Building Marketing Mix Models with Google Pixel Data Using Meridian

The primary business objective was to use a Marketing Mix Model (MMM) to better measure the causal effect of various marketing activities (like ad spend across different channels) and search trends (branded and generic queries) on Google Pixel sales performance. Key performance indicators (KPIs) included overall activations (number of phones sold and activated) and potentially revenue per activation.

Google sought insights into the effectiveness of their marketing spend and how different channels contribute to Pixel activations. The focus was on causal interpretation rather than purely predictive accuracy.

The approach involved two main phases: Exploratory Data Analysis (EDA) and Modeling. The team analyzed 2-3 years of anonymized, indexed daily data for the US market, covering sales (activations), marketing spend and impressions across 20+ channels, and engagement data like Google search queries.

Initial EDA focused on understanding trends in ad spend, activations over time by region, the relationship between marketing cost and activations, and the influence of non-marketing events (like Pixel release dates or sales events) on search interest and activations.

For modeling, the team utilized Meridian, an open-source MMM platform developed by Google, chosen for its capabilities in causal analysis. The process evolved from establishing a baseline model to refining it through hyperparameter tuning, focusing particularly on the number of ‘knots’ (see more here), as well as ROI priors (see more here) to improve model fit and granularity without overfitting. The team also explored using holdout data to test model performance on unseen data and experimented with expanding regions to improve generalization.

The project delivered insights into the effectiveness of different marketing channels on Google Pixel activations by applying and refining MMMs using the Meridian platform. Exploratory analysis revealed moderate positive correlations between overall marketing spend and activations, although some cost spikes didn’t yield proportional activation increases, suggesting external factors matter. Pixel launch events were correlated with activation spikes.

Through model tuning, particularly adjusting the number of knots (bi-weekly or 52 knots yielded good results) and experimenting with priors (μ and σ), the team improved the model’s fit (R² improved significantly from baseline) while managing overfitting. Testing on holdout data revealed challenges in predicting the most recent periods, indicating areas for further refinement. Combining data from multiple regions (e.g., LA and New York) helped create a more generalized model with better performance on test data. The final model provided channel-level contribution analysis.

Possible next steps identified include applying the tuned Meridian model to newer Pixel datasets or other Google products, further tuning model parameters (e.g., different prior distributions), and incorporating additional external data sources like regional demographics (income, education) to potentially enhance the model’s explanatory power.

The project provides Google with an understanding of how their marketing investments across various channels translate into Pixel activations.

These insights can inform Google’s future marketing budget allocation decisions, allowing for more efficient spending by focusing on channels demonstrated to be more effective drivers of sales. The methodology and learnings could potentially be scaled or adapted for other products or regions within Google, improving marketing measurement practices more broadly. While direct ecosystem or economy-wide impacts are harder to quantify from this specific project, optimizing marketing spend for a major product like the Google Pixel contributes to efficient market operation and resource allocation within the competitive consumer electronics sector.