Hey guys! Your AI cannot watch video. Ask Claude or ChatGPT about a TikTok and it answers from the title, the caption or a transcript. It never sees the picture. That is why every ad teardown it writes sounds like every other one. The obvious fix is to hand the model the frames. A minute of video is hundreds of pictures, and the model pays for every one. So we did the opposite: Playhead watches the video on our side and answers in words, with the second of every moment. The frames never touch your context. Our shot detection is new. Playhead names every cut and every transition in an ad, to the frame. Not "fast cuts". Cuts per second, the median shot length, and the second each claim lands. The part I did not expect was what made it actually useful for ads. The first version read one ad and wrote a teardown. It was accurate and it was worthless, because every ad opens on a face, cuts fast and shows the product by second three. So do the ads that lose money. A trait shared by every winner and every loser is a convention of the category, not a cause of anything. So a set now takes labels. Three winners and two ads you killed, and the question stops being "what do these have in common" and becomes "what do the winners have that the controls do not". One word in a prompt, and it turned a description into research. Every ad is still read blind to the others, so the comparison cannot talk itself into a pattern.

Playhead

Reverse engineer winning ADs at scale

Reverse engineer winning ADs at scale