First Impressions – Meridian: Google’s MMM Software is finally released

Google’s long-awaited Marketing Mix Modelling (MMM) solution, Meridian, was officially released overnight. 

Google had initially entered this space with the LightweightMMM package, launched around the same time as Meta’s Robyn. The consensus was that while LightweightMMM incorporated advanced modelling techniques, it lacked the feature richness and user-friendliness of Robyn. One advantage it did have was its Python-based implementation, making it more accessible to most people than Robyn’s R-based framework. However, it was clear that Google had not fully committed to the project, treating it as an experimental initiative rather than a robust industry solution. Despite its limitations, LightweightMMM showed promise, particularly in adopting Bayesian and causal modelling—approaches widely regarded as superior in marketing mix modelling. 

In March 2024, Google announced Meridian and shortly thereafter deprecated LightweightMMM. This signalled a potential leap forward: an evolution of LightweightMMM into a more polished, feature-complete solution capable of competing with Robyn, PyMC-Marketing, and other leading MMM tools. But does Meridian live up to the expectations? 

Here's a review from Jaywing's Matt Triggs:

The Essentials 

At a fundamental level, Meridian offers everything expected from a modern MMM: multi-geography modelling, adstock effects, decay and lag structures, and budget optimisation. These are now considered standard features. However, Meridian goes beyond this baseline, integrating advanced techniques that could set a new industry standard. 

Google’s Thought Leadership in MMM 

Google’s marketing scientists have been at the forefront of MMM methodology. Two seminal papers from 2017 [1,2] laid the groundwork for Bayesian approaches to marketing mix modelling, and subsequent research in 2021 and 2023 expanded these concepts to include trend and seasonality adjustments, as well as reach and frequency modelling using time-varying coefficients. 

Notably, the PyMC-Marketing team built upon this research to develop their own Bayesian MMM framework. I have long considered PyMC-Marketing’s approach to be among the best available: it embraces Bayesian and causal inference principles, includes a robust set of features, and, while complex, offers significant flexibility for advanced users. 

Closing the Gaps 

One of the major shortcomings of LightweightMMM, as highlighted in the PyMC-Marketing comparison table below, was its lack of time-varying coefficients and model calibration through lift testing. Meridian directly addresses these gaps, matching PyMC-Marketing’s capabilities and, in some areas, surpassing them. 

Model Calibration and Validation 

Calibration is a crucial element of modern MMM development. Rather than using controlled experiments (e.g., lift tests) solely for validation, Meridian enables their integration into the modelling process itself. In my view, this is a much better use of a very valuable data source. Additionally, the ability to generate out-of-sample predictions is a much-needed feature that enhances model evaluation and reliability. 

Specification of Model Priors 

The ability to specify model priors is another handy addition, particularly for advanced users. The hierarchical structure of Meridian’s modelling framework allows for both national- and geo-level models, incorporating control variables where necessary. However, one notable limitation is the absence of non-geographic segmentation (e.g., multiple product categories within a single model). While not a dealbreaker, this is a feature I would like to see in future iterations. 

For newer users, the option to specify priors based on ROI rather than direct beta coefficients offers (potentially) a more intuitive starting point. While experienced practitioners may not need this functionality, it could be a useful learning tool for those less familiar with Bayesian approaches. 

Reach & Frequency Modelling 

One of Meridian’s standout features is its inclusion of reach and frequency estimation. Traditionally, MMMs have struggled to capture these elements, making it difficult for econometricians to respond to stakeholder questions on the topic. The fact that Meridian incorporates R&F modelling directly into the framework is a major advancement which, to my knowledge, is unmatched in other tools. 

Documentation & Accessibility 

Meridian’s documentation is among the most thorough I have seen in an open-source MMM solution. Unlike many marketing analytics tools, which tend to abstract over mathematical complexities, Meridian’s documentation does not shy away from the details. This suggests that the software is designed with mathematicians and economists in mind rather than generalist marketing analysts. While this level of rigour is much needed in the MMM space, it may present a steep learning curve for some users. 

Conclusion 

Meridian is set to disrupt the MMM landscape significantly. By integrating cutting-edge research with Google’s extensive resources, Meridian has the potential to become the dominant open-source MMM solution. 

For anyone considering starting a Robyn project, now may be the time to pause and evaluate alternatives. While I continue to be impressed by PyMC-Marketing’s approach and would still recommend it over Robyn, Google’s entry into the space changes the equation. At first glance, Meridian appears to be the strongest free and open-source option available. 

Of course, as more econometricians and data scientists experiment with the tool in the coming months, its limitations and potential issues will become clearer. However, based on first impressions, it is evident that Google has produced a formidable competitor, one that may well set a new industry benchmark.