Meridian is an open-source Marketing Mix Modeling (MMM) package developed by Google, publicly announced in 2024 March. Meridian enables users to conduct Bayesian Marketing Mix Modeling, and is expected to replace Google's first Bayesian Marketing Mix Modeling package, Lightweight MMM. Below is a screenshot from Meridian's landing page:

With Meridian, skilled analysts and data scientists can create Marketing Mix Models with ready-made modeling tools:

Meridian makes it faster for data scientists to build Bayesian Marketing Mix Models, compared to developing MMM features, such as adstock or diminishing returns, themselves.

While Meridian builds on the achievements of Lightweight MMM, it also incorporates the long-standing research work done by Google on Marketing Mix Models. We at Sellforte are big fans of Google's work on the topic, and have incorporated many of the ideas from Google's papers in our work as well. Examples of Google's research on Marketing Mix Modeling below:

Meridian is built for:

Meridian's key features include:

Similar to Lightweight MMM, Meridian offers marketing mix modeling based on the Bayesian framework, which has become popular in the marketing mix modeling community. It should be noted though, that it is not the only approach. For example, Robyn by Meta does not use Bayesian approach.

Incorporating external knowledge in the form of priors. Bayesian modeling is the gold standard for marketing mix modeling for many reasons, of which one is the ability to provide external information to the model in the form of priors. Optimally priors are set to be as informative as possible, leveraging lift tests and other information about the behaviour of the medium.

Media saturation and lagged effects. Being a marketing mix modeling package, Meridian incorporates the industry standard MMM features, such as adstock and diminishing returns, to account for media saturation and lagged effects.

Evaluating and reporting model goodness of fit. Meridian provides tools for the data scientist to review model fit statistics to understand statistical quality of the model, as well as to evaluate and compare different models against each other.

Media budget optimization and estimation using what-if scenarios. Meridian enables users to get optimal budget allocations, based on the modeling results, as well as how media ROI behaves in different scenarios.

Based on Google's research paper Geo-level Bayesian Hierarchical Media Mix Modeling, Meridian enables geo-level modeling. Instead of having one national model, Meridian enables splitting the model into geographies, and modeling them together in a hierarchy. This enables getting more granular results (instead of only national results), as wells as getting more robust results as the models are connected.

User can input reach and frequency metrics as additional inputs to the model. The objective is to reach more reliable results and deeper insights, as proposed by this article: Bayesian Hierarchical Media Mix Model Incorporating Reach and Frequency Data.

Meridian provides the option to use Google Query Volume as a control variable in the model.

Meridian has supporting documentation available, as illustrated below:

Meridian has following limitations:

"The Meridian model is focused only at channel-level." Source

Campaign-level optimization is not available in Meridian. This is a critical shortcoming, because based on our research, optimizing spend level by campaign is the largest optimization lever in MMM: Campaign spend optimization enables eComs drive 2.9% more sales. Additionally, practical marketing execution typically happens in campaigns.

No data connectors. As Meridian is not designed to be an end-to-end MMM platform, some features that are available in many MMM SaaS platforms are not available. One example of this is the lack of data connectors. Getting the data together for modeling can be a huge undertaking if done manually. In larger MMM projects, time can be spent on data gathering than on the modeling. With Meridian, the user is responsible for orchestrating the data gathering. When setting up Meridian, our recommendation is to use data connector companies, such as Supermetrics, for the effort, as they also enable automated data updates in the future.

No data processing tools. After the dataset for MMM has been collected, the next big hurdle is to process the data to be ready for MMM. Raw data coming from ad platforms and sales systems is rarely usable as such if one wants to get the maximum benefit from the data. Data processing includes for example setting up media and sales data hierarchies, campaign mapping, data transformations, and enriching the data. For example, getting campaign-level MMM results requires advanced data processing.

Meridian is built for data scientists and skilled analysts, so it is no surprise that the modeling outputs are static plots, which the user can print out with python commands. This can be sufficient for the data scientist building the model. However, making the analysis accessible for the stakeholders making decisions based on the analysis, the marking team, can be challenging with this approach. One option is to make PowerPoint decks based on the analysis. Another option is to build a connection to a dashboarding tool, such as Tableau.

Meridian does not currently have an interactive tool for media budget optimization, and the optimization scenarios are built with python. Again, this can be sufficient for the data scientist, but makes scenario testing more challenging for the marketing team, as it takes time for the data scientist to translate marketing team's questions to code, and then communicate the results back to the marketing team, e.g. in a PowerPoint deck.

Meridian is a build-it-yourself modeling library, which means that to operate it in production, it requires a team: Data Engineers, Data Scientists, people who have strong business / marketing understanding to validate the results etc. This can make Meridian surprisingly expensive to onboard and maintain, even if the modeling library itself is free.