Clario Marketing Analytics
Request a demo
blog_tiny

Evidence-Based Marketing

Exploring Marketing Attribution - Part 2 of the  Evidence-Based Marketing Series

Posted by Matt Redlon on Sep 20, 2019

Attribution reveals which marketing efforts advance the customer along your desired journey most efficiently and effectively. Neither last touch attribution nor first touch attribution alone properly credit each interaction, and no single model works for every marketer. Learn more about Marketing Attribution models in this video in our Evidence Based Marketing series.

 

 

Video Transcription

This is the second video in our series on Evidence-Based Marketing. In this video, we're going to dive into marketing attribution. Attribution is the marketer's attempt to reveal which activities advance customers along their desired journey most efficiently and effectively. As you can see in the evidence-based marketing framework in the first video, we described the integration and enrichment of event and customer data. Since that's very foundational to modeling attribution, if you haven't watched that video yet, now would be a good time to go back and do that before proceeding with this video. 

To get started, let's bring back the classic retail case that we described in the first video. Remember, this is a customer who engaged with multiple marketing activities along their path to purchase. Foundational to understanding attribution is understanding the purchase funnel. The purchase funnel is a way to think about the consumer journey from initial awareness through an iterative process of product consideration based on features, and pricing, and competitive set, ultimately through to the purchase of the product. From the marketer's perspective, the purchase funnel can be thought of as casting a very broad net initially for awareness and then nurturing a customer through a journey of consideration that ultimately yields a purchase from a certain percentage of those customers. The marketer's activities vary depending on which stage of the funnel they're focused on. For example, a digital marketer might be focused at the awareness stage on video advertising on YouTube, or display advertising on Facebook or Instagram. Offline, a marketer might be engaged in radio, television, print, or even involved in billboards depending on the type of product, all of this to generate initial product awareness.

Marketing Attribution Examples
The consideration phase for marketers really is an attempt to understand how much purchase intent the customer has. For example, we might be looking at pay per click advertising, so search. We might be looking at having great SEO-rich content that will draw them in as they begin their iterative search process. We will continue to generate awareness because a customer might have learned of our product category from another vendor and yet be picking up our consideration stage advertising as well. Additionally, in the consideration phase, we're doing activities as a marketer to gather a little bit of information from those people who have demonstrated strong intent to purchase. From those individuals, we might get their email and that's how, in the example above, we began sending them emails, which they can click on, further driving them and nurturing them through that consideration process.

We might also gather their mailing address and begin to engage with them through direct mail. We might even gather digital identifiers from partners who place pixels on our website so that we can retarget these individuals across the web. With all that in mind, let's look at some of the common ways that marketers measure attribution for their programs today. The most common web analytics platforms such as Google Analytics and Adobe support a variety of attribution models. I'm going to talk to you about them in terms of single-touch models and multi-touch models. The most common single-touch models are first-touch and last-touch. Last-touch attribution simply says, the final touchpoint before the purchase gets 100% of the credit for that purchase, all other touchpoints receive 0%.

Last Touch Attribution Model(Last Touch Attribution)

There are a couple of challenges with last-touch attribution. The first is that it favors activities which are late in the purchase funnel. For example, retargeting with display ads, remarketing with email, branded search, where there's already high brand awareness, or even affiliates, where the customer intended to purchase and just went to the affiliate's site first to see if there was a coupon available. This can lead to very shortsighted budget allocation decisions if you use the last-touch attribution to allocate budgets and that results in an underinvestment in upper-funnel activity such as brand awareness. The second challenge is common to all digital-only attribution models and that is that things like email sends, direct mail sends, television, and radio are very difficult to incorporate and as such most of the models ignore them entirely.

First-touch attribution reverses this pattern and essentially says, let's take the first-touch and give it 100% of the credit instead of the last-touch. Unfortunately, the first touch model just switches the bias from lower-funnel activity to upper-funnel activity and still results in a misallocation of resources. As a result of these shortcomings, most marketers today are looking at some form of multi-touch model for their attribution. The most popular of these are position-based, linear, time decay, and algorithmic models and we're going to talk quickly about each.

First Touch Attribution Model(First Touch Attribution)

The key to understanding a position based model is to understand that the middle activities on the customer's path to purchase are going to share 20% of the credit for that purchase. Beyond this, we are just trying to decide how much credit to give to the first and the last events in the path to purchase. For example, with a U-shaped attribution, you will give the first and the last step in the path to purchase 40% each with the balance going to the middle and with a J-shaped, you will give the first 20%, the middle 20%, and the last 60%. With an inverse J, you're going to just flip that around and give the last 20,% the middle will share 20% and the first touchpoint will get 60%. In a linear model, all we're doing is taking the total number of touchpoints along the customer's path to purchase and giving each one an equal share of the credit for that purchase. In a time decay model, what we're doing is usually decaying the importance of each of the marketing touchpoints over some time period. Very commonly we use a seven-day half-life for each of them. As we go back in time with each of the marketing touchpoints, they get a little bit less and a little bit less and a little bit less.

