Marketing mix modeling without a data science team
Robyn and Meridian have made MMM realistic for £5M+ marketing budgets.
MMM used to mean a £150k consulting engagement. Open-source frameworks have collapsed the cost. A mid-market brand with two years of clean weekly data can now run a useful first model in a week and update it quarterly without external help.
The minimum dataset
Weekly spend and revenue by channel for 104 weeks. Add weather and seasonality. Skip macro data unless you're in travel or property. The most common failure mode is trying to model with 26 weeks — the seasonality just isn't there to learn from.
Tooling reality
- Meta's Robyn — most mature open-source option, ggplot-heavy output.
- Google's Meridian — newer, cleaner API, Bayesian.
- PyMC-Marketing — for teams already in the PyMC ecosystem.
- Avoid SaaS MMM tools under £50k/year — usually black boxes you can't validate.
What to do with results
Use MMM to set channel ceilings, not daily bids. Re-run quarterly. Pair with attribution for tactical decisions. The biggest mistake is treating MMM output as ground truth — it's a directional planning tool, not a measurement system.
Common confidence trap
Channel coefficients with wide confidence intervals are still useful for ordering but not for sizing. Don't reallocate 30% of budget on a coefficient with a 200% confidence interval.
