Marketing mix modeling with Meridian and Robyn: a primer

What an MMM really estimates, how much data an Indian brand needs before the numbers mean anything, and how Google's and Meta's free models differ.

A Staunton chess set in rosewood and boxwood laid out on a wooden board
Photo: Wilfredor, CC0, via Wikimedia Commons

Marketing mix modeling (MMM) is a regression on aggregate weekly data that splits sales into a baseline and the incremental contribution of each channel, after modelling how ad effects carry over (adstock) and flatten out (saturation). Google's Meridian is a Bayesian, geo-level model in Python; Meta's Robyn is an R package that searches thousands of ridge-regression models. Both are free, and both give wrong answers with too little data or no experiments to calibrate them.

Key takeaways
  • Adstock + Hill saturation is the core of both tools: carryover, then diminishing returns.
  • 1 / (1 − θ) is the total effect of one week of spend; at θ = 0.8 it is five times the first week.
  • 7 to 10 data points per parameter is Google's rule of thumb; two years of national weekly data gives about 4.
  • Meridian is Bayesian and geo-level in Python; Robyn is ridge regression in R with an evolutionary search.
  • Up to 50%: how far five near-identical models disagreed on sales at different budgets in a Google case study.
  • Calibrate with experiments: lift tests become priors (Meridian) or the MAPE.LIFT objective (Robyn).

Attribution tells you which click came last. MMM answers the CFO's question: without this quarter's ₹2 crore of TV, how many fewer orders?

It needs no cookies, user IDs or pixels, which is why it returned as tracking broke down.

What does a marketing mix model actually estimate?

An MMM estimates the incremental effect of each marketing input on a business outcome, usually weekly sales or conversions, while holding seasonality, price, distribution and other controls fixed.

The output you care about is not the coefficient but what it implies:

  • return on ad spend (ROAS) per channel;
  • marginal ROAS at the current spend level;
  • a response curve that tells you where the next rupee goes furthest.
The story

Through 2017, Procter & Gamble's chief brand officer Marc Pritchard publicly attacked waste in digital advertising, and the company acted on it. According to MediaPost's sources, P&G took about $200 million out of its ad expenses in the fiscal year that ended on 30 June, most of it digital and much of it in the final quarter.

On the earnings call of 27 July 2017, CFO Jon Moeller said the cuts targeted spending "where we were serving bots as opposed to human beings", and had not caused "a reduction in the growth rate" of revenues. His conclusion: the spending P&G cut was "largely ineffective".

That was a judgement read off one quarter's top line, not a measurement. Telling the dead spend from the working spend, before you cut it, is the job MMM is built for.

MediaPost: P&G makes its numbers, slashes ad expenses (27 July 2017)

eBay's 2012 field experiment showed the same trap from the other side: paid search that regression credited with over 1,400% returns turned out to lose money. That gap between a model's answer and an experiment's answer is the whole reason MMM is done with priors and calibration today.

The model has three moving parts:

  • Baseline. Sales you would get with zero media: trend, seasonality (Diwali, end-of-season sale, IPL), price, distribution and macro factors. Robyn uses Meta's Prophet library to decompose trend, seasonality and holidays. Meridian handles time effects with knots, and it can use Google Query Volume as a control so organic brand interest is not credited to paid search.
  • Adstock (carryover). A TV spot aired this week still sells next week. The model spreads each week's spend over later weeks with a decay function.
  • Saturation (shape). The tenth crore on Meta buys less than the first. A Hill curve bends the response so returns diminish.
Weekly spendper channel Adstockcarryover, θ Hill saturationshape, half-point × coefficient βmedia effect Baseline: trend, seasonality,price, distribution, controls Predicted weekly sales= baseline + Σ media effects Lift tests and geo experimentspriors (Meridian) or MAPE.LIFT (Robyn) Inside a marketing mix model
The standard MMM chain used by both tools. Calibration with experiments is what turns a correlation into something closer to a causal estimate.

How does adstock work, and what decay rates are realistic?

