Channel Mix Modeling Calibrated on Geo Holdouts

We will tell you inside the first week whether your data supports a media mix model, and about half the time the honest answer is no. The bar: at least one hundred and four weeks of weekly spend split by channel, real variation in that spend across the period, and few enough channels that the model has observations to spend on each. A channel that ran at a flat monthly budget for two years contributes nothing to the model, however much it cost you. When the data does not clear the bar we say so and design geo experiments instead, which answer a narrower question, faster, for less money. How the model gets built: Meridian or Robyn, depending on which your data team can maintain. That choice is theirs, not ours. Adstock and saturation curves are fitted rather than borrowed from a benchmark deck. The part that decides whether the model is worth anything is calibration: priors come from incrementality tests you have actually run. Where a channel carries real budget and no test exists for it, we run one before the model is trusted on that channel, and the deliverable marks which coefficients rest on evidence and which rest on a prior. Geo design in Mexico specifically: This is a hard geography for holdouts. The Valle de México carries a large enough share of national demand that removing it distorts everything else, so it is never the test cell. We build matched pairs from Guadalajara, Monterrey, Puebla, Querétaro, León and Mérida, matched on pre-period sales trend rather than on population, and the power calculation happens before the test rather than as an explanation for a null result afterwards. A test that cannot detect the effect size you care about is not worth the spend it interrupts, and we will say that before you spend it. What the model will not tell you: It will not tell you what to do next week; weekly noise is larger than most channel effects at ordinary budgets. It will not settle whether Meta or Google deserves the marginal peso at campaign level, because the resolution is not there, and anyone claiming otherwise has not looked at their own credible intervals. It will not replace incrementality testing. It organises what testing has already established, and it prices the channels nobody has tested as exactly that: unpriced. Handover: Code, priors file, data pipeline and holdout results, in your repository, with two working sessions for your analyst. If nobody on your side can re-run it, the model has a shelf life of one quarter and we would rather not build it at all. A model only we can run is a subscription we do not want to sell you. We turn down: Requests to calibrate against last-click numbers so the output agrees with the dashboard. Requests to model eight channels on eighteen months of data. Requests for one ROI figure per channel with no interval around it, since the interval is the finding. And attribution rebuilds, tag manager work or CDP implementation, none of which is what this practice does. Who this is wrong for: Companies whose media budget is small enough that a well-run geo test answers the question outright, which is most companies under about two hundred thousand dollars a quarter in working media. Teams with no analyst. And anyone who needs the answer this month, since the calibration tests alone run four to six weeks in market. Languages: English and Spanish. Model documentation defaults to English, since that is what data teams here read.