Problem
Thumbtack buys Google search ads as responsive search ads (RSAs). Each RSA holds up to 15 headlines (30 characters each) and 4 descriptions (90 characters each). The team wrote one template by hand, so 80% of the assets were identical across ad groups, and 11 of 15 headlines were the same everywhere. Ads were generic, repeated phrases inside an ad group, and poorly rated ads were never rewritten because hand-writing keyword-based copy for hundreds of ad groups was not feasible.
Example from the article: a search for "plumber to install pipe" could show the headline "Hire a reliable plumber", with no mention of the specific service.
Solution in brief
An LLM writes assets for each ad group from its name and keywords. A second LLM acts as a judge and keeps only assets that meet the requirements. A third step splits the survivors into distinct ads, and the SEM team reviews the result. Synthetic control measured the pilot, and Google Drafts and Experiments measured the full rollout.
Questions this case answers
- How to generate personalized text variants at scale under strict length limits while keeping human review affordable
- How to estimate the effect of a change applied to a whole account when per-unit metrics are too volatile for a quasi-experiment
- How to generate keyword-specific search ad copy for hundreds of ad groups without writing it by hand?
- How to filter LLM-generated ad assets automatically before a human reviews them?
Tools and methods
- LLM generation — writes headlines and descriptions per ad group
- LLM as a judge — keeps assets that meet the requirements
- Separation step — splits assets into distinct ads
- Validation checks — length and content checks on every LLM output
- Synthetic control — effect of the pilot
- Google Ads Drafts and Experiments (D&E) — LLM ads against manual ads, whole account
Google or a one-time email code. No password.