-
Yes, doing this with Claude Code against a 110+ account Google Ads portfolio (mostly real estate and local services). Read access goes wide with Google Ads API, GA4, pretty much anything and everything you can enable inside of a GCP and I let it pull whatever it needs to think.
Write access is where the boring engineering matters. Mutation Safety was the first skill I built when I added write access and it was modeled after how Google Ads Editor handles posting, where you stage changes locally and push them as a batch after review. Every mutation goes through a two-step approval pattern: the model proposes the change, dumps changes with the reason, and stages it. Nothing touches a live account until I look at it.
We do still use a good bit of bulk uploads and Google scripts as well as the stack has evolved over many years.
On CRM and offline conversions, I agree that Google Ads alone is not enough context. A keyword can convert at 3% in Ads and produce zero qualified leads in HubSpot. We’re pushing offline conversion integration as far as we can right now, currently testing implementation across a 70-account chunk. Better data leads to better outcomes is our motto.
Far from perfect though as most of the friction isn’t technical, it’s getting alignment between the client, their CRM, and their legal team on what data can actually flow back to Ads. But I think this is the future. The agencies that figure out the offline conversion + AI loop first are starting to look very different from the ones still doing monthly PDF reports.
How far am I letting it go overall: read everywhere, write nowhere without the staged approval. Even on negatives, where the cost of a bad add is low, the staging pattern is non-negotiable.
I open-sourced the Claude Code skills AI I’m comfortable sharing on GitHub if you want a look.
