because each new version takes hours instead of a new shoot.
Why creative testing matters more now
Ad platforms like Meta and Google now choose much of the audience automatically. The ad itself decides who stops and who scrolls past. So testing creative is how you "target" today. See Meta Advantage+ campaigns.
What to test first
In order of impact:
- The angle: which problem or desire you lead with. "Save time" and "look professional" can sell the same product to different people.
- The hook: the first two seconds of a video, or the headline on an image.
- The format: short video, UGC-style video, static image, carousel.
- The offer: free trial, discount, demo, guide.
- Details: colours, music, captions, call to action. Test these last.
How many versions?
Think in rounds, not a fixed number:
- Round 1: a few clearly different angles, each with one or two hooks. Enough to find a direction.
- Round 2: take the winning angle and make several new hooks and formats around it.
- Round 3: refine the winner: lengths, sizes, languages, small details.
The right number depends on budget: each version needs enough spend to collect real conversions. With a small budget, test fewer, bigger differences. With a larger one, you can test more at once.
How AI helps
- Scripts: draft many hooks and angles quickly, then pick the best by hand.
- Production: generate scenes, voices and product shots for each version. See how AI video ads are made.
- Resizing and cutting: turn one ad into vertical, square, horizontal, short and long versions.
- Languages: voice and caption the winner for other markets.
People still choose what to test and judge what's on-brand.
How to read results honestly
- Judge on the goal: cost per lead or sale, not likes or clicks.
- Wait for enough data. A few conversions can be luck.
- Change one big thing at a time between rounds, so you know what made the difference.
- Watch for fatigue. A winner slowly gets worse as people see it again; that's the time for new versions.
- Write down what you learn, like "problem-first hooks beat feature-first", so each round starts smarter.
Common mistakes
- Testing tiny changes before big ones.
- Spreading a small budget over too many ads.
- Picking winners on click-through rate alone.
- Letting AI publish ads nobody checked.
More on mixing AI and real people in AI ads vs UGC creators.