Geometric adstock carries a fixed fraction θ of last week's effect into this week, so a spend pulse decays as θ, θ², θ³ and so on.

Google's 2017 paper by Yuxue Jin and colleagues writes the weight at lag l as α to the power l, calls α the retention rate, and caps the carryover window at 13 weeks, since beyond that the weights in their simulation fall below one in ten million. The same paper adds a delayed adstock, where the peak arrives θ weeks after exposure, for media that take time to build.

Starting ranges for θ

Robyn's documentation gives rough ranges for θ from weekly models:

Mediaθ range (weekly)
TV0.3 to 0.8
Out-of-home, print, radio0.1 to 0.4
Digital0 to 0.3

Those are starting bounds for the search, not facts about your brand. The chart shows what three values from those ranges imply for a single week of spend.

Geometric adstock: share of week-0 effect still active
See the data as a table
theta 0.8 (long TV memory) (%)theta 0.5 (mid) (%)theta 0.2 (digital) (%)
Week 0100100100
Week 1805020
Week 264254
Week 351.212.50.8
Week 4416.30.2
Week 532.83.10
Week 626.21.60
Week 7210.80
Week 816.80.40

Source: Computed as theta to the power of the lag, the geometric adstock in Jin et al. 2017; theta ranges from the Robyn docs · Values rounded to one decimal.

The total effect of one week of spend under geometric adstock is 1 / (1 − θ) times the first-week effect. At θ = 0.8 that is five times; at θ = 0.2 it is 1.25 times.

Calculated in this post as the sum of the geometric decay weights, θ to the power k, from Jin et al., Google, 2017, which adds up to 1 / (1 − θ).

This is why adstock choices swing TV ROAS so much: a model that picks θ = 0.8 for TV credits it with four times the carried-over sales of a model that picks 0.5. Last-click attribution effectively sets θ to zero, which is why it underrates TV and YouTube.

What does the saturation curve do to your budget decisions?

The Hill function maps adstocked spend to a response between 0 and 1, with a half-saturation point K where the response is exactly one half and a slope S that sets whether the curve is C-shaped or S-shaped. Jin et al. borrowed it from pharmacology, where it describes drug dose against response.

In Robyn's parameterisation, alpha controls the shape (recommended range 0.5 to 3) and gamma sets the inflection point (0.3 to 1).

Average vs marginal return on a Hill curve Steep: next ₹1 still pays Dashed slope = average ROAS Flat slope = marginal ROAS Adstocked spend → Response →
At current spend (black dot) the average return, the slope from zero, still looks healthy, while the marginal return, the slope of the curve itself, is nearly flat. Budget optimisers chase the marginal number.

The budget implication is the gap between average and marginal ROAS. A channel can show a healthy average ROAS of 3 while its marginal ROAS at current spend is under 1, meaning the last ₹10 lakh lost money.

Both tools report marginal returns, and both budget optimisers move money from flat parts of curves to steep ones. The catch, which the Google paper is honest about, is that the individual shape parameters are hard to pin down even when the overall curve fits well.

A working 1970s-era television set on display in a museum
MMM was built for the television era, when weekly GRPs and sales were the only data advertisers had. It is back because digital tracking is getting weaker.Photo: JIP, CC BY-SA 4.0, via Wikimedia Commons

How much data does an Indian brand need for MMM?

For a national model you need roughly two to three years of weekly data, and even that is thin: three years of weekly data is only 156 data points.

That figure comes from Google's David Chan and Mike Perry, who note that modelling lag and diminishing returns takes three to four parameters per channel, and that a stable regression wants seven to ten data points per parameter. With 20 channels, the parameters can outnumber the data.

What the two tools' docs say

  • Robyn's analyst guide: weekly data as best practice, at least two years of it, four to five years if you only have monthly data, and about ten observations per independent variable.
  • Meridian's documentation: 12 channels, 6 controls and 8 time knots make 26 parameters; two years of weekly national data gives 104 points, about four per parameter, which it calls too low.
  • Meridian's answer is to go geo-level. With 105 geos and three years of weekly data you have 16,380 points, and between about 8 and 74 points per parameter depending on how much the model pools across geos.

