Your Rival's Growth Playbook Is Hiding in Plain Sight, and Your AI Agent Can Dig It Out

A company that makes short-video ad creatives open-sourced a marketing research tool called REM. Give it a company's website or App Store link, and it gathers the acquisition traces that company has left in public. Your AI agent then turns them into an analysis report with numbered evidence. The project ships with a sample breakdown of the calorie app Cal AI: 894 ads in the Meta Ad Library, 692 of them still running. The longest-running ones have been live for 185 days straight. All of them are videos, and nearly all send people straight to the App Store. The useful part is that every judgment can be traced back to its source through its number, and anything that could not be found is listed honestly.

First, the Teardown

The table below summarizes the Cal AI sample report that comes with the project. The data is as of October 2026.

StageWhat was foundWhat it suggestsPaid adsMeta shows 692 ads running. All 61 ads in the report's sample are videos. The longest-running one has been live for 184 to 185 days, and each has 10+ variationsMeta video ads are the main engine. Creative that runs this long has been beating the alternativesLanding page60 of the 61 ads go straight to the App Store. The website only has basic analytics installedThe website is not where conversion happens. It happens on the store page and inside the appSelling pointThe ads, the website headline and the App Store description all say "snap a photo to log your calories"One selling point is repeated everywhereConversionA lifestyle questionnaire comes first, then a paywall. The entry point is a 3-day free trialA classic pattern: the questionnaire builds a plan, and the free trial turns into a subscriptionWord of mouthMore than 370,000 App Store ratings at 4.8 stars. The homepage highlights fitness creators who use the appScale and creator endorsements drive conversionRecent activityOnly one new ad launched in the last 30 daysThey may have shifted from expanding to harvesting

The report also includes a scorecard for 12 channels. LinkedIn ads, Reddit marketing and Hacker News launches are all rated as "none." Each rating cites the sources that were checked.

The figures above come from the sample report bundled with the project. We did not verify them, and we did not install or run the tool ourselves.

Why It Works: Three Reasons

• Evidence and judgment are kept apart: The tool only collects "where we saw what." Judgment is left to your AI agent, and every judgment must cite an evidence number. A final check means a report with citations that don't match is not considered finished.

• "Not found" is its own kind of result: Items that could not be found are not written up as "they don't do this." They are listed in a separate section with links for checking by hand. Readers can tell which conclusions are solid.

• Watch how long ads run: An ad that runs for months is making money. Studying those winners first is far more useful than scrolling through hundreds of ads.

Follow Along: A Five-Step SOP

These five steps work for both people and AI agents. The install steps and commands live on the skill page reverse-engineer-marketing. Hand that page to your AI agent, then have it check each step against its "done when" criteria as it finishes.

Safety Rules

• Public information only: Don't log into any of their systems, don't get past paywalls, and don't use this to investigate specific individuals.

• Learn the playbook, don't copy the creatives: Their ad videos, images and copy are protected. Shoot and write your own.

• Keep keys out of the chat: Set up data service keys yourself in the terminal.

• Follow each platform's terms and local law. If you're unsure whether a channel is allowed, skip it.

Common Pitfalls

• The one-line README install fails: When this was written, the package had not been published to npm yet, so install from source. The skill page has the steps. Ask your AI agent to check before installing.

• Drawing conclusions without data: Without keys, Meta, Google Ads and social data all show as "unknown." Saying "the rival doesn't run ads" at that point is wrong.

• Mistaking a tracking tag for a live campaign: A platform's tracking tag only shows the company installed it. Trust what the ad library shows.

• Judging by ad count alone: The count comes from the first few pages of a sample. Run length and how fast new ads launch are more useful signals.

• Trusting auto-matched accounts: Founder and TikTok accounts are matched by search and can be the wrong person. Open and verify the important ones yourself.

To see how a batch of ads can center on one selling point, read the case Keeping each batch of AI ads to one selling point.