Google Meridian: Pricing, Fit and Review (India)

What Google Meridian is

Google Meridian is Google’s free, open-source marketing mix modelling (MMM) library — statistical software that estimates how much each marketing channel contributed to your sales, using aggregate weekly history rather than cookies or user IDs. It runs in Python, fits a Bayesian hierarchical regression, and is built around geo-level data, so it suits teams that can split spend and sales by state or region. It is Google’s replacement for its older LightweightMMM library, and the documentation lives at developers.google.com/meridian.

Key features

• Bayesian hierarchical regression fitted with MCMC (Markov chain Monte Carlo) sampling, written in Python on TensorFlow Probability.
• Geo-level modelling built in: each region gets its own effect while borrowing strength from the national average, so small markets do not produce wild estimates.
• Priors can be set directly from incrementality experiments — geo lift, conversion lift or holdout results.
• Search query volume can be supplied as a control variable.
• Handles channels where reach and frequency matter, such as YouTube, connected TV and display.
• Budget optimiser included, with average and marginal return curves per channel.
• Typical run time from minutes to hours; a GPU helps considerably.
• Released under a permissive open-source licence — free to download and run.

Who it fits

Meridian fits a marketing or growth team that already has a Python analyst and at least two years of clean weekly data. The practical minimum is 104 weeks of rows; 156 weeks is comfortable. If assembling a single spreadsheet of spend, impressions, price, promotions, seasonality and one stable outcome column would take you more than two days, fix the data plumbing before installing anything.

For Indian teams, the deciding question is whether you can split the data by geography. A D2C or fintech brand that can break spend and sales across 15 states gives the model far more variation to learn from than one national line, and that is exactly where Meridian is strongest. Heavy search and video spenders get the most out of it, because reach-and-frequency modelling and search query controls are where the library is opinionated.

Stage matters as much as size. A two-person team running a single channel has nothing to model. A team spending across six to ten channels, with festive peaks around Diwali, IPL windows, exam and admission cycles and monsoon effects to control for, has a real MMM problem — and at least one geo holdout test to anchor the model to.

Where it is weak

Meridian is a Bayesian model with genuine statistical choices in it, and the defaults will not suit every business. Budget roughly a week of an analyst’s time simply to understand what the priors are doing, before any modelling starts. Our own comparison of the open-source options rates its learning curve as medium-to-high — higher than Meta Robyn, lower than PyMC-Marketing.

The geo advantage is also its biggest constraint in India. National television and national out-of-home buys usually cannot be split cleanly by state, which is a real problem for brands with heavy offline media. If most of your spend arrives as one national number, you lose the reason to pick Meridian over the alternatives.

MMM in general will not tell you which creative won, attribute a specific order, or measure a channel whose budget never moved. Spend the same amount on radio for three years and no model can price it. Too many channels on too little history produces confident nonsense — a reasonable rule is to keep media variables below a tenth of your number of weeks.

On cost: the software itself is free, but the labour is not, and we have not yet published a rupee figure for what a first Meridian build costs an Indian team. The tiers above are scale guides, not quotes. What we can say from our own reporting is that a first credible model typically takes four to eight weeks of one analyst’s time, plus modest cloud compute.

Alternatives

• Meta Robyn — R-based, fits thousands of ridge regressions and uses an evolutionary optimiser to shortlist candidate models. The fastest route to a defensible first answer if your analytics team already works in R. See our open-source MMM comparison.
• PyMC-Marketing — Python, Bayesian, supplied as building blocks rather than a finished pipeline. Choose it when your business breaks the standard template, and only if someone can explain a posterior distribution. Covered in the same comparison.
• Incrementality testing instead of modelling — if you have one channel to settle rather than a whole budget, a geo holdout answers it faster and more precisely. Our guide to MMM vs attribution vs incrementality sets out when each one applies, and our explainer on how MMM works covers the mechanics.

Frequently asked questions

How much does Google Meridian cost?

The software is free. Meridian is released under a permissive open-source licence, so there is no vendor invoice, no seat fee and no usage tier. Your real costs are analyst time and cloud compute for the Bayesian sampling runs. We have not yet published an India-specific rupee figure for either, so treat any number you see elsewhere as unverified until you get a quote.

How much data do I need before Meridian is worth installing?

Two years of weekly data is the practical minimum and three years is comfortable — 104 rows, ideally 156. You also need spend and impressions per channel, one stable outcome column such as revenue or qualified leads, price and promotion history, and seasonality markers. Crucially, your spend must actually have varied. A channel held at a flat budget for two years cannot be measured by any model.

Is Meridian better than Meta Robyn?

Not universally — it depends on your team and your data. Meridian is the sensible default for Python teams with geo-level data and heavy search or video spend. Robyn gets an R-based team to a first answer faster and shows uncertainty more plainly by shortlisting several candidate models. Both include budget optimisers and both support calibration against lift tests, so the choice is practical rather than statistical.

Does Meridian work for an Indian brand with mostly national TV spend?

Less well. Meridian’s main advantage is geo-level modelling, which needs spend and sales split by state or region. National television and out-of-home buys often cannot be split cleanly, so that advantage disappears and you are running a national model that another tool handles just as well. If your offline media is national and unsplittable, compare options on team language and flexibility instead.