AI ad copy: start from scratch or feed it your winners first?

Most people ask an AI to write ad copy from a blank page and then wonder why everything sounds like a brochure. A better starting point is to paste in your five best-performing headlines and descriptions, plus the search terms they matched and a sentence on why you think they worked. The model then builds on a pattern that already got clicks instead of guessing what a clinic owner or a SaaS founder wants to read.

The catch is that you get variations on one angle. If your old winners were all about price, the new batch will be too. So it makes sense to run two prompts side by side, one with the winners and one blank, and only trust the blank-page drafts if they hold up in the test. Has anyone seen blank-page drafts actually beat the winners?

Calling the winners the ones "that got clicks" is shaky, Nova. Clicks are the easiest thing to win, and a headline promising a free trial or the lowest price can pull in plenty of curious people who never convert. Feed that set to the model and it'll happily write more lines that bring back the same tire-kickers. So before you paste anything, sort your old ads by cost per conversion or lead quality instead of CTR. Five headlines that produced cheaper signups make a much better brief than five that just got opened.

I'd keep the blank-page prompt too, but judge both batches on the same downstream number. A blank draft that beats your winners on CTR alone doesn't tell you much.

Agreed on the metric, but on Meta a winning ad often wins because of who saw it, not because of the words. The same headline can get a great cost per lead on a broad audience and a poor one on a narrow interest stack. If you paste those five winners in without context, the model learns the audience, not the copy. So tag each winner with its targeting, and give conversions a few weeks to settle first, since recent ads often look worse while attribution is still filling in.

It depends on conversion volume, so check that first. If an ad has 8 conversions, its cost per conversion is mostly noise, and you'd be teaching the model luck. Two good weeks on a small budget can look like a winner for the same reason. Once an ad has a few dozen conversions, the ranking starts to mean something.

When volume is thin, don't pick the single best ad. Look across the ads that performed reasonably well and find what they share, like a specific price point, a particular objection handled in the first line, or a call to action that shows up everywhere. Those repeating patterns are what the model can actually use. The one standout headline is usually the one you'd be wrong to copy, because you can't tell whether the words earned it or the audience did.

The repeating-pattern idea is the part I'd lean on hardest. When you pull those shared elements out, write them down as a short list of constraints (price point, first-line objection, CTA) and paste that list into the prompt along with the winners. Otherwise the model tends to copy the exact sentences and just shuffle the wording.