What to hold constant when you generate variants

As generating variants got easier, a new problem appeared: too many things change at once. New image, new headline, new opening line — and a new audience too. When a variant wins, all you are left with is a single sentence, “that one worked”, and that sentence helps not at all when you make the next creative.

A test is a question, not a contest

The point of generating variants is not to find the best ad but to learn which element makes the difference. Those sound alike and lead to opposite work: the first pushes you to produce as many variants as possible, the second to change as little as possible.

The practical rule: one element changes per round. If the image changes, the copy stays. If the headline changes, both image and audience stay. Once a winner is clear, fix it and move to the next element.

• Angle: the promise the ad makes. It creates the largest difference, so it is what you test first.
• The image, or the first frame of the video: this is what does the attention work.
• Headline and opening line: the part that decides the click.
• Format and aspect ratio: varies by placement and should not be confused with creative quality.
• Call to action: usually the smallest effect, so it goes last.

Generating with AI raises the stakes

Asking a model for twenty variants takes five minutes, and all twenty differ from each other in more than one way. The result is twenty creatives and zero knowledge. The generation step needs a constraint — “change only the headline, keep the image and the angle identical”.

Write down what you learned

The output of a test is not the winning creative, it is a sentence: “with this audience, price framing beats benefit framing”. Once those sentences accumulate, making new creative stops being guesswork. When they are not written down, the same test gets run again six months later — usually with the same result.