Writing
What we have learned building and running media mix models, including the parts vendors tend not to discuss.
A brand went dark for five years, came back with a big campaign, and the model measured almost nothing. The null result was the most valuable thing it bought all year, and what the model found next is why.
A GRP is reach multiplied by frequency, which lets a plan hit its target while buying almost entirely one or the other. The evidence favors reach, and average frequency hides how badly delivery is usually skewed.
The largest return from brand advertising is not incremental volume. It is a flatter demand curve, and a media mix model can measure it if you ask for the right things.
Meridian and Robyn made the statistics free. Most of what a media mix modeling engagement bills for was never the statistics, which is worth understanding before you buy one or build one.
In progress: how to tell whether your data can support a causal claim, and why quarterly refresh cycles are a pricing artifact rather than a methodological one.