Marketing Mix Modeling: Attribution Beyond the Click

Attribution is the tool that lets a media manager measure the return on ad spend. The most established method depends on clicks. And there are entire channels where clicks do not exist.

In this article, I explain how Marketing Mix Models (MMM) attack that blind spot, show the math behind the method and share what we saw when applying Meridian, Google's open-source MMM, in real Marktech projects. Along the way, a simulation with numbers to assess whether the model delivers what it promises.

• What click-based attribution can't see
• How a Marketing Mix Model works
• Meridian: Google's open-source MMM
• What the simulation showed
• The caveats with real data
• When MMM is worth it

What click-based attribution can't see

The most established attribution method is click tagging. It consists of embedding data in the ads' URLs according to how the media channels are configured. When a conversion happens, those parameters make it possible to identify the origin: channel, campaign, creative.

It is an effective approach. It is how media operations usually decide, for example, how to split the budget between Google and Meta.

But it has an important limitation: how do you measure the contribution of channels that do not generate clicks? Offline media is the classic case. Radio, TV and out-of-home influence conversions, but they carry no tagged URL. For click-based attribution, these channels simply do not exist.

How a Marketing Mix Model works

MMMs (Marketing Mix Models) are an alternative to that traditional methodology. They take a less granular view: instead of tracking each conversion individually, they correlate aggregate spend, total conversions and control variables over time, using regression techniques to estimate each channel's contribution.

In short, conversions on a given day t are modeled as follows:

y_t = β₀ + f(x_t; β) + ε_t

Where:

• x_t represents the vector of spend (and other variables potentially related to conversions, such as weather, organic searches, among others) for day t;
• ε_t corresponds to the noise;
• β₀ and β are the model's parameters.

The goal of the MMM framework is to fit the parameters β₀ and β from the history of observed data (x_t and y_t), seeking the best possible representation of the relationship between spend and results.

Note what the model does not need: no clicks, no cookies, no individual journeys. Just historical series of spend, conversions and context. That is what opens the door to including offline channels in the math.

Meridian: Google's open-source MMM

In 2025, Google released an open-source MMM, Meridian. At Marktech, we have already applied this framework in a few real projects and came away with mixed impressions of its practical applicability: the promise holds up under controlled conditions, but the method demands data conditions that not every operation has today. I will come back to that below.

A question that always comes up: if Meridian is Google's, wouldn't it favor Google's own channels? There is no way it could. You are the one who defines the labels of each channel sent as input. Meridian has no way of knowing which channels belong to Google in order to, hypothetically, favor itself in the process.

What the simulation showed

To illustrate, I ran a simulation with synthetic data, representing a plausible scenario of marketing spend and conversions: three channels (Google, Meta and programmatic media), with daily spend. The advantage of synthetic data is that each channel's true attribution is known. You can compare what the model estimates against the right answer.

In the simulated scenario, Google accounted for 60% of conversions, and Meridian estimated 64%. Meta accounted for 33%, and the estimate was 25%. Programmatic accounted for 7%, against an estimated 11%.

In other words: even without direct attribution available, the MMM derived attribution percentages considerably close to the true values, showing it was able to deduce the channels' contributions satisfactorily by looking only at the aggregates.

The caveats with real data

Synthetic data is a lab test. When dealing with real data, a few additional points need to be considered.

First: the model may require a significant amount of data to be trained well. A short historical series produces unstable estimates, and drawing conclusions from too little data is a trap from the same family as the false positives in A/B tests ended too early.

Second: it can be quite sensitive to how much spend varies over time. If spend is practically constant from day to day, the model will struggle to correctly infer the lead generation function. The intuition is simple: regression learns from variation. If the budget never changes, there is no information about what would happen if it changed.

When MMM is worth it

With proper data structuring, MMMs can become a useful tool to broaden the view of marketing spend, especially in the media channels where traditional click-based attribution cannot see.

In practice, MMM does not replace click tagging: it complements it. The click remains the best lens for the day to day of digital channels, and the aggregate model comes in to see the whole, offline included. With both views on the table, the decision of how to allocate the budget across channels no longer depends on a single instrument.

Want to assess whether an MMM makes sense for your operation? Talk to us.