You’re looking for something, anything, to improve your marketing attribution, and you’ve heard of Marketing Mix Modeling (MMM), but it makes you nervous. It’s a black box to you – you’re sure it works for some people, but you don’t know enough about the mechanics to know if it’ll work for you. Do you have the right data? Do you have enough of it? Sound familiar?
These are common questions that I’ve run into over the past year talking to people while building vexpower.com. The truth is that marketing mix modeling sounds like rocket science, often because the people who offer it want to get paid like rocket scientists do! Marketing mix modeling isn’t just for the Fortune 500, but it is a holistic exercise: you’re going to have to incorporate a lot more than just what you did in marketing to get an accurate model.
To see all the sources of data that go into a typical marketing mix model, check out my data request checklist template (no signup needed).
It’s called ‘Marketing Mix’ Modeling because the idea is you incorporate factors from all of the 4 P’s of the Marketing Mix. Product, Price, Place and Promotion. Many marketers these days only control that last P – advertising spend, public relations, affiliate partnerships, discounts / offers – but in the golden age of marketing you got to control the other 3 P’s too. Growth marketing has brought things like Product back under our locus of control, but just because you don’t control a factor doesn’t mean it doesn’t affect sales.
These projects can be quite big exercises – hundreds of variables across months of work. However the projects I work on tend to be much smaller in size: anything from an afternoon in GSheets, to a couple days or weeks in Python. For smaller projects like this, you typically only need a handful of the most important variables to get you 60% to 70% of the way there.
One thing the statisticians don’t tell you: with MMM you have exactly the same number of data points as a Fortune 500 company!? How is this possible? Well MMM is based on time series, which means if you are running your model for a 3 year period, you have 1,095 data points, just like anyone else does! Of course there are other things bigger brands can afford to do to increase their accuracy, but they also have far more complex businesses. If you're only running a handful of channels and have one or two main products, you should be able to predict your sales with not too complex of a model.
I highly recommend starting small with your model, and seeing how much you can predict with one or two variables (total marketing spend + main keyword search trend is a good bet). That can give you a directional steer, and then when it’s time to add more complexity to improve accuracy, you can go back to collect and clean more data to use later.
However even when I am starting with a small, ‘minimum viable model’, I find it helps to run through a checklist of all the data I have found useful to include in other models. Checklists are a great way to see if you’ve missed anything, and they help jog your memory or inspire you to include other alternative data you otherwise wouldn’t have thought of.
For example if you were selling chess sets when The Queen’s Gambit came out on Netflix, you’d have to include viewing figures to get an understanding of why your sales spiked that month. Or if you are a travel company, swings in exchange rates can govern how many people decide to vacation in a specific location.
Not all of these data sources are relevant, and most variables you test won’t end up being significant, so don’t get overwhelmed. Modeling is about confidence, and covering all the bases is only going to help you give a stronger presentation to your boss or client when you finish the project.
If you didn’t catch it earlier in this post, now would be a good time to check out my data request checklist template (no signup needed). Feel free to use this as a template for requesting data from clients, or internally from the rest of your organization, when building your next marketing mix model. Happy modeling!
