Marketing Mix Modelling (MMM)

What is Marketing Mix Modelling and the extent to which it resolves the attribution problem.

What is MMM

Marketing mix modelling (MMM) is a statistical model that quantifies the incremental impact of marketing activities on the number of conversion events* (e.g. sales). MMM models are effective at providing a holistic understanding of attribution across different categories/channels of marketing activity.

*A conversion event is any action a prospect/customer takes which generates revenue or increases the probability of generating revenue examples include a website visit, subscription, signup and checkout.

How MMM works in simple terms

Using engagement to estimate sales

MMM begins with establishing a dependent variable which is typically the number of conversion events (in this example we'll use sales). The number of sales fluctuates across time due to a number of factors such as:

  • Seasonality (e.g. demand is higher in summer than winter)
  • Discounts
  • Engagement with TV advertising
  • Engagement with digital performance marketing

The first two factors are examples of control variables, they are factors exogenous to marketing. Where as, the final two factors are examples of independent variables, they are determined by alterations to the marketing strategy. MMM models are trained by predicting the value of the dependent variable (sales) in each period (typically a week) based upon the values of the control variables and independent variables. The MMM model attempts to minimise the difference between the predicted number of sales and the actual number of sales in each period. In making these predictions each dependent variable and control variable has its associated coefficient. The coefficient for an independent variable can be interpretated as an estimation of the incremental impact of the channel on sales.

Interaction Effects

Marketing activities do not work in isolation, companies typically run a multitude of concurrent marketing activities. The effectiveness of a particular marketing activity is dependent on the performance of other marketing activities. This interdependence comes in the form of two effects:

  • Halo effect - running the two activities simultaneously results in more sales than if they were run in isolation e.g. brand awareness adverts have a significant positive impact on the conversion rate of search adverts
  • Cannibal effect - running the two activities simultaneously results in fewer sales than if they were run in isolation e.g. branded search advertising is capturing users who would have visited the site through an organic result

MMM models are able to capture both of these effects by introducing an interaction variable which is just another independent variable. When the model is run, interaction variables get an associated coefficient which can be interpreted as the halo/cannibal effect.


The impact of marketing activity isn't instantaneous and the impact of a marketing activity can persist beyond the advert running. Adstocks represent how engagement with an activity impacts sales across time. Adstocks can be described by two features: a lag which is the average time between engagement and conversions, a half-life which is how quickly the impact of engagement on conversions diminishes over time. To understand adstocks it is best to consider comparing a Top of the funnel Brand Awareness TV Advert and a Bottom of the funnel Google Shopping advert.

  • Awareness Advert: A large pool of prospects see the advert, of which few are familiar with brand and even fewer are considering purchasing, a very small proportion of the pool purchase and these purchases are spread across an extended timeframe, the adstock has a very long lag. The advert is particularly memorable meaning the awareness and consideration of the brand persists for an extended period, the adstock has a very long half-life.
  • Shopping Advert: A small pool of high intent prospects see the advert, a large proportion of this pool convert mostly and this conversion typically takes place within the same session, the adstock has a very short lag. The advert is not particularly memorable and the purchasing intent of buyers diminishes significantly over time (bought alternative or forgotten), the adstock has a very short half-life.

Adstocks explain why the impact of an advertising activity on sales as reported by the MMM might not perfectly align with the engagement. MMM models must use representative adstocks to accurately attribute sales.