Marketing ops · Production

Thumbtack automates Google RSA ad copy generation with LLMs for personalized SEM campaigns

The problem

Thumbtack's SEM ad copy was produced from a single template, leaving over 80% of assets identical across all ad groups with no keyword-level personalization—making tailored ads for hundreds of ad groups nearly impossible at scale.

First attempt

The templated approach produced identical, repetitive ad copy—11 of 15 headlines were the same across all ad groups, device-specific variants were absent, and poorly performing ads were never updated due to the manual effort required.

Workflow diagram · grounded in source
1
Ad group inputs assembled
trigger
“The LLM inputs for generation include the ad group name, keywords, ad copy best practices, and character limits”
2
LLM generates ad assets
ai_action
“Using LLMs, ad assets are automatically generated based on best practices and character count restrictions. This methodology focuses on creating highly personalized ads for all ad groups, at scale, by including top-performing keywords an…”
3
LLM judges and selects assets
ai_action
“LLMs are leveraged not only for generation but also "as a judge" to choose preferred assets from the generated options. This capability significantly reduces the manual review time by filtering out poorly drafted assets and only outputti…”
4
Validation and parsing
validation
“involves parsing the initial generation output to filter out irrelevant and undesirable content, or assets that do not adhere to specific character limits or best practices”
5
Assets separated into distinct ads
ai_action
“a critical aspect of the proposed solution is the ability to separate the selected assets into distinct ads, when multiple ads are needed for each ad group. The goal is to replace existing manual ads with new AI-generated ads, aiming for…”
6
SEM team human review
human_review
“The SEM team initially reviews the LLM output and selects the best ad copy from the filtered options. Although the aim is to eventually enable the LLM to write all ad copy with minimal manual review, human oversight remains crucial, part…”
7
Ad assets uploaded to Google
output
“the final selection of assets for upload to Google still involves a manual review process”
Reported outcome

The LLM approach saved several months of review time and enabled personalization at scale across all ad groups (previously impossible); the PoC produced a 20% lift in traffic and 10% lift in conversions, and phase 4 results showed statistically significant improvement in impressions, CTR, and conversions.

Reported metrics
Ad assets tailored to ad groups (baseline)less than 20%
Generic assets identical across all ad groups (baseline)80%
Identical headlines across all ad groups (baseline)11 of the 15 headlines
traffic lift (PoC phase 1)20%
Show all 9 reported metrics
ad assets tailored to ad groups (baseline)less than 20%
generic assets identical across all ad groups (baseline)80%
identical headlines across all ad groups (baseline)11 of the 15 headlines
traffic lift (PoC phase 1)20%
conversion lift (PoC phase 1)10%
review time savedsaved several months for review
CTR (phase 4)higher CTR
conversion value per request (phase 4)higher conversion value per request
impressions, CTR, conversions improvement (phase 4 overall)statistically significant improvement
Reported stack
Large Language ModelsGoogle Responsive Search Ads
Source
https://medium.com/thumbtack-engineering/automating-ad-generations-with-llms-8ac887f02e0a
Read source ↗

Frequently asked questions

What did this team achieve with this AI workflow?

The LLM approach saved several months of review time and enabled personalization at scale across all ad groups (previously impossible); the PoC produced a 20% lift in traffic and 10% lift in conversions, and phase 4 r…

What tools did this team use?

Large Language Models, Google Responsive Search Ads.

What results were reported?

Ad assets tailored to ad groups (baseline): less than 20%; Generic assets identical across all ad groups (baseline): 80%; Identical headlines across all ad groups (baseline): 11 of the 15 headlines; traffic lift (PoC phase 1): 20% (source-reported, not independently verified).

What failed first in this deployment?

The templated approach produced identical, repetitive ad copy—11 of 15 headlines were the same across all ad groups, device-specific variants were absent, and poorly performing ads were never updated due to the manual…

How is this marketing ops AI workflow structured?

Ad group inputs assembled → LLM generates ad assets → LLM judges and selects assets → Validation and parsing → Assets separated into distinct ads → SEM team human review → Ad assets uploaded to Google.

Related marketing ops cases