How marketing mix modeling measures spend without tracking individual users

Marketing mix modeling, or MMM, estimates how much each channel contributed to sales by comparing spend and results over time.

Google's open-source Meridian model is one public example of how the method is set up. Its documentation shows what data a team needs, how long a history is required and where experiments fit in.

What a marketing mix model does with your data

An MMM takes historical sales and marketing data across channels and regions and estimates the incremental effect of each channel. Google's description of Meridian says it uses aggregate data and Bayesian methods, and that it updates its estimates as new data arrives.

Google states that Meridian does not need individual identifiers or cookies.

Google describes Meridian as measuring online and offline channels together, including TV ad performance and Search.

How much history and how many regions it needs

Google's overview says Meridian needs at least two years of historical data, and that users begin by defining their KPIs. The detailed documentation says it prescribes no single minimum ratio of data points to parameters. Its most reliable test is to run the model and check how wide the credible intervals are, meaning the range of values the model considers plausible for each estimate.

The documentation sizes data as regions multiplied by time periods, and compares that count with the number of parameters the model must estimate. It warns that regions and time periods are not independent, so a count of data points is not the same as a count of independent observations.

For a national model, the worked example has 26 parameters: 12 channels, 6 controls and 8 knots. Two years of weekly data gives 104 points, about 4 per parameter, which the documentation treats as too few.

The fixes it lists are combining or dropping low-spend channels, using fewer knots (the default is 1) and removing controls that are not true confounders. Adding data also helps. A trimmed example with 3 channels, 5 controls, 2 knots and three years of weekly data gives 156 points for 10 parameters, about 15 each. The documentation says that might yield some directional insight.

Weak variation in spend is a separate problem. The documentation notes that insufficient variation in media spend also weakens national models.

Why regional data tightens the estimates

A model built on regions has far more observations than a national one. The documentation's geo example uses 105 regions and 156 weeks, which is 16,380 data points, with 12 media channels, 6 controls and 100 knots.

Whether that is enough depends on how much the regions share information. Counted as if every region were estimated separately, the model has about 2,094 parameters, or about 8 points each. Counted as if all regions behaved identically, it has about 222 parameters, or about 74 points each.

The real position falls between those two, and the documentation says it depends on variance parameters that are only known after the model is fitted. For that reason it sets no single minimum ratio.

When data is scarce, the documentation suggests combining or dropping channels, or using a tighter prior on those variance parameters so the regions share more information.

Where experiments calibrate the model

Meridian can take outside experiment results as priors.

Google's Meridian page describes a tool called GeoX that runs geo experiments and can convert and adjust the results into priors for Meridian. Google's overview also says the model can incorporate incrementality experiment results.

For paid search, the Meridian page says search query volume data can be included to control for organic demand.

The model can also use reach and frequency rather than impressions alone. Google's overview says Meridian accounts for reach and frequency, not just impressions.

Meridian measures marketing effectiveness using aggregate data and Bayesian methods.

What the outputs are for, and what they cannot do

Google says Meridian produces budget optimisation recommendations, which refresh as data is added. Its overview contrasts it with traditional MMMs on optimal frequency and regional nuances.

Google says most teams need a data scientist or a third-party partner to get the most from the model. For Google's own channels, the Meridian page says the MMM Data Platform can help gather campaign data across Google Ads and Display & Video 360, and its MMM Data team will follow up with access details within 3 business days of a submission.

The documentation generally doesn't recommend campaign-level data. The model works at the channel level, and hard campaign starts and stops risk losing the memory of the Adstock. For detail below channel level, it points to data-driven multi-touch attribution for digital channels.

Google names Pandora, a brand present in more than 70 countries, as one of the first users of Meridian.

What to check before commissioning one

The documentation points to a short list of checks. The first is history: at least two years of data, and three years of weekly data for a national model with few channels. The second is regional breakdown. Google suggests adding geo granularity to the data and using a geo model.

The third is spend variation. The documentation notes that insufficient variation in media spend adversely impacts national models. The fourth is experiment history. Past geo tests can be converted into priors with GeoX.

The last is scope. Combining small channels before modeling is one of the remedies the documentation gives for too few data points. Reading the credible intervals once the model has run is the check it recommends on whether the data was enough.