With the rise of e-commerce, online advertising, and the ability for companies to perform fine-grained digital attribution at the customer level, media-mix modeling has fallen in popularity and many marketers of the newest generation are not familiar with this very important tool for optimizing a marketing program.
In this blog post I’ll review the history of media-mix modeling, discuss why customer level attribution based on click-tracking is insufficient, review the challenges with media mix modeling in the context of “programmatic” channels, and finally, I’ll discuss how Recast addresses these problems. Let’s dive in!
A brief history of MMM
The problem to solve
Imagine that you run the marketing department for a CPG brand in the 1980’s — your products are distributed through third-party retailers like Sears, grocery store chains, or a growing chain called “Walmart”. Your primary advertising channels are television, radio, direct mail, and print advertising in newspapers). Critically, outside of the use of coupons (which have their obvious downsides and limitations), there is no good way to tell which people saw your advertisements and which of those went on to make a purchase.
“Half the money I spend on advertising is wasted; the trouble is I don’t know which half.” — John Wanamaker, retail pioneer
Enter the econometricians
In the late 80s and early 90s, econometricians began developing techniques to estimate the impact of these different advertising channels on sales. By using sophisticated regression techniques, they were able to link how changs in spend in the “media mix” (say, one week where spend in television increased and spend in radio decreased) to changes in the number of total sales. While some of the math under the hood can get tricky, the concept is straightforward: the models take advantage of “natural experiments” in the levels of media spend in order to understand how the different channels impact downstream sales.
This type of modeling became a critical tool in the marketing departments of large consumer-facing brands, and a number of companies and consulting firms stepped up to further develop this technology and provide the models as a service to brands. To this day, most large brands (especially those that are primarily distributed through brick-and-mortar stores) either do this type of modeling in-house or contract with one of a few large providers in order to drive their planning and budgeting process.
Online advertising and attribution
With the rise of the internet and, especially, direct-to-consumer internet brands, the story changed. Brands like Warby Parker and Casper that started in the early 2010’s sold directly to customers online and were driven primarily by online advertising. Critically, when advertising in channels like Facebook and Google search, you can know exactly which customers saw your ads and whether they converted or not. So rather than having to use a complicated statistical model to estimate the impact of your advertising spend, you can see the “conversion rate” and ROI right at the top of the dashboard of any digital advertising platform.
Problem solved, right?
Unfortunately, digital advertising tracking turned out not to be the panacea that some had hoped. While we don’t have space for a deep-dive into the nitty-gritty details, we’ll cover most of the generally-accepted flaws with click-tracking-based attribution methods.
Last-click attribution over-credits bottom-of-funnel channels
Most companies that do click-tracking based attribution use what’s called “last-click” attribution. That is, when doing their ROI calculation, they attribute the sale to the last advertisement the customer clicked on before converting. While in general this is a great way to get started measuring advertising efficacy, very obviously this methodology will over-credit channels that are closer to the bottom of the advertising funnel (like paid search) and will under-credit more awareness-building top-of-funnel channels (like video advertising on YouTube).
How to integrate offline and online channels?
Most big brands today, even direct-to-consumer brands that only sell online, are advertising in a mix of online and offline channels: TV, facebook, radio, google search, podcast, banner ads, direct mail, etc. This mix of channels exacerbates the problem with trying to use a pure click-based attribution model: you try to “track” customers coming from offline channels using coupon codes or surveys and then you use a rule-of-thumb smudge to try to come up with an ROI estimate.
The more channels a company is operating in, the more these types of guesswork analyses can lead the marketing team astray and lead to millions of dollars wasted on channels that don’t actually drive incremental sales.
Correlation vs. Causation
This leads us to one of the subtler — but no less important — points. Even if a marketing team were to overcome all the measurement challenges across different types of channels that target different stages of the consumer journey, they would still be left with a thorny problem: attribution models don’t, even in principle, tell the marketer where they should allocate additional spend to drive incremental sales.
Why is this? Attribution models don’t correct for something econometricians call endogeneity, and you might think of as feedback loops. We think of increased spend on marketing as causing increased sales (and that’s the effect we want to isolate), but we have to remember that increased sales can also cause increased spend on marketing, and other factors can cause increases in both at the same time. For example, someone that was already going to buy your product might click on a Google ad because it’s the simplest way to get to your site, or a marketer might increase spend in anticipation of a seasonal bump that was going to increase your sales anyway.
Without correcting for endogeneity, attribution numbers are correlations, not causal effects, and should bear a large flashing warning of “past performance does not guarantee future results”.
The future is omni-channel
Most businesses have come to recognize that the future of retail is omni-channel. Challenger brands like Casper, Warby Parker, and Harry’s have all moved into physical retail by opening their own brick-and-mortar locations or distributing through retailers like Walmart and Target, and brands like Nike have invested heavily in building a powerful e-commerce engine to complement their brick-and-mortar presence. In the omni-channel world, using a purely click-based attribution model begins to break down very quickly as it is not clear at all how to link impressions on Facebook to in-store conversions at Walmart.
This blend of omni-channel advertising and omni-channel purchase options has, justifiably, left modern marketers in a real bind when it comes to deciding how to improve the overall performance of their marketing program.
Can Media Mix Modeling can Help?
Unfortunately, things aren’t as simple as they were in the olden days, and the statistical models developed in the past are no longer valid in a world where large amounts of money are spent on programmatic advertising.
The programmatic puzzle
If you’re not familiar with the term, “programmatic” channels are advertising channels like Facebook or Google search where advertising spend happens dynamically in response to user behavior: if customers click on your ad a lot, Google and Facebook will both charge you and serve your ad to more customers.
It’s the last part that causes a problem for traditional media mix models — in a programmatic channel, the amount of spend is chosen by the platform, not by the advertiser. Additionally, the platform could react to outside causes — if an excellent TV campaign builds a lot of trust in your brand, Facebook might register increased engagement with your ads and subsequently ramp up spend to show your ads to more people. A traditional Media Mix Model will be biased toward allocating credit to Facebook for those conversions, even though it was the TV spend that really did the heavy lifting, and one might imagine that many of those people might have purchased anyway whether they saw the Facebook ad or not.
Because with programmatic advertising purchasing behavior can drive advertising spend in addition to advertising spend driving purchasing behavior, the assumptions in traditional Media Mix Modeling are broken, and the estimates are biased.
So, the whole situation is hosed and no one knows anything?
Recast’s Incrementality Model
We’ve developed a proprietary solution to solving this problem and have built Recast as a modern alternative to traditional media mix modeling. Recast was built from the ground up to use a statistical model to estimate the true incrementality of marketing spend in both programmatic and non-programmatic channels.
Recast is the only media mix model designed to operate in a true omni-channel world — to learn more, reach out at email@example.com or get started here.