Home » Connecting Data with Google Meridian for MMM using Windsor.ai [Tutorial]

Connecting Data with Google Meridian for MMM using Windsor.ai [Tutorial]

Connecting data with Google Meridian for MMM using Windsor.ai

Google Meridian is an open-source library developed by Google to implement Bayesian Marketing Mix Modeling (MMM). It helps marketers understand the effectiveness of their advertising channels by estimating how different media inputs (like ad spend or impressions) impact a key business metric (KPI), such as conversions or revenue. 

Meridian utilizes probabilistic modeling to provide insights into return on investment (ROI), channel attribution, and forecasted outcomes, enabling businesses to make data-driven decisions about budget allocation.

Using Windsor.ai data integration platform allows you to automatically fetch, transform, and prepare your marketing data for use with Meridian. With just a few API calls, Windsor.ai connectors aggregate data from various ad platforms, enabling seamless integration with the MMM pipeline in Meridian.

How to use Windsor.ai data with Google Meridian for MMM

This tutorial explains how to connect Windsor.ai’s marketing data with Google’s Meridian library to build a Marketing Mix Model (MMM). 

You will learn how to:

  • Collect and fetch marketing data from Windsor.ai
  • Prepare the data in the proper format for Meridian
  • Configure the Meridian model
  • Fit the model to analyze your marketing performance
  • Generate a summary report

Prerequisites:

  • Python 3.11+
  • Libraries: tensorflow, tensorflow_probability, psutil, pandas, requests, and meridian
  • A Windsor.ai API URL with key, date_preset, and fields
  • Knowledge of Google Meridian MMM

Step 1: Fetch data from Windsor.ai

You can get the Windsor.ai API URL for your data from the dashboard and use it to fetch daily performance metrics across your marketing channels. 

These metrics typically include:

  • Date
  • Clicks
  • Spend
  • Impressions
  • Conversions

Important:

Aggregate your data by date (daily frequency), and fill any missing values with zero to ensure compatibility with Meridian.

In our example, we use data from Google, Facebook, Bing, and Reddit Ads, and the key performance indicator (KPI) we aim to model is Conversions to Spend.

Step 2: Set up Meridian input format

To use your Windsor data with Meridian, you need to do the following things:

  • Define how the columns in your dataset map to what Meridian expects (e.g., date, KPI, media spend, etc.).
  • Indicate the media channels and their corresponding spend columns.
  • Load the data into Meridian using this script:
coord_to_columns = load.CoordToColumns(
   time="date",            # Timestamp column
   geo="geo",              # Optional: region/location (if available, else keep as None or a dummy column)
   controls=[],            # Any external control variables (economic indicators, etc.)
   kpi="conversions",      # Your Key Performance Indicator (target variable for prediction)
   revenue_per_kpi=None,   # Optional: Revenue generated per conversion (for revenue models)
   media=["facebook_impressions", "google_impressions", "reddit_impressions", "bing_impressions"],  # Media signals (e.g., impressions)
   media_spend=["facebook_spend", "google_spend", "reddit_spend", "bing_spend"],  # Spend per media channel
)

# Maps your impression column to its corresponding media channel (important for attribution and budgeting)
correct_media_to_channel = {
   "google_impressions": "Google_Ads",  # replace with your media channel
   "facebook_impressions": "Facebook_Ads",
   "reddit_impressions": "Reddit_Ads",
   "bing_impressions": "Bing_Ads"
}

# Maps spend column to the correct media channel
correct_media_spend_to_channel = {
   "google_spend": "Google_Ads",  # replace with your media channel
   "facebook_spend": "Facebook_Ads",
   "reddit_spend": "Reddit_Ads",
   "bing_spend": "Bing_Ads"
}

# -------------------------------
# STEP 3: Load Data into Meridian
# -------------------------------
# This loads and prepares the dataset based on the mappings and parameters defined above.
loader = load.CsvDataLoader(
   csv_path=csv_data,                              # CSV data from Windsor or your source
   kpi_type='non_revenue',                         # Use 'revenue' if you provide revenue_per_kpi
   coord_to_columns=coord_to_columns,
   media_to_channel=correct_media_to_channel,
   media_spend_to_channel=correct_media_spend_to_channel,
)
data = loader.load()

Important note:

For more optimized results:

  • Consider adding control variables like:
    • Seasonality and time effects
    • Economic and market indicators
    • Competitor & market activity
    • Product or price changes
    • Organic demand proxies
  • Consider adding organic media and non-media treatments (promo, etc.)

Step 3: Configure model priors

Set prior expectations for ROI using a LogNormal distribution.

For example, if the expected ROI is ~0.5 conversions per $1 spent:

# ROI prior: Used to inform the model of expected returns on ad spend.
# LogNormal is used since ROI is strictly positive.
# roi_mu: ln(ROI). If you expect ~0.5 conversions per $1, ln(0.5) ≈ -0.6931
# Assign roi_mu and roi_sigma values according to your dataset
roi_mu = -0.6931
roi_sigma = 0.5  # Uncertainty/spread in the ROI prior

prior = prior_distribution.PriorDistribution(
   roi_m=tfp.distributions.LogNormal(roi_mu, roi_sigma, name=constants.ROI_M)
)

# Wrap the prior into the model specification
model_spec = spec.ModelSpec(prior=prior)

# This creates a Meridian model object using the input data and model specifications
mmm = model.Meridian(input_data=data, model_spec=model_spec)

# First, sample from the prior distribution (before seeing data). Good for understanding priors.
# Adjust these values according to your dataset
mmm.sample_prior(100)  # 100 samples is enough for exploration

# Then, sample from the posterior (after observing the data)
# Smaller values (n_chains=2, n_adapt/burnin=100) make it faster for small data or experimentation.
mmm.sample_posterior(
   n_chains=2,    # Number of independent chains
   n_adapt=100,   # Adaptation steps for tuning the sampler
   n_burnin=100,  # Number of warm-up iterations (burn-in)
   n_keep=200,    # Number of posterior samples to keep
   seed=1         # Random seed for reproducibility
)

Important note:

  • If your baseline is negative. It means the model is attributing all KPI to media channels. You can solve this by:
    • Tighten ROI priors for paid channels – Lower and narrow ROI priors to reduce over-attribution.
    • Add control variables – Include seasonality, organic demand, and economic factors to explain non-media effects.

Step 4: Summarize and export results

After fitting the model, you can generate a full performance summary that covers:

  • Channel effectiveness
  • ROI estimates
  • Attribution breakdowns
# Generate an HTML report with parameter summaries, media channel attribution, ROI, and more.
# The date arrange should be inside the used dataset
mmm_summarizer = summarizer.Summarizer(mmm)
file_path = './'                     # Save location for summary report
start_date = '2024-04-24'            # Analysis start date (adjust based on your data)
end_date = '2025-04-23'              # Analysis end date

# Export the model summary to an HTML file
mmm_summarizer.output_model_results_summary('summary_output.html', file_path, start_date, end_date)

The summary helps you understand how well each marketing channel is performing and whether it justifies a major investment.

Conclusion

That’s it with using Windsor.ai data connectors with Google Meridian for marketing mix modeling.

Ready to optimize your marketing spend? Get started with Windsor.ai to automatically connect your data to Google Meridian and unlock valuable insights for smarter decisions!

Explore the source code in our GitHub repository to try it on your own: https://github.com/windsor-ai/WindsorMeridian.