Over twelve months we shipped 340 AI-generated ads for one ecommerce account. Roughly one in nine cleared the account's ROAS floor and stayed on. Here is what separated them.
The first two seconds carry everything
Hooks that named a specific irritation outperformed hooks that named the product, consistently and by a wide margin. "I replaced our product shoots with this" beat "Meet the new Helio lamp" every time we ran them against each other.
The winning openers were almost all first-person and mid-thought, as though the viewer had walked in on a conversation already happening.
Volume only helps when it is tagged
Sixty untagged variants a month is noise. The same sixty tagged by hook type, format and offer produce a readable table after four weeks, and the next batch of scripts writes itself from the winners.
This is the actual argument for AI creative: not that it is cheap, but that it makes a testing cadence possible at all.
Fatigue is faster than you think
A winning AI ad in this account burned out in about eighteen days at scale. Human UGC in the same account lasted a little longer, but took six weeks to produce. The maths favours the format you can refresh weekly.
Runs the hook library and the ad formats behind 340 shipped creatives for Helio Home, and designed the interfaces on the Nordwind and Veyra builds.