Check your own data

Use the calculator to see where your own data sits. It uses the national view; treat a geo-level result as optimistic, because Meridian's docs warn that geos and weeks are not independent observations.

Is there enough data for your MMM?

The default inputs (two years, national, eight channels, five controls) give about 3.6 points per parameter: typical for a mid-sized D2C brand, and why the model needs priors and experiments.

InputGranularityIndian specifics to watch
Outcome (KPI)Weekly, by geo if possibleCombine D2C site, marketplace and quick-commerce sales, or model them as separate KPIs; a marketplace-only spike can be a sale event, not media
TVWeekly GRPs or impressions plus spendBARC data by market; split regional language channels if spend varies independently
Meta and YouTubeWeekly impressions and spend; reach and frequency for videoSplit prospecting and retargeting when spend and impressions diverge, as Robyn's guide suggests
SearchBrand and generic as separate channelsBrand search mostly harvests demand others created; Meridian can use Google Query Volume as a control
Retail media and quick commerce adsWeekly spend per platformSales and ads come from the same platform, so watch for double-counting with the KPI
Price and promotionsWeekly average selling price, discount depthDiwali and end-of-season sales overlap with media bursts; without a promo control, media gets the credit
DistributionWeekly pin codes served, stores, availabilityExpanding to new cities raises sales whatever the media did

How do Meridian and Robyn differ?

Meridian is a fully Bayesian model that prefers geo-level data and takes experiment results as priors; Robyn is a ridge regression tuned by an evolutionary optimiser that picks from many candidate models on fit and plausibility. Both are open source, both run adstock and Hill saturation, and both include a budget allocator.

FeatureGoogle MeridianMeta Robyn
LanguagePythonR core; Python API in beta since Nov 2023, which still needs the R package
EstimationBayesian causal inference, posterior credible intervalsRidge regression; Nevergrad searches hyperparameters across thousands of models
Data levelNational or geo; docs recommend geo when possibleNational time series, usually weekly
AdstockGeometric carryover with priorsGeometric, Weibull CDF or Weibull PDF (for lagged peaks)
SaturationHillHill (alpha and gamma)
CalibrationExperiment results become ROI priors; GeoX runs geo tests that feed themMAPE.LIFT objective penalises models far from lift-test results
Model selectionOne posterior; check credible-interval widthTrade-off between NRMSE (fit) and DECOMP.RSSD (spend share vs effect share)
VideoReach and frequency inputsSpend or impressions
Licence and statusOpen source, generally available since Jan 2025MIT licence, labelled experimental

Which one to start with

The practical split is skills and data shape:

  • Python analysts with state or city-cluster order data should start with Meridian's geo model.
  • A national weekly series and an R analyst point to Robyn, whose automated search and one-page summaries give a faster readout.
Bayesian Methods for Media Mix Modeling with Carryover and Shape EffectsYuxue Jin et al. · Google Inc. · 2017

The paper that set the adstock-plus-Hill template both tools now use. The authors simulated 500 datasets with three media channels and one control, then fit the model with two years and with sixty years of weekly data. With two years, the estimated response curves were biased low by 18% to 40% at the points they checked (for example, minus 39.7% for one channel at half the spend range); with sixty years, bias fell to within about 10%. They also fit 2.5 years of weekly data from a shampoo advertiser with TV, magazines, display, YouTube and search, and found the data did not carry enough information to pin down carryover and shape. The limit: it is a methods paper on simulated and one real dataset, and its message is that priors drive the answer when data is small.

Read the paper (free)
A Brunsviga 15 mechanical calculator with its covers removed to show the gears
Fitting a regression once meant cranking machines like this Brunsviga. Robyn now fits thousands of candidate models per run.Photo: CEphoto, Uwe Aranas, CC BY-SA 3.0, via Wikimedia Commons

Why do two well-fitting models give different budget advice?

