How Does Marketing Mix Modelling Work?

Marketing mix modelling works by comparing how your sales moved over time with how your marketing spend and other outside factors moved over the same period, then using statistics to estimate how much each channel contributed. It does not follow individual people. It reads patterns in aggregate data, which is why it keeps working as tracking becomes harder.

The core idea: sales move for many reasons at once

A business rarely changes only one thing at a time. Spend on search goes up, an email campaign runs, a discount appears, the season shifts, and sales change. Marketing mix modelling treats sales as the outcome and every one of those influences as an input, then asks how much of the movement in sales each input can reasonably explain.

The inputs usually include spend by channel, pricing and promotions, seasonality, and external factors such as holidays. The more consistently these are recorded over a long stretch of time, the more clearly the model can separate one influence from another.

The steps a model goes through

The work starts with data. Sales or revenue, spend by channel and commerce records are gathered into a consistent time series, usually daily or weekly. Gaps and mismatched definitions are cleaned up, because a model can only be as sound as the numbers it is given. This is why connecting ad platforms, analytics and commerce data cleanly matters so much before any modelling begins.

Next, the model is fitted. Statistical methods estimate the relationship between each input and sales, while allowing for effects that real marketing has. Advertising often keeps working after the spend stops, and returns tend to flatten as a channel is pushed harder. A good model accounts for both the lingering effect and the diminishing return, and then it is checked against periods it has not seen to confirm it predicts sensibly.

What you get out of it

The output is an estimate of each channel’s contribution to sales, and from that a view of channel level return on investment. Because the model also describes how returns change as spend rises, it can be used to test budget scenarios before committing money, such as moving spend from one channel to another and seeing the predicted result.

This is where marketing mix modelling differs from platform reporting. Ad platforms report the conversions they can attribute to themselves, which can overstate their own effect. A mix model looks at the total business outcome, so it is better placed to show what spend actually added rather than what would have happened anyway.

Where incrementality testing fits

Incrementality is the question of what a channel caused, as opposed to what it simply sat next to. A model gives an estimate based on observed patterns. A controlled test, where some audiences or regions see the advertising and others do not, gives direct evidence. The two approaches check each other, and using both gives estimates that hold up better than either alone.

Seeda’s platform brings marketing mix modeling, incrementality testing, channel level ROI and budget optimisation together, with your data connected in one place. If you would like to see how it measures true incremental performance on your own numbers, you can book a no obligation demo.

Frequently asked questions

What data does marketing mix modelling need?

It needs sales or revenue over time, spend by channel over the same period, and information about other influences such as promotions and seasonality. A longer, consistent history generally gives the model more to learn from.

How is marketing mix modelling different from attribution?

Attribution traces credit to individual touchpoints along a customer’s path, which depends on tracking. Marketing mix modelling uses aggregate data and statistics, so it does not rely on following individuals and can include offline and external factors.

Can marketing mix modelling tell me how to split my budget?

It can estimate how returns change as spend on each channel rises, which lets you compare budget scenarios before you commit. The results are estimates, so they are best treated as guidance that you confirm with testing.