Marketing mix modelling (MMM) used to be something only big brands paid consultancies for. Now Google and Meta both give away open-source MMM tools, and small teams are asking whether they should run one. For most advertisers spending modest sums across two or three channels, the answer is “not yet”. The model won’t have enough data to tell your channels apart, and a cheaper experiment answers the same question better.
What MMM does, in one paragraph
MMM is a regression model. It looks at how your sales moved over time and asks how much of that movement lines up with spend in each channel, after accounting for seasonality, pricing, promotions and trend. Modern versions add adstock, where an ad’s effect lingers for weeks after it runs, and saturation, where each extra pound or rupee buys less than the last. Out comes an estimate of each channel’s contribution and a response curve you can plan budgets against.
It works on aggregated weekly or daily totals, so it needs no cookies, user IDs or tracking consent. As browser and privacy changes erode click-based attribution, that is its big appeal.
The free tools at the time of writing
Google Meridian
Meridian is Google’s open-source MMM. It went to selected users first and became generally available in January 2025. Google’s FAQ confirms the library is free and open source.
Its GitHub page at the time of writing (October 2026) describes a Python library that requires Python 3.11 to 3.13 and recommends at least one GPU. It uses Bayesian causal inference, so you can feed in prior beliefs, including results from experiments. It is built for geo-level data (sales and spend by region, say) but also supports national models, can use reach and frequency data to help plan ad frequency, and includes scenario planning for budget allocation.
Google has also released Meridian GeoX, an open-source library for designing and analysing geo experiments. Its results can become priors that calibrate a Meridian model.
Meta Robyn
Robyn is Meta’s open-source MMM package, which Meta calls “experimental” and “semi-automated”. Its GitHub page lists it as an R package on CRAN, with a Python option in beta that still requires the R package. Robyn’s site points out that it needs no personal data and can calibrate against “ground-truth” methods such as geo tests and Meta lift studies.
PyMC-Marketing
PyMC-Marketing comes from PyMC Labs under the Apache 2.0 licence. It is a Python library for Bayesian MMM with adstock, saturation and budget optimisation, plus customer lifetime value models, and suits teams already working in Python who want more control over model structure.
All three are free to download. None is free to run well: you need clean data, someone who understands Bayesian modelling, and time to check whether the outputs make sense.
How much data MMM needs
Most small-budget projects fall over here. Row count is the wrong measure. Ask how many data points you have for each thing the model must estimate.
Google’s Meridian documentation gives a clear worked example. A national model with 12 media channels, 6 control variables and 8 seasonality knots must estimate 26 effects. Two years of weekly data gives 104 data points, roughly four per effect, which the documentation says is too low to estimate reliably. Group the 12 channels into 3, cut the knots to 2, drop a control and use three years (156 weeks), and you are down to 10 effects with about 15 points each. Google says this “may allow some directional insight”.
Read that again. Three years of weekly data, three channel groups, and the reward is directional insight.
Geo-level data helps a lot, because each region adds data points. The same page’s geo example uses 105 geos over 156 weeks for 16,380 data points. But geo data means spend and sales split by region for every channel, which many small advertisers cannot produce, least of all for influencer seeding or podcasts.
The documentation also stresses variation in spend. If you have spent roughly the same on Meta every week for two years, the model has nothing to learn from. It cannot tell what happens at a spend level you have never been at.
Signs MMM is not worth it for you yet
Tick three or more of these and park MMM for now.
- You have less than two years of consistent weekly spend and sales data.
- One channel takes most of your budget, and the others are small or sporadic.
- Your spend per channel has barely changed week to week.
- You cannot split spend and sales by region.
- A few big launches or promotions dominate your sales.
- Nobody on the team can explain a prior or a credible interval, and there is no budget to hire help.
- You change products, prices or positioning every few months, so last year’s data describes a different business.
None of these is a failing. They are the conditions under which a statistical model cannot separate cause from noise, and a model built under them will still draw confident-looking charts. That is the danger.
₹6 lakh a month across three channels
A D2C skincare brand spends ₹6,00,000 a month: ₹4,50,000 on Meta, ₹1,00,000 on Google Search and ₹50,000 on influencer gifting. It has 18 months of data, and spend on each channel has moved by perhaps 20% month to month.
