The problem
Why describe a messy, non-linear business with simple, linear, deterministic models? That is traditional attribution. Last-click, even “data-driven”: toy models of the complex thing that is your business. Every channel takes credit for the sale, the credits add up to more than the revenue, and nobody can say what would have happened without the spend.
A general AI model does not fix this. Upload the deck and it reads it well. Ask why the number moved and it has no context. Connected beats clever.
How we measure
- 01Geo-lift
The gold standard. Turn it off here, leave it on there, compare. We pick the regions, hold some out, run the treatment in the rest, and read the difference with a confidence interval.
- 02Synthetic control
When a clean holdout is not possible, we build a control from the markets the spend never touched. Weaker than a holdout, and we say so.
- 03Causal reads on what already happened
Granger causality, causal impact, correlation with the lag tested. For questions like whether Meta spend moves branded search, and for how long.
- 04Marketing Mix Models
Or, as we call them, geo-lift interpolation models. They connect the dots the geo-lifts drew. Calibrated on experiments, not on last-click. Where a model would be guessing, it says so.
- 05Good attribution
Pixels still have a place, and with server-side integrations they are not bad at all. Good for day-to-day tactical calls. Plan the quarter on them? No. Compare today with yesterday? Sure.
Why the automation can be trusted
AI is good at diagnosis and weak at execution. Ask a model to analyze an account and it finds real things. Let it act and it changes things without knowing the business. So teams pick one: a human runs everything, or a model does. Neither is right.
Our answer is the middle. In EPYC, only the policy, what you want done written in plain words, comes from the model. The inputs are code, so it reads exact data and cannot invent it. The actions are code, so it can only do what your rails allow: never more than a share of a budget you choose, never negate the brand, always ask first. After enough approved suggestions, you can let it run. Until then, it asks.
What we refuse to claim
- Lift without a control. No holdout, no synthetic control, no lift. Just a change.
- A number without its interval. Every result carries its confidence. An interval that includes zero is reported as exactly that.
- Long-term effects we did not measure. A two-day halo is a two-day halo.