Steal the Winners: Building Your Own Competitor Ad Spy Tool With AI

Most people write ads from imagination. They open a blank page, ask an AI to invent a hook, run it, and then wonder why the leads are junk. That is exactly how this mastermind session started: with a hundred dollars of AI-written Facebook creative that produced nothing, sitting next to another member’s ads that were quietly converting. The fix was not a better prompt. It was evidence. So the session turned into a live spying exercise, and then into something more interesting: instead of paying a spy platform every month, we built our own competitor ad research tool from a single API key while everyone watched.

The Case Against Writing Ads From Scratch

Generating creative with AI feels productive. You get ten hooks in ten seconds. The problem is that none of them have ever met a real audience. Meanwhile, thousands of ads that have met a real audience, survived, and are still spending money are sitting in public, free to look at.

• A hundred dollars, no traction – AI-invented Facebook creative produced replies that never made it past the second message
• The comparison that stung – a member’s home-improvement ads, tested rather than imagined, were converting in the same week
• Obvious in hindsight – winning angles look intuitive once you see them, which is precisely why you need to see them
• Someone has always tested more than you – scaled spend is the market telling you what already survived

Rent the Scrapers, Don’t Build Them

The reason most marketers never touch this data is friction. The official channels are either gated or narrow. TikTok’s Commercial Content API requires an application through a developer account, and the dataset currently covers EU countries only.^{[2]Commercial Content APIdevelopers.tiktok.com} Google’s Transparency Center is browsable by hand, with filters for advertiser, date and location, but it is a website, not a firehose.^{[4]Announcing the launch of the new Ads Transparency Centerblog.google}

The session’s shortcut: API marketplaces where independent developers sell access to scrapers they already maintain and keep working.

• One catalog, every platform – short-form video, photo feeds, ad libraries, page data and search results, all behind a single key
• Cheap relative to what you get – a per-platform subscription in this space runs tens of dollars a month, not the hundreds a spy suite charges
• You skip the approval queue – a reseller hands you working access today instead of an application form and a waiting period
• Test before you commit – most listings include a playground, so you can confirm what a single request actually costs you
• Know what you are buying – some sellers are brokering access they were granted; others are running real scraping infrastructure

Building the Spy Tool Live

This was the part of the session that made the room go quiet. One credentials file, one paragraph describing the interface, and a working competitor-research tool was open in Chrome with playable video previews. No boilerplate, no framework decision, no afternoon lost to setup.

• The brief was plain English – describe the interface, the niche, the country, and what each result should show
• It built its own backend – a small local proxy so the API key stays on your machine and never reaches the browser
• Real results, not a mockup – live competitor ads came back with working previews, including foreign-language variants of the same angle
• Then you just keep asking – bulk download, sorting, ads-versus-organic filters, click a handle to pull that advertiser’s whole library
• The features you would never have specified – tell it to run ten improvement passes and walk away; some is filler, some is genuinely better than your spec

Modern agentic coding tools read a codebase, edit files and run commands themselves, and they now run from a terminal, an IDE, a desktop app or a browser.^{[5]Claude Code Overviewcode.claude.com} That last part matters more than it sounds. The same docs describe moving a running session to your phone, so you can approve the next build step from the couch.^{[5]Claude Code Overviewcode.claude.com} We spent a few minutes on that in the session, because it quietly removes the last excuse for not building your own tools.

Two Models Are Better Than One

A habit that came up mid-build and is worth stealing for every project you touch: after your primary AI writes the code, have a second, different coding model review it before you run it.

• The instruction is one line – ask your assistant to verify its own work with the other model
• It finds real problems – typically a handful of correctness, security or performance issues per pass
• Adversarial by design – two models trained differently look at the same problem and notice different things
• Measurably better first-try success – roughly a fifth more code that works on the first run, across a long project
• Cheaper than reaching for a bigger model – two mid-tier passes beat one expensive one on cost per working feature

Analyze the Footage, Not Just the Script

Downloading a competitor’s ads is the easy half. Understanding why a particular video works is the part that actually changes your results, and it takes two separate jobs.

• Transcribe everything first – speech-to-text APIs return word-level start and end timestamps, so you can pattern-match scripts across dozens of winners^{[6]Pre-Recorded Audio Transcriptiondevelopers.deepgram.com}
• Then analyze the visuals separately – cuts, transitions and pacing never show up in a transcript
• Go second by second – break the video into per-second frames and pair them with the transcript so the model sees construction, not just content
• What you are looking for – how fast they cut, and what the picture is doing at the exact moment the hook lands
• Give the AI real access – it needs to download the file locally and run the analysis itself, not guess from a description

Turning a Competitor’s Winner Into Your Own Ad

Once the tool is full of proven creative, the workflow is refreshingly unglamorous.

• Rip the concept, not the file – a human made that video, so take the structure and the hook, not the footage
• Freelance editors are the cheap path – send reference clips, point at the parts that matter, pay a small flat fee for a rebuild
• Cross-platform arbitrage – an angle proven on one network can often be run on another where your category has more room; ad policies differ by platform and market
• Watch the calendar – most verticals have a season where everybody piles back in, and the library shows you when
• Clean usually beats gimmick – incentive-heavy angles win volume, straight offers win lead quality
• The tool itself is an asset – bolt on a few more data sources and the interface becomes something you can sell

The Real Takeaway

The gap between “I have an idea for a tool” and “the tool is open in my browser” is now minutes and one API key. That changes what is worth building. Instead of renting a spy platform whose feature set someone else chose, you describe the exact research workflow you want, let one AI build it, let a second one check it, and keep bolting on features until it does something no off-the-shelf product does. Your competitors’ ads were always public. What changed is that turning them into your own intelligence system stopped being a project and became a conversation.