Because media spend is a weak signal in sales data, many models fit history almost equally well yet disagree on which channel works.

Chan and Perry show this with a real US advertiser: five plausible models, each with R² of 0.98 or 0.99, and four of them with an out-of-sample error of 6 to 8% on the last 12 weeks. Their predicted sales at different budgets differed by up to 50%, and they disagreed on how much to put into the advertiser's largest channel.

Three causes

  1. Channels move together. Brands raise TV, YouTube and Meta in the same festive weeks, so the regression cannot tell them apart.
  2. Selection bias. Search ads appear when people are already looking to buy, so search gets credit for demand it did not create, exactly what eBay's experiment showed.
  3. The range problem. If you have always spent between ₹40 and ₹60 lakh a month on Meta, the model knows nothing about ₹1.5 crore.
Fit statistics do not validate an MMM. Out-of-sample prediction, stable parameters on refit, and agreement with experiments do.The lesson of Chan and Perry, Google, 2017
Challenges and Opportunities in Media Mix ModelingDavid Chan and Mike Perry · Google Inc. · 2017

A frank account from Google's own measurement team of why MMMs often fail. It covers limited data (three years of weekly national data is 156 points, against 3 to 4 parameters per channel and a rule of thumb of 7 to 10 points per parameter), correlated inputs, selection bias from targeting and seasonality, and model uncertainty. Its case study of five near-identical fits with budget recommendations that disagree is the best single argument for calibration. Its proposed fixes include Bayesian priors, category-level models, geo models and validation against experiments. The limit: a position paper, not a benchmark; the advertiser is anonymised.

Read the paper (free)

How do you calibrate an MMM with lift tests?

You run an experiment that measures a channel's causal effect directly, then force the model to agree with it. In Meridian, the experiment result becomes a prior on that channel's ROI; in Robyn, it becomes a third objective, MAPE.LIFT, which penalises candidate models whose predicted contribution is far from the test result.

Google's 2024 paper on calibration reparameterises the model so that ROAS itself is a parameter, which makes it straightforward to put a prior from an experiment on it.

Robyn's documentation cites a third-party white paper by Analytic Edge reporting that uncalibrated models differed from ground truth by about 25% on average. It also cites a 2022 hackathon simulation in which calibration brought ROAS estimates closer to the truth, with accuracy improving as more channels and up to 10 studies per channel were added.

Two kinds of test feed this:

  • Conversion lift tests inside Meta or Google, which hold out users. They work only for digital and only measure that platform.
  • Geo experiments, which hold out or increase spend in some regions and compare them with control regions. These cover TV, out-of-home and any platform. Meridian GeoX, generally available since September 2026, needs daily geo-level data and at least 10 geos for a single-cell design, with 50 or more recommended.

India suits geo tests better than many markets: states and metro clusters differ, regional TV and language campaigns are already bought by market, and quick-commerce and D2C order data come with pin codes. Robyn's guide adds two cautions: match the test's period and KPI to the model, and skip low-confidence results.

Media Mix Model Calibration With Bayesian PriorsYingxiang Zhang et al. · Google · 2024

Proposes writing the model so that each channel's ROAS is a parameter with its own prior, which lets an experiment result or expert view be put in directly instead of guessing which coefficient prior to tighten. The simulation uses 100 geos and two years (104 weeks) of weekly data and compares calibrated and uncalibrated geo-level hierarchical models. Calibration produced tighter credible intervals and lower bias on average ROAS, and reduced bias substantially when an unobserved confounder was present. The limit: simulated data built to match the model's own assumptions; real confounding can be worse.

Read the paper (free)

Where should MMM sit in an Indian brand's measurement stack?

MMM sets the quarterly budget split; experiments verify the biggest channels; platform attribution steers daily optimisation. None of the three replaces the others.

LayerQuestion it answersCadenceTrust order
Experiments (lift and geo tests)Did this channel cause extra sales?A few per year1st
MMM, calibrated to experimentsHow should the budget split across channels?Quarterly2nd
Platform attribution and server-side trackingWhich ad or audience inside a channel works?Daily3rd

I would build it in that order of trust: experiments first, MMM calibrated to them second, platform-reported ROAS last.

