Google Meridian vs Meta Robyn: Which to Pick in India

The short answer

Pick Meta Robyn if your analytics team already works in R, or if most of your media is national television and out-of-home (OOH) that cannot be split by state. Pick Google Meridian if your team works in Python, you can split spend and sales by region, and search and video carry most of your budget. Both are free to download, so the real decision is about your team’s language and your data shape — not about price.

Marketing mix modelling (MMM) is statistical modelling that estimates how much each channel contributed to sales using aggregate weekly history instead of cookies or user IDs. Our explainer on how MMM works covers the mechanics if you need them first.

How they differ on cost

They barely differ at all, which is the honest answer. Both are released under permissive open-source licences, with no vendor invoice, no seat fee and no usage tier.

To be explicit: we have not costed the Growth or Enterprise tier for either tool in rupees, and you should treat any figure you see elsewhere as unverified until someone quotes you. 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 — and that is true of both libraries.

What actually drives the bill is everything the Starter tier excludes. Someone has to build the weekly table of spend, impressions, price, promotions, seasonality and one stable outcome column. Someone has to run a geo holdout or lift test so the model can be calibrated against reality. Someone has to refit quarterly and track whether the estimates are stable. Neither project ships a hosted environment, managed pipelines or a support contract, so all of that is your payroll. If you want a sense of how these internal data builds tend to balloon, our breakdown of what a CDP implementation actually costs follows the same pattern: the licence line is the small number.

Where Google Meridian: Pricing, Fit and Review (India) wins

Meridian’s advantage is geography. It fits a Bayesian hierarchical regression in which each region gets its own effect while borrowing strength from the national average, so a small market does not throw out a wild estimate. A brand that can break spend and sales across 15 states hands the model far more variation to learn from than one that supplies a single national line.

It is also the more opinionated choice for search and video. Reach-and-frequency modelling for YouTube, connected TV and display is built in, and search query volume can be supplied as a control variable — useful when brand search rises because of your own television burst rather than because of your search budget.

Two more practical wins: priors can be set directly from incrementality experiments such as geo lift or conversion lift tests, and the budget optimiser reports both average and marginal return curves per channel, which is the number you need when deciding where the next crore goes. If your analysts are Python-first, the whole thing lands inside the stack they already use, and a GPU cuts sampling time considerably.

Where Meta Robyn: Review & Costs (India) wins

Robyn wins on speed to a defensible first answer. It fits thousands of candidate models using ridge regression with an evolutionary hyperparameter search, then hands you a shortlist to choose between. That shortlist makes uncertainty visible in a way a single chart never does — five plausible models is a more honest briefing document than one confident one.

Its learning curve is also gentler. We rate Meridian’s as medium-to-high, higher than Robyn’s, and Meridian needs roughly a week of analyst time just to understand what the priors are doing before any modelling starts. Robyn asks fewer statistical choices of you up front.

For Indian brands there is a structural win too. Robyn’s default is national, which is simply honest about the constraint that most national TV and OOH buys cannot be split cleanly by state. Geo modelling is possible but is not where the tool is strongest — and if you have no clean geo splits, that costs you nothing. Adstock and saturation are modelled per channel, calibration against lift tests is supported, and the one-page decomposition and allocation output is genuinely usable in a stakeholder meeting.

Which to pick at your stage

Two-person team, one or two channels. Neither. You have nothing to model. Run a geo holdout on the channel you are arguing about — our guide to MMM vs attribution vs incrementality explains why that answers one question faster and more precisely than modelling every question approximately.

Growing D2C or app business, six to ten channels, one analyst. Let the language decide. R team, Robyn. Python team, Meridian. Do not make an analyst switch languages to run their first MMM; the tooling tax is real and the Python interface for Robyn is still in early development.

Multi-state retailer or fintech with regional splits and heavy search and video. Meridian, clearly. This is the case it was designed for, and the geo structure is worth the steeper learning curve.

Mass-market brand with large national television and OOH buys. Robyn. Meridian’s strongest mode is unavailable to you anyway, so you would be running a national model that Robyn handles just as well.

On budget. At any of these stages the licence is ₹0 and the commitment is four to eight weeks of one analyst’s time. If you cannot protect that analyst’s calendar for a month or more, buy neither yet.

What neither does well

Both are MMM, and MMM has hard boundaries. Neither will tell you which creative won, neither will attribute a specific order, and neither can price a channel whose budget never moved — spend the same on radio for three years and no model on earth can value it. Both also break down when you ask too much of too little history; a reasonable rule for either is to keep media variables below a tenth of your number of weeks, which in practice means six to ten channels on 104 to 156 weekly rows. Push past that and you get confident nonsense.

Both also leave the hard part to you. Neither ships data engineering, neither offers a vendor SLA (service level agreement, a contractual support guarantee), and neither provides a hosted dashboard. Open-source pace risk applies to both: check recent commit activity before you build a quarterly reporting habit on either repository.

And the India-specific gap is shared. Clean geo splits for national TV and OOH are listed as an exclusion on both tools’ Growth tiers. Robyn is more upfront about it, but neither one can manufacture state-level data that your media buy never produced.

What this means for you

• Decide on language first. It is the single filter that eliminates one option in most teams.
• Budget the analyst, not the licence. Four to eight weeks of one person’s time is the number to defend internally.
• Run one geo holdout before or alongside the model. Both tools accept experiment results as calibration, and an uncalibrated MMM is a forecast with a nice chart.
• Fix the data plumbing first. If assembling one clean weekly spreadsheet would take more than a couple of days, that is your actual project — see our note on first-party data strategy for small teams.
• If you want PyMC-Marketing in the frame too, read our three-way open-source MMM comparison.

Frequently asked questions

Is Google Meridian or Meta Robyn cheaper?

Neither, on licence. Both are free open-source software with no vendor fee, no seat pricing and no usage tier, so the Starter tier for each is ₹0. The real cost is identical in shape: roughly four to eight weeks of one analyst’s time for a first credible model, plus modest cloud compute. LearnMarketing has not published a rupee figure for that labour at the Growth or Enterprise scale for either tool.

Which one should an Indian brand with mostly national TV spend choose?

Meta Robyn. 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 usually cannot be split cleanly, so that advantage disappears and you end up running a national model anyway. Robyn defaults to national, which matches the data you actually have. Both tools list clean geo splits for national TV and OOH as something they cannot supply.

Do both tools need the same amount of data?

Effectively yes. Both treat 104 weeks of clean weekly data as the practical minimum, with 156 weeks comfortable. Both need spend and impressions per channel, one stable outcome column such as revenue, orders or qualified leads, price and promotion history, and seasonality markers. For Indian teams that means Diwali and festive sale windows, the IPL, exam and admission cycles, and monsoon effects. Monthly data will not work for either.

Can we switch from one to the other later?

Yes, and the switching cost is mostly in skills rather than data. The weekly dataset you build — spend, impressions, outcome, price, promotions, seasonality — is the expensive asset, and it serves either library. What does not transfer is analyst fluency: Robyn is an R package and Meridian is Python on TensorFlow Probability. Build the dataset well and the tool choice stays reversible.