Open-source MMMs are inspectable. That doesn't make them correct.

A car with a glass hood is still a bad car if the engine is built wrong. You can watch every piston fire, trace every belt, follow every wire, and still end up stranded. The problem was what was under the hood, not visibility. The same logic applies to open-source marketing mix models. The code is public so anyone can fork the repo, but none of that tells you whether the model is actually measuring your marketing correctly. And that’s what truly matters for brands making real budget decisions.

The open-source argument tends to go: you can see the code, so you can trust it. But inspectability and correctness are not the same thing. An open-source MMM can be fully readable and still be built on assumptions that don't hold in real marketing environments. When those assumptions are wrong, the outputs are wrong, and you can see exactly how they got there.

Key takeaways

• Open-source MMMs are inspectable, meaning anyone can read the underlying code. But inspectability is not the same as accuracy or correctness.
• Most open-source MMMs share a foundational assumption that marketing effects can be cleanly separated from baseline demand. This assumption breaks down when spend is correlated with seasonality and events, which it almost always is.
• When a model's bedrock assumptions are wrong, more data doesn't fix the problem. The errors are baked into the model's design.
• The practical consequences include systematic underattribution of upper-funnel channels, inflated baseline estimates during high-spend periods, and budget recommendations that can be significantly off.
• Research comparing open-source MMM baselines against a mechanistic alternative found attribution errors three to four times larger in the open-source models, even under controlled conditions with known ground truth.
• A model being proprietary doesn't make it a “black box.” What matters is whether the team behind it understands its strengths, limitations, and the decisions built into it.
• When evaluating an MMM, the more useful questions are about how the model handles cross-channel interaction, correlated spend, and non-stationary efficiency, not whether its code is publicly available.

What "open source" actually means in MMM

Open-source MMMs—tools like Meridian from Google and Robyn from Meta—make their code publicly available. You can read the modeling logic, inspect the priors, and see exactly how the model gets from inputs to outputs. That level of transparency is valuable for academic research, building intuition about how MMMs work, and auditing whether a vendor has made changes from the base model.

What it doesn't tell you is whether the model's foundational design is right for your marketing environment. That question requires something different: an understanding of whether the assumptions built into the model reflect how marketing actually works. (And, yes, every single MMM is built on assumptions.)

The assumption that breaks things

Most open-source MMMs are built on the shared assumption that your total revenue can be decomposed into independent contributions from baseline demand and each marketing channel. In other words, paid social contributes X, paid search contributes Y, baseline demand contributes Z, and together they add up to your results. Clean, simple, inspectable.

But real marketing environments don't work that way:

• Marketing spend is deliberately synchronized with seasonality and peak periods (you intentionally ramp up spend during BF/CM, for example). 
• Upper-funnel investment changes how lower-funnel channels perform. 
• Efficiency shifts over time. 
• Effects compound across channels rather than stacking independently. 

When a model assumes everything is separable and additive, it can't actually tell the difference between what your marketing drove and what would have happened anyway, especially during critical seasons like BF/CM when spend and demand move together.

This is what researchers call an identifiability problem, and it's not something that resolves with a bigger dataset. The errors are built into the model itself, which means they're predictable and consistent, and they persist regardless of how much data you feed the model.

What bad assumptions look like in practice

Understanding the theory matters less than understanding what it actually costs you. When a model can't reliably separate marketing effects from baseline demand, a few things tend to happen:

Upper-funnel channels—brand awareness, video, connected TV—get systematically underattributed. Because their effects are diffuse and delayed, a separable model has a harder time isolating their contribution, so it often absorbs that signal into baseline demand instead. You end up with a measurement picture that makes your top-of-funnel investment look weaker than it is.

During high-spend periods like the holiday season, the problem gets worse. Spend and demand move together almost perfectly during these windows, which means the model has even less ability to separate them. Estimated holiday effects frequently end up overstated, and the marketing-driven contribution gets blurred.

Research from Prescient AI’s data science team evaluated two widely used open-source MMM baselines against a mechanistic alternative using a synthetic environment with known ground truth, meaning the true marketing contribution was observable and measurable. The open-source baselines produced attribution errors 3–4 times larger than the mechanistic model. And when those attribution errors were carried into budget optimization during peak periods, one baseline recommended overspending by up to 81% relative to the known optimum.

The point isn't that the models produced wrong answers because they were implemented incorrectly. They produced wrong answers because their design wasn't capable of representing the actual dynamics of the marketing system being measured.

This is why we call Prescient a “glass box”

The conversation around MMM transparency tends to collapse into two buckets: open source (good, trustworthy) and proprietary (suspect, opaque). But that framing leaves things out.

There's a difference between a model that's publicly inspectable, a model whose developers genuinely understand every structural decision they made and why, and a model that no one—internally or externally—can fully explain. The first and second aren't the same thing, and neither are the second and third.

Cody Greco, Prescient's CTO and co-founder, has written about this distinction directly. His framing is this: the right term for a well-understood proprietary model isn't “black box,” it's “glass box.” The team building it knows exactly what the algorithms are doing and why. The fact that they don't publish the code doesn't mean they don't understand it. In fact, it was precisely because Greco could read and understand the existing open-source models that he was able to evaluate their limitations and decide none of them were right for the measurement problems he was trying to solve.

A model being proprietary means the code isn't public. It doesn't mean the model is a mystery to the people who built it, and it doesn't mean their decisions are worse than the ones baked into an open-source alternative.

What to actually ask when evaluating an MMM

If code transparency isn't the right filter, what is? A few questions tend to get closer to what actually matters:

How does the model handle cross-channel interaction? If it treats channels independently that's a limitation marketers need to understand. Marketing channels influence each other: upper-funnel investment shapes lower-funnel performance, and a model that can't represent that will misattribute both.

How does the model handle correlated spend? If your brand spends more during high-demand periods (and nearly every brand does), the model needs to be able to separate marketing-driven lift from demand that would have materialized anyway. This is where separable models tend to fail most visibly.

Has it been validated against ground truth? It's difficult to validate any MMM against real-world ground truth because we rarely know what the true marketing contribution actually was. Synthetic environments with known parameters are one way researchers test this directly, and they tend to surface failures that wouldn't show up in a standard hold-out test or a match against platform-reported numbers.

What are the acknowledged limitations? Any honest measurement provider should be able to tell you where their model is most and least reliable. If the answer is “our model handles everything equally well,” that's worth probing.

Where Prescient comes in

Prescient was built by a team that started by reading the existing open-source models and identifying what was wrong with them—not in the sense of implementation bugs, but in the sense of foundational design decisions that were incompatible with how real marketing systems behave. OMEN, Prescient's underlying model, is mechanistic and nonlinear by design. It's built to account for cross-channel interaction, non-stationary efficiency, and the correlated dynamics between spend and demand that cause separable models to break down. That design reflects a deliberate set of choices, grounded in research, about what an MMM needs to represent in order to be accurate.

If you're evaluating your current measurement approach and want to understand where open-source or separable models may be falling short for your specific channels and spend patterns, Prescient's team can walk you through it when you book a demo.