The honest first step

For most Indian brands below ₹100 crore in revenue, the honest first step is not Meridian or Robyn but data hygiene: a single weekly table of sales by state, spend by channel, price, promotions and distribution, going back two years.

If your tracking is still unreliable, fix it first with server-side tracking through the Conversions API, and read how to build an attribution stack after cookies for the layer MMM sits on top of.

It settles the brand argument

MMM also settles an argument that most Indian marketing teams have every year: how much to put into brand. Because it measures carryover, it can show TV and YouTube paying back over months, the core of the brand versus performance balance.

It also changes unit economics: a channel with long carryover can look expensive on CAC payback while being the cheapest source of incremental customers. For how the industry got from last-click to here, see the history of performance marketing.

A tall Doordarshan television transmission tower in Kasauli, Himachal Pradesh
Indian television is still bought market by market, which makes regional geo experiments easier to design than in many countries.Photo: Sahil Dhiman, CC BY-SA 4.0, via Wikimedia Commons

If you would rather have someone set this up end to end, from the data table to the first calibrated readout, my work page shows the kind of marketing systems I build.

Questions people ask

Is MMM better than multi-touch attribution?

They answer different questions. Multi-touch attribution assigns credit across tracked clicks and views for one user and misses offline media and untracked journeys. MMM estimates the incremental effect of each channel from aggregate data, including TV and out-of-home, but only at weekly and channel level. Meridian's own docs advise against feeding it campaign-level data and suggest attribution for that granularity.

Can a small D2C brand run Meridian or Robyn?

Yes, technically, since both are free. The constraint is data: with two years of weekly national data and eight channels, you have under four data points per parameter, well below the 7 to 10 that Google's researchers suggest. Cut the model to four or five channels, use geo-level data if you have order pin codes, and calibrate with at least one experiment.

How often should an MMM be refreshed?

Most teams refit quarterly and use the response curves for the next quarter's budget. A refit with four extra weeks doubles as a stability test: if a channel's ROAS swings sharply, the model was not well identified.

Should I trust MMM tools built by Google and Meta?

The code is open, so you can read exactly what it does. The risk lies in defaults and priors accepted without checking; running both tools on the same data and comparing them with a geo test is cheap insurance.

Sources

  1. Jin, Wang, Sun, Chan, Koehler: Bayesian Methods for Media Mix Modeling with Carryover and Shape Effects, Google (2017)
  2. Chan and Perry: Challenges and Opportunities in Media Mix Modeling, Google (2017)
  3. Zhang et al.: Media Mix Model Calibration With Bayesian Priors, Google (2024)
  4. MediaPost: P&G makes its numbers, slashes ad expenses (July 2017, source for the story)
  5. Blake, Nosko and Tadelis: Consumer Heterogeneity and Paid Search Effectiveness: A Large Scale Field Experiment, NBER Working Paper 20171 (2014; Econometrica 2015)
  6. Google Meridian documentation: About Meridian (2026)
  7. Google Meridian documentation: Amount of data needed (2026)
  8. Meta Robyn documentation: Features (adstock, saturation, objectives, calibration) (2026)
  9. Meta Robyn: Analyst's guide to MMM (2026)
  10. Robyn GitHub repository, MIT licence (2026)
  11. TechWyse: Google takes Meridian GeoX to general availability globally (2026)
  12. PPC Land: Meridian lands inside Analytics 360 (2026)
  13. Think with Google India: Google Meridian marketing mix modelling (2026)
  14. PPC Land: Google opens Meridian marketing mix model (2025)
marketing mix modelingMMMMeridianRobynadstockincrementalitymeasurement
Sapta

Sapta (Saptarshi Sarkar) is a mechanical engineer turned marketer in Bengaluru and a solo touring rider. He writes the seven Saptaverse blogs. Work with him or read more of Marketblog.

Keep reading in Marketblog

From the other blogs