To gain more clarity on marketing performance, many brands are increasingly investing in first-party data and exploring alternative measurement frameworks. It differs significantly from cookie-based click attribution, in that attribution models attempt to take a specific sale and assign credit for it to specific clicks or web visits. Meanwhile, media mix modelling (MMM) takes the total sales and total investments, looking for statistical correlations between them.
This is why attribution models are sometimes referred to as bottom-up (starts from the click and then added together) and MMM-style analysis as top-down (starts from the total sales and is then portioned out).
How does MMM work?
When we talk about ‘statistical correlations’, it can often sound like there is something horrifically complex happening. However, the basic logic of MMM is pretty simple: if you spend more, you should get more back.
Take this ad account in isolation as an example, with weekly spend and revenue plotted out on a scatter chart. There is a clear trend that as more is spent, more revenue is driven. This line of best fit, or ‘linear regression’, is of the format: y = mx+c.
In this case we could also write it as:
The ‘Incremental ROI’ represents how much additional revenue each £ of investment generates and the ‘Baseline Rev’ represents the revenue that would be achieved even if you spent nothing.
In essence, this is the world’s simplest MMM. If you’re investing in a single channel, you can export the data, put it in a spreadsheet and – hey presto – you’re done, you’ve modelled the impact of your investment.
Now, the majority of businesses are actually investing in multiple channels and want to understand the relative impact that each channel is having in order to reallocate funds. Due to the interconnectedness of marketing, we can’t just do the same thing twice. What if we ramped up spend in our brand-awareness focused channels and it drove more own-brand searches?
The other important thing to remember here is that we’re modelling the effect of the spend on the overall revenue. The whole point of MMM is to find out how much influence each channel has had on sales.
So, now we have two spends and one revenue, we can plot this as a 3D graph. Instead of a single line of best fit, we have a plane of best fit:
This outputs an equation that says to get our modelled weekly revenue we follow: 
Implying a 3:1 Google Ads return on investment (ROI), a 10:1 Facebook Ads ROI and a £4.5k per week input from non-paid sources.
Here, we aren’t limited to just two channels. We can add as many extra channels in as we like and keep making multiple linear regression analyses on it; we just run out of the ability to plot it visually.
However, the output would still come out something like:
Improving the model
The above steps are how the model would work, in theory, if there were no external factors impacting your business.
Of course, we know there are many – chief of which is usually seasonality. Even if you’re not a traditionally seasonal business, there are going to be periods where outside factors are going to affect your investment and return, such as Christmas or other public holidays. In our example above, both platforms reduced spending heavily over Christmas; at the same time, revenue also dropped. This needs to be controlled for in the model either with a seasonality series alongside the spend series, or by excluding outlier data points.
Other factors you might want to consider in the model are large exhibition events, notable PR coverage; anything that might materially impact the bottom line of your business should be included within the model for the best accuracy.
Finally, there are other statistical measures you might want to consider such as the level of co-dependency between variables or the amount of uncertainty in the regression.
What’s the point of MMM vs Data Driven Attribution Models?
Data-driven attribution models (DDA) are great as, because of the bottom-up approach, they can give all the granularity of detail that you need. If you want to understand the performance of a keyword or an ad, then you need a DDA model to do so. However, they have some flaws.
If you were to take a web visit model such as via Google Analytics or some equivalent platform, then it will successfully measure and attribute out the web visits. So what’s the flaw? It will not be able to include in that model the effects of users viewing any ads. So all your brand awareness campaigns will always look worthless.
But there’s view-through attribution in ad platforms I hear you say? And indeed there is, it’s great. However, you then have the problem of no centralised location doing the attribution. You can’t add up what Google says you’ve made and what Meta says you’ve made, or else you’ll end up claiming you’ve made more than you have.
So, a strategic measurement system like MMM is what you use to look at big-picture, channel vs channel performance. It seeks to answer one question and one question only – when you invest money in an area, does it increase overall sales and by how much?
Summary
MMMs are not here to replace your DDA models that you use in day-to-day optimisation of your marketing efforts. They need large amounts of data to work, stretching over at least a year, ideally two-three. As such they cannot help you make rapid granular decisions.
However, when it comes to making your big strategic decisions about where to invest your limited budgets, MMM is a valuable tool to have on your belt to help you look beyond click-based attribution models.
