Ori Feinberg All work

Reverse-Engineering a Competitor's Best Facebook Ad — With No Performance Data

How do you tell which of a competitor's Facebook ads is working best, when Facebook shows you none of the numbers — no spend, no views, no results?

You can't measure it, so I inferred it. The idea: a company's own behavior gives it away. When an advertiser runs the same style of ad again and again, they're telling you — with their budget — that it works. Nobody keeps paying to run an ad that flops.

So I built a tool that pulls a competitor's entire set of live Facebook ads, sorts them by type, and ranks each type on two things: how often the company runs it, and how long those ads stay up (advertisers pull losers quickly, so the ones still running are the proven ones). The "best" ad isn't the flashiest — it's the formula the company keeps betting on. Then the tool takes that winning formula and rebuilds it as a ready-to-shoot creative brief and storyboard for a different brand.

The output: a ready-to-shoot storyboard. The tool doesn't just name a winning ad — it produces a finished creative brief and a full shot-by-shot storyboard for the new brand, frames and all. Here's the real output from a real run:

Shot 01
Shot 02
Shot 03
Shot 04
Shot 05
Shot 06
Shot 07
Shot 08
Shot 09
Shot 10 · End card

Unofficial concept. Frames are generated by the pipeline; Guidde is used as the example target brand. Not affiliated with or endorsed by Guidde.

How it works

Five steps, each handing off to the next:

  1. Collect — scrape the competitor's live Facebook ads.
  2. Read — AI vision tags each ad's structure (format, hook, angle, call-to-action).
  3. De-duplicate — Facebook shows the same ad under many different IDs, so I removed exact copies by comparing the actual image and video files. 38% turned out to be duplicates — which would have quietly skewed the whole ranking if left in.
  4. Rank — score each ad type and pick the winner.
  5. Rebuild — generate the new brief and storyboard for the target brand.

What I learned

What actually makes the winning ad work: it turned out to be a video testimonial — a person talking to camera. My tool analyzes images, so it could see the ad but not hear it — and in a testimonial, what the person says is half of what makes it land. So I studied the winning ad by hand, and the pattern was clear: it's mostly story and barely product (the product only appears for about 10% of the video, right in the middle), and it ends on a feeling of status — "be the person who knows how things work" — not the usual "save time" promise. That structure is what I rebuilt for the new brand. Same formula, new ad.

Built with

  • Python
  • Claude (AI vision + copywriting)
  • Apify (ad scraping)
  • OpenAI (image generation)

Built end-to-end by directing Claude Code.

View the code on GitHub