The first thing I heard about Marketing Mix Modeling (MMM) was Bayesian statistics - along with tons of exciting but slightly spooky formulas and theorems. Not because they are overly complex, but because I don’t have years in reserve to learn them from scratch.
While I’m still in the process of understanding prior and posterior distributions, I decided to rely on the PyMC documentation and explore MMM in practice.
So, what can MMM and PyMC actually do for you?
They help answer one of the biggest questions in marketing (if not humanity): what to do about attribution when every channel tries to claim the biggest piece of the pie?
Here’s the unexpected spoiler: you don’t need attribution models at all.
The only data you really need is:
-Your total backend revenue (or conversions)
-Marketing costs by digital and/or offline channel
You can also (actually, you'd better include) include “free” channels (such as email or CRM) and occasional events like Black Friday or other promotions.
What MMM can give you:
-Channel contribution. How much each channel contributes to total sales or revenue.
-Posterior ROAS. You don’t need platform-level attribution. With only channel costs and total revenue, MMM bypasses platform reports and directly estimates each channel’s return on ad spend.
-Adstock (carryover effect). Marketing impact doesn’t end immediately. Part of today’s spend influences future periods (e.g., brand recall). MMM applies a decay function so recent spend has a stronger effect, while older spend gradually fades.
-Saturation (diminishing returns). Each additional unit of spend becomes less effective at higher levels. Early spend drives strong gains, but eventually the response flattens. MMM models this using nonlinear functions such as logistic or Hill curves.
-Seasonality effects. Weekly, yearly, and holiday-driven demand patterns.
-Scenario forecasting. “What happens if we increase Meta spend by 20%?”
Maybe I’m overconfident, but I don’t think the PyMC documentation is too hard to try on your own - especially with the help of AI. My biggest challenge was fitting the model: it took 48 hours on my old PC. Still, I highly recommend doing it yourself. I’ve been working in digital marketing for over six years, and machine learning has been the best tool for solving many of the problems I face in my work.