Our Google Ads brain,
as code.
The AI Experiment Engine runs on Claude Code skills — reusable, auditable workflows we built for our own accounts first. This is the library, in the open: what each one does, exactly as we run it. The public repo ships this autumn.
Full audit of one account: campaigns → weak ad groups → wasteful search terms → PMax asset groups → best and worst products. Ends in steering actions, not observations.
Every search term of a campaign categorised by intent, a compressed negative list (phrase patterns, collision-checked), and the structural causes flagged — TOFU-aware, so exploration doesn't get killed.
Measures how much brand traffic is hiding inside PMax, seals it with campaign-level negatives, and feeds the missed brand variants back into your branded campaign.
A complete branded defense campaign: target impression share bidding, exact + phrase + misspellings, and a 15/4 RSA aimed at Excellent ad strength. Ships PAUSED — you flip the switch.
An asset-less Performance Max build: feed-only, text generation off, URL expansion off. Optionally split by product type on rising search demand.
Cross-references the products getting spend today with real Keyword Planner demand: is your budget on rising SKUs or dying ones?
Last year's Q4 from the account (or Keyword Planner demand if there is no history) → a dated BFCM / holiday campaign calendar with owner per task.
Upcoming seasonal and pop-culture demand that fits a store's niche — validated with real search volumes for five markets before anything gets built.
The Allowance, properly: contribution margin → break-even CPA → payback-derived target CPA with sensitivity bands. Script-computed, so client numbers are auditable.
Why give this away? Because the skills are the engine room, not the edge. The edge is the judgment layer on top — the pre-registered criteria, the payback windows, the senior who signs every euro. You can run the tools; the standard is the product.