I see marketing mix modeling (MMM) becoming easier to access, but I do not think it has become easy to get right.
After several conversations about MMM adoption, I keep hearing the same concern: “We believe in MMM, but we do not know how to get started.”
My answer is that open-source platforms have lowered the barrier to entry in a meaningful way. What they have not lowered is the level of expertise required to produce results that are trustworthy, explainable, and useful for decision-making.
Open-source MMM has changed the starting point
I am seeing MMM adoption accelerate because marketers need more durable measurement methods. Almost half of U.S. marketers expect to invest more in MMM over the next year, and many now rank it as one of the most reliable measurement approaches available.
The open-source shift is real. Three production-grade libraries now give teams a practical way to approach MMM across a wide methodological spectrum.
- Robyn (Meta, R): I see this as the most approachable starting point because it includes automated hyperparameter search through Nevergrad, Pareto frontier model selection, decomposition, and response curve plots. It is also the one I use most often because it is highly customizable.
- Meridian (Google, Python/TensorFlow): I view Meridian as a more rigorous option, especially because it uses Bayesian inference, geo-level priors, and principled uncertainty quantification. The tradeoff is a steeper learning curve.
- PyMC-Marketing (PyMC Labs, Python): I consider this the most flexible path. It offers a full probabilistic model that comes closest to academic-grade Bayesian MMM, but it also demands the most statistical fluency.
This generation of tools has removed the old $150,000 to $500,000 consulting gate that used to be the primary path into MMM. A team with R or Python expertise and reasonably clean historical data can now run a model in-house.
The caveat I always make explicit is this: “free tool” does not mean “free model.” The software may be free, but the domain expertise needed to configure it correctly is not. That expertise is a major part of the value.
The vendor landscape is crowded and complicated
I also see a fast-growing SaaS layer built on top of open-source MMM. To evaluate it clearly, I find it helpful to separate vendors into a few practical groups.
Data-layer-first vendors
Platforms like Rockerbox and Northbeam started with attribution and data collection, then added MMM. Their advantage is usually pipeline speed and data access, not deep modeling flexibility or customization.
Measurement-first vendors
Platforms such as Measured, Analytic Partners, Ekimetrics, and Nielsen Gracenote tend to offer more rigorous modeling and enterprise-grade capabilities, usually at a higher price point.
Google Meridian and GA360
I think Google’s decision to open-source Meridian is both a generous contribution to the field and a strategic move. When a walled garden funds and packages a measurement methodology that can be used to evaluate its own channels, I believe it is worth maintaining healthy skepticism about priors, defaults, and assumptions, even when the code is transparent.
The practical vendor question I keep coming back to is simple: who owns the data layer, and does that ownership create conflicts in the modeling layer?
Challenge 1: Data access can quietly break MMM
I think data access is the most underappreciated MMM implementation blocker. A well-specified model needs more than a quick export from a dashboard.
- I usually want two to three years of weekly data as a baseline, so the model can capture at least two full seasonality cycles and enough spend variation to learn from.
- I need consistent channel-level spend granularity, not just a broad “digital” bucket. Search, social, display, video, and other channels need to be separated.
- I need offline channels such as TV, OOH, radio, events, and direct mail, even though they often live in different systems, belong to different teams, and use incompatible time periods.
- I need external covariates, including macro indicators, competitor activity, pricing data, and product launch calendars.
- For B2B, I often need even more history because longer sales cycles and lower conversion volumes make the data requirements more demanding.
In practice, I often find that the real blocker is the six-week data archaeology project that happens before modeling begins. Finance owns revenue. The brand team owns TV. The agency owns digital spend. A spreadsheet from 2021 may be the only record of trade promotions.
The model is only as good as the data archaeology behind it, and that is rarely the part anyone highlights in a vendor demo.
Challenge 2: I still have to roll up my sleeves
AI assistants have lowered the syntax barrier. They can scaffold a Robyn run, generate a Meridian configuration, or help debug a PyMC model. What they cannot reliably do yet is make the judgment calls that determine whether an MMM is credible.
- I still have to decide where to land on a Pareto frontier across hundreds of model solutions, balancing NRMSE against DECOMP.RSSD tradeoffs.
- I still have to know whether Nevergrad’s optimizer has meaningfully converged or simply landed in a local minimum.
- I still have to configure adstock transformation parameters, including Weibull shape and scale or geometric decay, so they reflect realistic channel behavior.
- I still have to diagnose why a model gives a channel an implausible contribution and decide whether the fix is a prior, a data correction, or a variable exclusion.
In other words, if I try to vibe code my way into MMM, I may end up with a model that appears to work but is wrong in ways I will not catch. The scripting is not the hardest part. The real work is validating the output, including using channel-specific incrementality experiments to calibrate the model.
Challenge 3: Human expertise is not optional
Even if the tools mature enough for AI to run a competent default MMM, I still see human expertise as essential. The irreplaceable work is encoding business context that no model can infer from the data alone.
- Adstock and carryover context: I need to know whether a TV buy carries over for four weeks, paid search carries over for three days, or a brand awareness campaign decays over months. That knowledge usually lives with channel experts, not inside the dataset.
- Saturation curve shape: I need to recognize when a channel is probably approaching diminishing returns before the model says so, and I need to question the model when it suggests something implausible.
- Guardrails and anomaly handling: I need to explicitly model or flag COVID troughs, product launches, pricing shifts, and macro disruptions as structural breaks. AI does not automatically know a client had a pricing crisis in Q3 2022.
- Interpretation sanity checks: If a model assigns 40% of contribution to TV for a brand spending $2 million on TV, I need the experience to say, “That feels wrong,” and investigate. That intuition is earned, not computed.
- Organizational translation: A technically correct model has little value if I cannot explain why it recommends moving 15% of search budget to CTV in language a CMO and CFO will act on.
I start with the groundwork before the model
The best place to begin is not the model itself. I start by understanding what data is needed, who owns it, and who can help interpret it in the context of real marketing decisions.
None of that is quick or easy, but it is essential if I want meaningful insight from MMM, whether I choose an open-source library or a subscription-based platform.
As a practical first step, I would download Robyn’s demo script and experiment with sample data before applying MMM to my own business data. That kind of hands-on testing makes the strengths, limits, and judgment calls much clearer.
Inspired by this post on Search Engine Land.
