A Marketing Mix Model Is Not Ready When the Spreadsheet Is Full

An MMM can estimate channel contribution, ROI, response curves, or scenarios, but those outputs are not interchangeable. Decide which outcome, population, time range, spend intervention, and counterfactual matter to the budget decision. A model built for national revenue may not support a regional pipeline question.

Google’s Meridian documentation frames MMM as causal inference and recommends geo-level data when available. That is not a promise that geography removes confounding. It creates useful variation only when the media, outcome, controls, and business operations are measured at a compatible grain.

Reconcile totals to finance, media platforms, and the operational source before transforming anything. Look for missing periods, duplicated geographies, currency changes, outliers, tracking migrations, promotions, supply constraints, and organic demand shocks. Each repair needs a recorded rule rather than an invisible cell edit.

Controls deserve causal reasoning. A variable that affects both media planning and the outcome may reduce confounding; a mediator created by the advertising can absorb the very effect being estimated. More columns do not necessarily mean more truth.

Meridian explicitly cautions that out-of-sample prediction metrics cannot directly validate causal inference. A model can predict sales while assigning the wrong contribution to correlated channels. Review residuals, convergence, parameter plausibility, baselines, posterior uncertainty, and sensitivity to reasonable specifications.

Keep a model card that states the estimand, data window, transformations, control rationale, priors, exclusions, diagnostics, and unsupported uses. The person receiving the budget chart should be able to see what would make the recommendation fragile.

Experiment results can inform Bayesian priors when their population, treatment, outcome, and time horizon are compatible with the MMM estimand. Google research on MMM calibration shows why alignment matters: an experiment is not generic ground truth that can be pasted onto any channel coefficient.

Refresh on a cadence supported by decision frequency and data stability. Compare new posterior estimates with prior runs, investigate structural breaks, and rerun scenarios under realistic constraints. A model is an operating instrument, not a quarterly slide that becomes authoritative because it contains uncertainty bands.