Marketing teams have access to more data than ever, but understanding what actually changed is still surprisingly manual.
A marketer might have to open Google Ads, Meta Ads, GA4, Search Console, and ecommerce dashboards just to answer a simple question:
Why did performance change yesterday?
That is the problem we are trying to solve with Marvyn.
Marvyn connects marketing platforms and uses AI agents to investigate performance across them.
Instead of only displaying metrics, the goal is to help answer questions like:
What changed?
Where did it change?
Is it actually unusual?
What could have caused it?
What should the marketer look at next?
The interesting technical problem
The difficult part isn't simply connecting APIs.
Each platform has a different data model, attribution logic, terminology and reporting structure.
A change in Google Ads might only make sense when compared with a change in GA4 conversions. A traffic drop might have nothing to do with advertising and instead be related to tracking or the website.
So we are working on building a context layer that lets agents investigate across multiple sources rather than treating every API independently.
Why we kept humans in the loop
We are also working on execution.
For example, Marvyn could detect a campaign issue and prepare a change, or detect a tracking problem and prepare the fix.
But we deliberately don't want the AI making those changes autonomously.
Advertising involves real money, and LLMs can still make mistakes.
Our current approach is:
AI investigates → AI prepares an action → human reviews → human approves
This gives the user the speed of an agent without giving it unrestricted control of an ad account.
What we're learning
One of the biggest lessons so far is that building an AI agent around business data is less about giving the model more information and more about giving it the right context at the right time.
We are still early and actively testing Marvyn with real marketing accounts.
If you have worked with Google Ads, Meta Ads, GA4 or agent-based systems, I'd especially love feedback on how you would approach cross-platform reasoning and safe execution.
Marvyn: https://www.marvyn.io/
Top comments (0)