Marketing mix modeling on a mid-size budget, a realistic scope
Open-source MMM tools make modeling cheaper to try. What a team spending a few hundred thousand dollars a year can reasonably expect from one, and what to prepare first.
Marketing mix modeling used to be a project for brands with large budgets and an analytics department. That is changing, because the tools are now free to download. Meta has an open-source model called Robyn, and in March Google announced Meridian, an open-source model offered in limited availability, with general availability planned but no date given. Google cited a Kantar study in that announcement saying 60% of US advertisers use marketing mix models and 58% of non-users are considering them.
Free software does not make the work easy. A mid-size advertiser can get value from a mix model, but only with a narrower scope than the vendor decks suggest.
What a mix model can and cannot do
A mix model uses aggregate data, typically weekly spend and outcomes by channel, plus outside factors such as seasonality and promotions, to estimate how much each channel contributed. It does not need user-level tracking, which is why interest has returned as privacy changes make user-level measurement patchier. Google makes the same argument for Meridian, describing mix models as a privacy-durable option.
It is not a precise instrument. A model estimates a range, and its estimates move when you change the inputs. Treat the output as a structured argument about where money is working, and check it against experiments.
Check whether you have enough variation to model
The question is not budget size by itself. It is whether your history contains enough movement for a model to separate channels. A few honest tests:
- Do spends change independently? If you raise and lower Apple, Meta and Google budgets together every month, the model cannot tell them apart. Channels that move at different times give it something to learn from.
- Do you have a long enough history at a steady grain? We would want a couple of years of weekly data as a starting point. Less than a year is thin.
- Is the outcome clean? Define one target, such as paying users or revenue, and keep its definition fixed across the whole period.
- Are there enough channels to matter? With two channels and a stable split, a simple comparison of periods may serve better than a model.
If you fail the first test, the cheapest fix is to start varying spend on purpose, in planned steps, so that next year's model has signal.
Calibrate with experiments, or do not trust it
Google lists calibration with incrementality experiments among Meridian's features, and that is the idea to copy whatever tool you use. A geo holdout or a time-based pause gives you a measured lift for one channel. Compare that number with what the model says for the same channel. If they disagree widely, the model is the one to doubt. Suppose a holdout suggests a channel drove roughly 10% of paying users and the model credits it with 30%. That gap is a reason to revisit the model's inputs before moving budget.
A sensible scope for a first model
- Pick three to five channels, not fifteen. Combine small ones into an "other" group.
- Include the non-media drivers you know about: pricing changes, major releases, seasonality, and promotions.
- Use the output to answer one decision, such as whether to shift a tenth of the budget from one channel to another, rather than to produce a full allocation.
- Rerun it when you have a new quarter of data and compare with the previous run. Stability is a signal in itself.
Build or buy
The open-source route costs analyst time rather than license fees. If nobody on your team has done this before, a first model will probably take longer than planned, and a bad first model is worse than none because people will quote it. Hiring a specialist for the first build, then maintaining it in-house, is a reasonable split. Whichever you choose, ask to see how the model handles uncertainty and which assumptions it makes about lagged effects.
Practical takeaway
Before you open any tool, export two years of weekly spend and outcomes by channel and look at them on a chart. If channels move together, plan experiments that vary them apart. If they move separately, you have the beginnings of a model, and one calibration test will tell you how far to trust it.