In one line
A marketing mix model is a regression of weekly sales on media, price, deals and outside factors such as season. A Bayesian one says what each euro of media returned with a credible range, not just a single number.
What it does
- Carry-over (adstock): an ad still sells in the weeks after it runs, fading at its own rate. Saturation: each extra euro in a channel buys less, so the response curve flattens. Modern models such as Google's Meridian, PyMC-Marketing and Meta's Robyn estimate both for every channel.
- The model learns from how sales moved when spend moved. If spend barely varied, or two channels always ran together, it cannot tell them apart. If spend rose when demand was high anyway, plain least squares hands the season's sales to media.
- Priors carry what earlier studies found: hundreds of estimates put the short-term advertising elasticity at about 0.12 (Sethuraman et al. 2011). Pooling across the brands or regions of a category lets a brand with thin data borrow from the rest. With little data the prior steers; with a lot, your data wins.
- A lift test switches a channel off in some regions and compares. Fed into the model as a measurement it must agree with, it narrows the range more than another year of observational data.
Business and brand
Impact on the businessP&L, cash, share price
- Read the 80% range before the point estimate: if it crosses break-even, the model can't yet say the channel pays back.
- Budget decisions need the next euro's return from the response curve, not the average return.
- Click-based attribution credits the last ad clicked and misses TV and posters; a mix model sees every channel, but only as precisely as your spend varied.
Impact on the brandMemory, loyalty, pricing power
- A weekly model mostly sees the short run. Published long-run advertising elasticities are about 2 times the short-run ones, so a model judged on the quarter alone leans to activation.
- Binet & Field's 60:40 brand to activation split is an average across IPA cases. Your own best split depends on your margin, how saturated each channel is and how long you are judged over.
- A mix model sees euros, not ads: a stronger campaign looks like the channel working harder unless creative quality is measured and put in the model.
Try it
Fit a model
We simulate two or three years of weekly sales with a known true return, then fit a model. Change how much spend varied, whether it was timed to the season, whether the model uses a prior, and whether you ran a lift test.
Estimate
Estimate per €1€1.82
95% range−€0.86 to €4.49
Notice: Wide: the true return is inside, but so is almost anything. Too vague to move budget on.
Runs Countermove's market engine: the reference soft-drinks market, three brands, rivals holding their plans so only your change moves the result.
Sources
- Jin, Wang, Sun, Chan & Koehler (2017), Google: Bayesian methods for media mix modelling with carryover and shape effects
- Chan & Perry (2017), Google: challenges and opportunities in media mix modelling
- Sethuraman, Tellis & Briesch (2011), Journal of Marketing Research: advertising elasticity meta-analysis
- Assmus, Farley & Lehmann (1984), Journal of Marketing Research: how advertising affects sales
- Google Meridian documentation (2025): ROI priors and calibration with experiments
- PyMC-Marketing documentation (2025): lift test calibration
- Binet & Field (2013), The Long and the Short of It, IPA