Meta Robyn: Review & Costs (India)

What Meta Robyn is

Meta Robyn is a free, open-source marketing mix modelling (MMM) package built by Meta. MMM is statistical modelling that estimates how much each marketing channel contributed to sales, using aggregate weekly data instead of cookies or user IDs. Robyn runs in the R programming language, fits thousands of candidate models and hands you a shortlist to choose between — which makes it the quickest route to a first defensible MMM if your analytics team already works in R.

Key features

• Ridge regression paired with an evolutionary hyperparameter search (Nevergrad) across thousands of candidate models
• Adstock (carryover, where an ad seen this week still drives a purchase later) modelled per channel
• Saturation curves, so the model knows the tenth crore of spend buys less than the first
• Calibration against experiment results — geo lift, conversion lift or holdout tests
• Budget optimiser and response curves included
• A one-page output covering decomposition, response curves and budget allocation
• Typical run time of minutes to a couple of hours, because it fits many models rather than one
• Geo-level modelling is possible; the default is national
• Permissive open-source licence, documented at facebookexperimental.github.io/Robyn

Who it fits

The filter that matters is language, not headcount. If your analysts live in R, Robyn is the sensible starting point. If they live in Python, the setup cost is real and you should look at Google Meridian or PyMC-Marketing first.

Beyond that, it fits a team that can assemble two to three years of clean weekly data. The practical minimum is 104 weeks; 156 weeks is comfortable. You also need spend and impressions per channel, your sales or lead outcome, price and promotion history, and seasonality markers. A useful rule: keep the number of media variables under a tenth of your number of weeks, which in practice means six to ten channels.

For Indian teams there is a structural reason Robyn often lands better than its geo-heavy rivals. National television and national out-of-home buys usually cannot be split cleanly by state, so the geo-level modelling that makes Meridian strong is unavailable to exactly the brands with heavy offline media. Robyn’s national default is honest about that constraint rather than fighting it.

Seasonality is the other India-specific input. Diwali and the big festive sale windows, the IPL, exam and admission cycles, and monsoon effects for weather-sensitive categories all need to be in the model. Leave them out and the model will credit your media for the calendar.

Stage-wise, Robyn suits brands spending across several channels where last-click attribution has stopped answering the budget question — typically a direct-to-consumer (D2C) brand, an app business or a multi-channel retailer with an analyst who can own a quarterly refresh.

Where it is weak

The shortlist is Robyn’s best and worst feature at once. Seeing five plausible models makes the uncertainty visible in a way a single chart never does — and it also lets a junior analyst quietly pick the model that flatters the CMO’s favourite channel. If you adopt Robyn, write down your model-selection criteria before you look at the results.

• R-only in practice. A Python version is in early development, so Python-first teams pay a tooling tax.
• National by default. Geo modelling is possible but is not where the tool is strongest.
• Dense output. The one-page report is good for a stakeholder meeting and heavy going for a first-timer.
• Labour, not licence, is the cost. Budget roughly four to eight weeks of one analyst’s time for a first credible model, plus modest cloud compute.
• It cannot measure what never moved. A channel with a flat budget for two years is invisible to any MMM, Robyn included.
• Open-source pace risk. Check recent commit activity on the GitHub repository before you build a quarterly reporting habit on it.

Robyn also will not tell you which creative won or attribute a specific order. That is not a flaw so much as a boundary — see MMM vs attribution vs incrementality for where each method belongs.

Alternatives

• Google Meridian — Bayesian hierarchical modelling in Python, built around geo-level data and channels where reach and frequency matter. The better pick if you can split spend and sales by state. Compared side by side in our open-source MMM comparison.
• PyMC-Marketing — Bayesian MMM as configurable building blocks in Python, plus customer lifetime value models. Choose it when your business breaks the standard template; skip it if nobody can explain a posterior distribution. Also covered in the same comparison.
• Incrementality testing — a geo holdout measures one channel precisely rather than every channel approximately. Cheaper to start, and it makes any MMM you build later more credible. Background in how MMM actually works.

Frequently asked questions

Is Meta Robyn free?

The software is free. Robyn is released by Meta under a permissive open-source licence, so there is no vendor fee, no seat pricing and no usage tier. The real cost is labour and compute: roughly four to eight weeks of one analyst’s time to reach a first credible model, modest cloud compute for the fitting runs, and ongoing work to refresh the data each quarter. LearnMarketing has not published an India rupee figure for that labour.

Do I need to know R to use Meta Robyn?

Effectively, yes. Robyn is an R package, and a Python version is only in early development. If your analytics team already works in R, that is Robyn’s main advantage over the alternatives. If it works in Python, the language switch is a genuine obstacle and Google Meridian or PyMC-Marketing will usually get you to a usable model faster.

How much data does Meta Robyn need?

Two years of weekly data is the practical minimum and three years is comfortable — 104 to 156 weekly rows. Monthly data will not work, because it gives you only 24 to 36 rows. You need spend and impressions per channel, one stable outcome column such as revenue or orders, price and promotion history, and seasonality markers. Your spend must also have varied; a flat budget cannot be measured.

Should I choose Meta Robyn or Google Meridian?

Choose by team and data, not by brand. Robyn suits an R team that wants a fast, opinionated first answer. Meridian suits a Python team that can split spend and sales across regions or states, and that spends heavily on search and video. For Indian brands with large national television or out-of-home buys that cannot be split by geography, Meridian’s strongest mode is often unavailable anyway.