Google Ads AI

Ad account analysis is skilled, repetitive work: export performance data, eyeball which messages win, guess why, write variants, decide what to test next. Done well it takes an analyst a day per account, so it happens monthly at best — and the reasoning lives in one person's head.

The interesting question wasn't 'can AI write ad copy?' — it obviously can. It was: can AI do the analyst's job of explaining performance and choosing what to test next?

The pipeline treats analysis as a chain of narrow jobs. Structured campaign data flows in from the Google Ads API; the model receives normalized performance tables — not raw exports — and works through the same sequence a strong analyst would, one step per prompt.

The output isn't a chat answer. It's a ranked experiment queue: each recommendation carries the messaging theme it exploits, the evidence behind it, generated ad variants ready to launch, and an expected-impact score used for ranking.

The naive version — paste a CSV into a chat window and ask 'what should I do?' — produces confident garbage. The model has no baseline, invents causality, and averages away every interesting signal. What made this work was refusing to let the AI be a generalist.

Each stage does one analyst task with exactly the context that task needs. The performance-analysis prompt sees baselines and variance, not ad copy. The theme extraction prompt sees copy and outcomes, not budgets. By the time the model ranks experiments, it's synthesizing conclusions it derived stepwise — the same way a human analyst builds a recommendation they can defend.

The result reads like an analyst's memo, arrives in minutes instead of days, and runs weekly instead of monthly. The human's job moves up a level: approve, launch, and feed results back into the next cycle.