WTF is marketing mix modeling?

The job of the CMO has become even more challenging.

More and more, success is tied to hard metrics like financial results. CMOs are having to show how various advertising and marketing tactics led someone to purchase their company’s product, as well as whether they drove softer metrics like raising people’s awareness of a brand. That proof has to convince CFOs — who still see marketing as a “cost center” despite CMOs’ best efforts — to maintain marketing budgets.

One approach is to use marketing mix modeling, which allows CMOs to show business leadership how their efforts help the bottomline. “CFOs love it because a lot of analysis is done in silos,” said Jon Turner, global chief analytics officer at Mediahub, adding that those silos can add discrepancies into reporting. “With marketing mix modeling, you look holistically so it can’t explain more than what your sales actually are. It explains all the sales and allocates them to various marketing drivers.” 

Sure but what is marketing mix modeling?

It’s a way of using statistical analysis as a tool to look back at sales over a period of time to determine what exactly caused those sales. Essentially, it’s a way of helping marketers and agency execs contextualize what’s working and what’s not. For example, say a marketer who typically spends the majority of their ad dollars on TV reallocated that spending to digital channels and offered a discounted product price. If that approach accounted for higher sales figures, that marketer could then take that analysis, tweak their approach and optimize it to spend more of their budget on what’s working and less on what’s not.

Sounds like an obvious thing to do. How does it work? 

Marketers and agency execs input data to the analysis based on not only the marketing tactics they are using but each activity that a brand may deploy or encounter. So they’re not only accounting for digital, TV, out-of-home, radio, podcast and social media advertising but the price of a product and various promotions that are being run. Of course, that’s not all. That’d be too easy. They’re also accounting for things like inventory levels, seasonality, even shifting weather patterns — basically anything and everything that could impact sales. That data is then compared to previous sales data, often at least three years’ worth, to show how sales have changed and give a reason as to why they have changed. It’s correlation over causation.

If that sounds like a vague synopsis, well, that’s because it is one. The model is specified for each brand and has to account for anything that would cause sales peaks for valleys. 

OK so it’s just another attribution method. Big whoop. 

Well, yes and no. While it is a way for marketers to point to a reason for sales, it’s also a predictive model to help marketers make decisions for the months ahead. Marketers will use the analysis — often on a quarterly basis — to see the shifts that are happening and move dollars around to hopefully continue positive trends. Should the model show that a particular channel is working more, they’ll likely move more marketing dollars there. Take out-of-home, for example. As people returned to travel and commuting following lockdowns, it’s become a more useful channel again so marketers are spending more there.

But you just brought up the pandemic. Doesn’t that throw a wrench in the whole thing? 

In some ways but not really. That’s why marketers use a few years’ worth of data for marketing mix modeling. “When you have a shock to the system like Covid, having years’ worth of data becomes even more important,” explained Larry Davis-Swing, evp of advanced analytics at Spark Foundry. “By having plenty of data before it and plenty of data after you can start to understand and isolate all of the stuff you saw happening during Covid.”

Davis-Swing continued: “When markets shut down, we saw consumer behavior shift. People went from going to restaurants to doing takeout and delivery. We saw delivery explode. So we can account for that initial explosion, not because of advertising or marketing, but because consumers had to change their behavior.” 

So yes, data from mid-March 2020 to the end of 2020 — maybe even summer 2021 — is a bit of a wash as consumer behavior changed significantly, making it harder for predictions to come to bear. However, as people get back out of their homes and return to pre-pandemic activities, marketers can then weigh the data from 2019 higher and factor more normal behaviors in to help future predictions be more accurate.

That’s why you have to make sure the inputs are correct.

Exactly. Marketers and agency execs have to think through everything that might account for sales variation so the model can work properly and help with predicting how they should be allocating their marketing mix. If you have a model that’s trying to explain the variation in champagne sales, you’re going to have to input a peak on New Year’s and Valentine’s Day, explained Trisha Pascale, group director of analytics at The Many. If you don’t account for that, the model could be inaccurate and the predictive element of it useless.

Accounting for shifts in marketing and advertising strategies is important too. With the turnover of one CMO to another, which tends to happen every 18 months or so, there’s often a shift in strategy. If you haven’t accounted for more digital advertising or whatever the change may be in the marketing mix modeling, then it won’t show how that shift is working.

Unlike multi-touch attribution, marketing mix modeling isn’t run at the consumer level, so the more personalized data that could go away with the death of the third-party cookie isn’t as important for marketing mix modeling. 

“We’re talking about really big trends, and we’re not building these models at the consumer level,” said Michael Salemme, svp of analytics at Zenith. “There are ways to run aggregate data to continue to run [marketing] mix modeling. We’re trying to explain changes in sales typically at a national or regional level, so we just need to know approximate exposures.”

https://staging.digiday.com/?p=441088
Digiday Top Stories
Ad rendering preventing in staging

Ad position: web_bfu