Marketing mix models used to be an enterprise project. Eighteen-month engagement, six-figure budget, three consultants in a room. The output was usually a slide nobody trusted.
That world is over. With modern tooling, a small in-house team can stand up a working MMM in three to four weeks. Here is the exact sequence I have used twice in the last year.
Week one — clean the inputs
Every model is downstream of its data. Most MMMs fail not at the modelling step but at the data step. Spend the first week getting paid media, organic, email, and offline spend into a single weekly cadence with a consistent currency and a stable definition of revenue.
A good MMM is mostly bookkeeping with a regression on top.
— Engagement notes, 2025
Week two — the model
I use Robyn or Meridian, depending on the team’s appetite for Python versus a hosted UI. Both will give you a working model with reasonable adstocks and saturation curves out of the box. The point of week two is not to optimise the model; it is to see the model.
Week three — calibrate
Every model has prior beliefs about what is roughly true. Use a small geo-holdout test to calibrate the most contested channel — usually display or programmatic — and feed the result back as a prior. This is the step that separates an MMM from a horoscope.
Week four is presentation: a one-page executive view, a two-page channel view, a reproducible notebook. Then keep running it monthly. The model will get smarter the longer you live with it.