Any of these multi-touch models are significantly better than the single-touch model we described before. The biggest challenge I find with these models is that they aren't customized for each marketer, each customer cohort, or each purchase channel, like web desktop versus web mobile. Additionally, and somewhat controversially, channel managers and agencies are going to be incented to pick and choose which of these models they implement based on their financial incentives. For example, an agency may push for a last-touch attribution model if they know that they are incented based on ROAS and can fairly easily change spending to inject non-incremental touchpoints just prior to the purchase via retargeting or branded search.

One aspect we haven't discussed yet is the concept of a lookback window. Whether you're looking at the first-touch attribution model, or linear, or position-based, or time decay; in all of these models, we have to understand how far back in time we're going to look to draw a cutoff and determine that only events subsequent to that cutoff can be considered for credit. If we assume, in our example above, that this was a 14-day lookback window, you can see how if we shift this to a seven-day lookback window, all of these events would lose their credit and that credit would then be shifted to these two more recent events.

Marketing Attribution 7 Day Lookback Window(Example of 7 Day Lookback Window vs. 14 Day as shown above)

Equipped with the understanding of a lookback window, we can understand another challenge to these methodologies, which is that all marketing programs or other events which we are using to predict the outcome, purchase or otherwise, in an attribution model are going to be given the same treatment via the lookback window. We have to assume that an email click, Facebook display ad or engagement, interacting with a form on your website, all have an equivalent influence and longevity of influence per these methodologies.

The final methodology to discuss is the algorithmic methodology, which attempts to address a lot of the shortcomings of the other multi-touch attribution methodologies we discussed. The goal of an algorithmic model is to use rules or statistical algorithms to try and give each marketing touchpoint credit reflective of its underlying incrementality. Incrementality is a fancy way of saying it incrementally moved the customer towards an ultimate purchase.

I'm going to be describing a model built on survival analysis, but there are other techniques as well such as Markov chain. At Clario, we decided on survival analysis simply because it incorporates the time that has elapsed from each marketing indicator until purchase, not just the sequencing and interaction of those indicators. One of the most powerful tools the data scientist has when building an algorithmic attribution model is the decay function or half-life created uniquely for each marketing indicator utilized by that marketer.

The decay function serves two primary purposes. First, it replaces a global lookback window with a lookback window specific to each marketing indicator. Perhaps in this example, an email send has a lookback window of seven days, meaning that when the data scientists looked at all of the purchase events, the non-purchase events, and the email send indicators together, they determined that there was a meaningful difference in outcome back to seven days prior to the event, but in the same model, an email click might have a 14-day lookback window.A combination of business rules and statistical analysis might determine that a Facebook impression or click, when it is the first event for a stitch identity could have a period of influence of 30 days for example.

Lookback Window Specific to Marketing Indicator(Example of three different lookback windows vs. a global lookback window. Purple line represents email send, 7 days. Burgundy line represents  email click, 14 days. Green line represents: Facebook impression or click as first event, 30 days.) 

The second purpose of a decay function is to represent mathematically the rate at which that particular marketing indicator is influencing purchase each day prior to the desired event, say, purchase. In our example, email sends might have a very steep decline in their influence over time. Email clicks might have a more modest, and Facebook, in the case where it is the first event in and stitched identity, could have a very modest or slow burn type impact over the course of its period of influence. Finally, and in sort of non-statistical algorithm terms, we're simply looking for the existence of these marketing indicators prior to the purchase event or the other desired event, and then the amplitude or magnitude of the decay function at the point of occurrence for that marketing indicator.

Now, I didn't draw in all of the decay functions, but hopefully what you understand in using this algorithmic approach, you're able to set a weighted influence for each of the marketing events on the path to purchase in a highly customized way. One that is specific to customer population, like newly acquired customers versus loyal retain customers, versus lapsed customers, specific to the purchase channel, say, on the web, or via the phone, or via the retail store, and inclusive of events whether they were online or offline, meaning they're outside the scope of traditional digital marketing attribution solutions. Additionally, these models are flexible enough to incorporate all forms of evidence, meaning things that have resulted in joint tests with you and your agencies, with you and your media mix modeling vendor. All of that data can be incorporated into and influence the custom nature of the model that has been produced to deliver algorithmic attribution.

I hope this video has done a good job introducing you to the general concepts of marketing attribution and really demonstrated how these concepts build on the framework of event data and customer snapshot data we presented in the first video in the series. I hope you'll join us next time where we will continue building on the first two videos in this series and get to customer lifetime value, both predicting and acting on it. And as always, I really appreciate you taking this evidence-based marketing journey with us. Thanks.

Matt Redlon

Written by Matt Redlon

Search

Subscribe to our blog

Join over 3,000 retail leaders and get access to news, ideas, and best practices.