Google Meridian: How Causal MMM Works at Enterprise Scale

For years, media-mix models (MMMs) tried to answer with regressions that correlated spend and sales. They worked—until complexity exploded. Today, spend spans dozens of channels, crosses regions, and interacts with macro cycles. Simple regressions confuse correlation with causation, rewarding the loudest channel rather than the most effective one.

Google's Meridian project reframes MMM as a causal inference problem. It brings academic rigor Judea Pearl's backdoor criterion into a scalable Bayesian framework that runs across hundreds of geographies. The result: credible causal estimates of incremental impact—without requiring controlled experiments in every market.

At its core, Meridian is a hierarchical Bayesian causal MMM. It models how paid media, organic media, and non-media drivers (pricing, promotions, distribution) influence a KPI such as sales, while adjusting for confounding factors like seasonality or macro trends.

| Component | Role | |-----------|------| | Geo-time decomposition | Treats each geo as an independent observational unit. The same DAG applies to every geo, ensuring parallel scalability. | | Adstock / lag structure | Captures delayed effects of paid and organic media (TV, Search). Non-media drivers are assumed instantaneous. | | Controls | Seasonality, macroeconomic indicators, weather, or holidays—variables that affect both spend and sales. | | Bayesian priors | Share information across geos for stability and shrinkage, allowing small markets to "borrow strength" from larger ones. |