Around this time last year, I wrote a blog to briefly introduce the concept of Media Mix Modelling, or Marketing Mix Modelling, also known as MMM. A year has gone by, and given how the digital landscape has drastically evolved versus a year ago, I felt the need to write another blog about this powerful tool and share with you what’s changed in the world of MMM. If you have no idea about MMM or are interested in knowing more about machine learning in general, please check out my last blog before reading this one!
So… what’s new?
The biggest game-changer that revolutionized the whole of MMM is the official release of Meridian in January, a robust open-source model developed by Google. Technically, it was initially launched in March 2024, but was exclusive to a certain group of digital marketers and data scientists to evaluate how their marketing efforts impact sales, or any other key business metrics.
What I like about Meridian
First and foremost, it encourages users to input multi-market data, analyze the underlying regional trend, and identify a national pattern from multiple geographies. Canada is a huge country, and provinces could be vastly different for businesses. It is inevitable for national companies to develop various tactics accordingly. From both a strategic and operational perspective, some markets are simply more critical, and hence more attention as well as marketing efforts. Running a multi-geo model is more sophisticated than running separate models for each geo or a single, undifferentiated or aggregated national model, because it provides a larger volume of data for machine learning.
Another wonderful feature would be the inclusion of the population in a multi-geo model. Why does population matter? Let’s think about the ratio and density here. 1 million impressions in a province of 500,000 residents, i.e. 2 impressions per person on average, is a completely different level of exposure compared to 1 million impressions in a province of 10 million people (0.1 impressions per person). By factoring in the population, instead of having the raw media spend and impressions, the model is interpreting metrics into units like “Spend per 1,000 people” or “Impressions per capita.” Therefore, it compares the intensity of your marketing efforts across regions on a relatively fair basis. Apart from media input, population size across markets is also potentially skewing the KPI metrics. For instance, sales in The Maritimes are very likely to be lower than in Ontario or British Columbia, because of the market size. A province with a larger population naturally has a larger potential customer base and, all else being equal, higher baseline sales. If the model doesn’t have visibility on the population, it might incorrectly attribute the higher sales in a large province to the media.
Google also incorporated two other important elements, namely Google Query Volume (GQV) and Reach and Frequency (R&F) Data into Meridian. Google’s ecosystem is its biggest competitive edge, which allows marketers to understand customer behaviour in the video platform and search engine. GQV essentially serves as a proxy for organic consumer interest or product demand. By including it in your model, you can help separate organic search behavior from the incremental lift of the paid search campaigns. This reduces the risk of misattributing sales and helps you get a more accurate ROI for your paid search efforts. R&F is another interesting add-on that makes the model more robust. In traditional MMM, it solely relies on the impression metric, which does not account for how many unique people saw an ad or how often they saw it. However, it is obvious that it’s different between a campaign that reached 10 people once and a campaign that reached the same person 10 times. With the R&F data, the model can give you a more accurate picture of video and display effectiveness on YouTube.
Lastly, the built-in visualization module of Meridian is informative and amazing, which saves my fellow data nerds some time on visualizing the model outputs. Sometimes, it is quite time-consuming to transform the numbers into graphs or charts because the insights behind the data are what matter.
Final Thoughts
Similar to every data science project or machine learning model, data quality is always the key to success. Please make sure the dataset is clean, complete, and consistent. The input features will determine the model outcome; that’s why we always say “Garbage In, Garbage Out”. Now, it is time for me to get back to my Meridian project, but please do not hesitate to reach out to us if you are interested in learning more!