AI-drafted RSA headlines: let the assistant write them or mine them from search terms?

Pasting a whole catalog into an AI assistant and asking for fifteen RSA headlines looks productive, but I doubt it beats an hour with the search terms report. The model tends to spit out plausible, generic lines that could belong to any company in the category. Then Google rotates them, and the reporting won't tell you which headline actually earned the click, so you learn very little.

I'd give the assistant a narrower job instead. Pull the phrasing customers really use from search terms, reviews and support emails, then have it draft headlines from that language. Keep pinning off until you have enough clicks to judge. Where do you draw the line between AI drafts and copy you write yourself?

Nova, you can't see click-level attribution per headline, true, but "the reporting won't tell you which headline earned the click" goes a bit too far. The asset report still gives you impressions and a performance label for each headline, and that label is built from conversions. So the real question is whether your conversion tracking is clean enough for the label to mean anything. If you're counting form fills that include junk leads, or a conversion fires on page load, the labels end up ranking headlines on noise. Fix the conversion setup first and the asset report becomes a usable signal.

I agree with mining search terms and reviews for language, but I'd keep the AI draft step in the middle. Search terms show what people typed, which isn't the same as what makes them click an ad. A headline in the customer's own words can still lose to a plainer one, so let the data decide the ranking instead of assuming the customer's wording wins.

Jun's right that the labels come from conversions, but they're relative, not absolute. Google rates each headline against the others competing for the same position, so a "Low" can be a perfectly decent line that just lost to a stronger one. Pinning muddies it further, since a pinned headline gets fixed exposure and the others in that slot are judged on less data. I'd treat the labels as a ranking inside one test, not a verdict on the copy, and check the conversion counts behind them before cutting anything.

It comes down to how many conversions each campaign gets. If an ad group only sees a handful a month, the asset labels are mostly noise, whether the copy came from search terms or from a model. In that case I'd write the headlines yourself and use the data just to kill the weakest ones. Once you have enough volume for the labels to settle, let the AI draft from the search-term language and let the winners decide.

Atlas, agreed on volume. When you hand search terms to the assistant, put the conversion count next to each term instead of sending a bare list. Without it the model treats every phrase as equal and writes headlines around the long, chatty queries that never buy. Filter down to the terms that actually convert first, then ask for drafts from that shortlist.