Why open-source models fall short for next-generation MMM

The marketing industry is currently buzzing about “next-generation” marketing mix modeling (MMM), applauding how yesterday’s solutions are evolving to meet the needs of today’s marketers. Recent advancements in MMM are often attributed to the availability of open-source models from companies like Meta and Google—and most of today’s so-called “next-gen” MMM solutions are anchored on adaptations of these open-source models.

When we started Prescient AI in 2019, we thoroughly evaluated available open-source MMM solutions, and determined they wouldn’t suffice for what we intended to deliver for marketers. Our analysis of existing models revealed critical shortcomings: 

• Architectural limitations prevented campaign-level insights. These models were designed to operate at the channel level, lacking the granularity to optimize specific campaigns.
  
• Computational inefficiency made frequent refreshes impossible. The models required significant computing power and time to run, making daily or even weekly refreshes impractical. This meant insights were always looking backward at data that might be weeks or months old.
  
• Simplified assumptions contradicted marketing reality. We observed that these models imposed uniform assumptions about marketing effects that directly contradicted what all marketers know to be true from real marketing experience.
  
• Limited causal understanding. Perhaps most critically, these models focused primarily on correlation rather than “cause and effect.” They could achieve seemingly impressive accuracy while completely misunderstanding what actually drove performance, leading to potentially damaging recommendations for optimization.

These weren’t minor limitations or superficial issues that could be patched—they were structural constraints embedded in the mathematical foundations of traditional MMM. We could either work within the constraints of these models, accepting their fundamental limitations, or we could develop something entirely new. To deliver the caliber of insights marketers truly needed, the choice was clear.

Since our initial analysis, new open-source models like Meta’s Robyn (released in 2021) and Meridian by Google (2025) have included notable improvements like machine learning elements and better handling of diminishing returns, but they are simply enhancements to the same outdated models. They still operate on the same mathematical foundations as earlier marketing mix models, inheriting the same architectural limitations that prevent them from accurately capturing the complex realities of marketing today.

By the time Robyn and Meridian were released, Prescient AI was already operating with an entirely different approach—one based on core principles that govern how marketing actually works in the real world. We call these the “Laws of Marketing” because they represent observable realities that marketers describe if you ask about their lived experience running paid campaigns. Many measurement systems (including most MMM solutions) make simplified assumptions that directly contradict these existential marketing realities. 

For MMM solutions to deliver recommendations marketers can trust won’t lead them astray, the underlying system needs to be a reflection of what’s really happening. That requires more than tweaking existing models.

By building the models behind Prescient AI from scratch, we designed an architecture to overcome past limitations. We incorporated advances in machine learning and causal inference that simply didn’t exist when MMM was first developed. We aligned our technical approach with what we knew to be true about real-world marketing dynamics. 

True marketing measurement requires distinguishing between mere correlations and genuine causal relationships. 

• Correlation identifies variables that tend to move together: when A increases, B also tends to increase (or decrease).
• Causation identifies variables where one directly influences the other: changes in A cause changes in B.

Without this distinction, marketers risk making decisions based on illusory patterns rather than actual drivers of performance. 

We weren’t satisfied with correlation. We knew our platform needed to get to the root of cause and effect related to the marketing spend of our clients. When brands rely on platforms using open-source MMMs, they risk systematically misallocating marketing investments—over-investing in campaigns that merely correlate with success while under-investing in efforts that actually make it happen.

Of course, we aren’t the first or only measurement provider to preach the correlation versus causation distinction. As incrementality testing has grown in popularity, causality has been a common message. While incrementality testing does its best to follow the scientific principles of random control testing (RCT), the marketing landscape is rife with potential variables. Creating a truly controlled environment is incredibly difficult to accomplish. 

Recently, many incrementality testing providers have added MMM solutions to their product offerings (and many MMM providers are tacking on testing), suggesting that testing results can validate, calibrate, and bring causality to marketing mix models. You see where I’m going with this? Once you understand the limitations of traditional MMMs and incrementality testing, you can see why getting them from the same source would result in a “grading their own homework” situation—and potentially disastrous conclusions. 

We have a lot more helpful content in the works discussing how incrementality testing and other attribution tools fit into the measurement ecosystem and how to make sure they are additive, not degrading your MMM—but the purpose of this post is simply to explain why existing open-source models, however much improved upon, will never be able to meet the demands of marketing in the future.

Prescient’s unique methodology provides reliable, consistent optimization guidance that remains valid when conditions change, accurately identifies which marketing activities truly drive performance, and enables granular campaign-level optimization—when and where marketers actually need to make decisions.