Google Meridian

Definition

Meridian is Google's open-source marketing-mix-modeling library, released January 2024. It implements hierarchical Bayesian MMM in TensorFlow Probability (TFP) + JAX, with built-in adstock (carryover), Hill-saturation curves, geo-hierarchical pooling, and MCMC sampling (NUTS by default) that produces full posterior distributions over channel contributions, ROI, and mROI rather than the single point estimates frequentist MMMs produce. The model spec follows Jin et al. 2017 ("Bayesian Methods for Media Mix Modeling with Carryover and Shape Effects") with extensions for reach/frequency, treatment of paid search, and incrementality-experiment calibration (priors informed by lift studies, geo-experiments, or holdouts). The catch — and the reason "Meridian" and "Bayesian MMM platform" are not the same thing — is that Meridian gives you a model definition and a Python package. It does not give you (a) ingest from non-Google ad platforms (you write your own Meta/TikTok/LinkedIn/Amazon/Flipkart connectors), (b) an optimizer (translating posterior ROI distributions into a budget reallocation is left as an exercise), (c) an execution engine (pushing reallocations back into ad-platform budgets), (d) a UI for non-technical marketers, or (e) per-tenant ops (auth, billing, multi-brand isolation, SLA monitoring). You also own the compute — Meridian's MCMC run on a 2-year weekly dataset across ~10 channels typically needs a GPU and 30-90 minutes per refresh; Google's published guide assumes Vertex AI. Managed MMM platforms close the gap by shipping the connectors, optimizer, execution engine, and ops on top of an equivalent-quality model. The DIY-vs-platform tradeoff is the same one as Apache Airflow vs. Astronomer, Postgres vs. RDS, or Kubernetes vs. ECS: Meridian wins on flexibility for in-house data-science teams; a managed platform wins on time-to-first-insight and operational cost for marketing teams who want the answer, not the codebase.

How it applies in India

Meridian runs on Indian data without modification — the model is country-agnostic. The gaps that show up in practice are operational: (a) TFP + JAX install pain on the laptops most Indian marketing-analytics teams use, (b) zero out-of-the-box connectors for Meta Ads India INR billing, Google Ads INR conversion currency, JioMart Ads, Flipkart Ads, or Amazon India Sponsored Products, (c) no Hindi/regional-language creative-fatigue tracking, and (d) the GPU-on-Vertex-AI assumption is a cost item Indian SMBs and mid-market brands rarely justify. For a 5-person India D2C team, getting to first useful Meridian output is realistically a 6-8 week engagement with an external Bayesian-statistics consultant.

Frequently asked questions