Feed that into an MMM and three problems arrive together. There are about 78 weekly data points. Meta dominates, and because Search mostly moves with it, the model will struggle to separate the two. Influencer gifting is so small and lumpy that its estimate will mostly reflect whatever prior you chose.
You would get a chart, a response curve and a recommended budget split, with very little reason to believe any of them.
Compare a geo holdout on Meta. Switch Meta off in a group of matched regions for four weeks and compare sales against similar regions where ads kept running. That one test tells you how much of your largest line item is incremental, which is the number that moves your budget most.
What to do instead
Fix measurement basics
MMM, attribution and experiments all depend on clean inputs. A consistent UTM naming convention and a tracking audit before scaling cost far less than a model and improve every report you already use.
Track blended CAC and payback
Divide total marketing spend by new customers each month, then set that against contribution margin per customer using the LTV to CAC payback framework. Crude, yes. But it rests on real money and attribution windows can’t game it.
Run one incrementality test a quarter
Pick the channel with the biggest spend, or the one you doubt most. Open-source tools help here too: Meta’s GeoLift walkthrough uses 90 days of daily pre-test data across 40 cities and recommends a test covering at least one full purchase cycle. Write a clear hypothesis before you start, so you know which result would change your budget.
Create the variation MMM will need later
If you expect to run MMM in a year or two, start varying spend on purpose now. Raise one channel by 30% for a month, cut another, and log the dates. Planned variation plus geo tests is exactly the data a future model needs.
Ask at checkout
A free-text “How did you hear about us?” at checkout picks up channels that your tracking and an MMM may both miss, such as word of mouth and private messaging.
When MMM does make sense
MMM starts to earn its keep when most of these are true:
| Condition | Why |
|---|---|
| Two to three years of weekly data | Enough points per effect, per Google’s worked example |
| Four or more channels with meaningful spend | Budget split decisions are complex enough to need a model |
| Spend varies across weeks and regions | The model needs variation to learn from |
| Offline or upper-funnel spend (TV, outdoor, podcasts, video) | Click-based attribution cannot see these |
| At least one completed incrementality test | Gives the model a ground truth to calibrate against |
| Someone who can own and refresh the model | MMM is a process, not a one-off report |
If that is you, start with a simple model on the free tools, calibrate it with your experiment results, and treat the first version as a hypothesis to test before it becomes a budget order.
The count that decides it
Export weekly spend by channel and weekly sales for as far back as you can, then count the weeks and the channels. Divide the weeks by the number of things the model would have to estimate (channels plus controls plus seasonality terms). Google’s example called about four points per effect too few and about fifteen enough for a direction. If you land nearer four than fifteen, or spend has been flat in most channels, skip MMM for now: pick one channel for a holdout or geo test, write the hypothesis down, plan deliberate spend changes for next quarter and report blended CAC and payback monthly from real revenue. Count again in six months, with an experiment result in hand.
Questions people ask
Is Google Meridian free?
Yes. Google's Meridian FAQ says the library is free to use and open source on GitHub. You still pay for the people, compute and data work needed to run it, and Google recommends a GPU for faster model fitting.
How much data do I need for marketing mix modelling?
There is no single minimum, but what counts is the ratio of data points to things you are estimating. Google's Meridian documentation shows that two years of weekly national data with 12 channels and several controls is too thin, and suggests fewer channels and three years of data for directional insight.
Is Meta Robyn still maintained?
At the time of writing (October 2026), Robyn is available on CRAN and GitHub as an R package, with a Python option in beta that still needs the R package installed. Meta describes it as experimental and semi-automated.
What can I do instead of MMM if my budget is small?
Run a simple geo or holdout test on your largest channel, add a post-purchase survey question, and track blended customer acquisition cost from real revenue and spend. Together these answer most budget questions a small team faces.
Sources
- Google for Developers: Meridian amount of data needed
- Google for Developers: Meridian general product information (FAQ)
- GitHub: google/meridian
- Google for Developers: Introduction to Meridian GeoX
- Meta Open Source: Robyn
- GitHub: facebookexperimental/Robyn
- PyMC Labs: PyMC-Marketing documentation
- Meta Open Source: GeoLift walkthrough
- Marketech APAC: Google rolls out open-source marketing mix model Meridian for general availability
