Ingredients for a Successful MTA Implementation

We are still grappling with the problem we call marketing attribution a century after Wanamaker uttered that quote. Although modern technology has given us much better tools for measuring marketing efficacy, we also have much more complex consumer journeys and many more channels to analyze. To be able to accurately assign credit to marketing channels and campaigns is tremendously powerful and allows a marketer to efficiently allocate resources and derive in-depth insights into their customers.

In digital marketing specifically, multi-touch attribution (MTA) tools are designed to assign credit across multiple channels algorithmically by finding patterns in successful (converting) and unsuccessful (non-converting) customer journeys. This is much more sophisticated than the traditional heuristic approaches familiar to most marketers: last-click, first-click, linear, U-shaped, etc. It is unnecessary to dissect each heuristic approach individually, as they all suffer from the same two basic problems:

• Human bias – heuristics, by definition, are artificial rules that will reflect the bias of the person who has defined them.
• Oversimplification – as an example, last-click measurement does not account for any of the advertising that ultimately led to a conversion. Upper-funnel channels and tactics are substantially undervalued relative to lower-funnel channels. The assumption behind last-click is essentially that demand-driving campaigns have no effect on customer behavior.

Moving from a heuristic attribution model to MTA is a large organizational shift both in mindset and technology. From a technical perspective, there are two components to successful MTA: (1) tracking and measurement and (2) choice of algorithm.

Tracking

Without knowing the exact series of touchpoints each customer has interacted with, it is impossible to appropriately attribute credit, as we do not know how the channels interact with each other and influence customer behavior. A robust multi-channel tracking solution will be the foundation of any successful MTA tool. If the data is collected properly, algorithms can always be experimented with and changed at-will. Without good data, there is no point to applying an algorithm and therefore no MTA.

The majority of tracking and measurement is performed through cookies which are read and set through pixels and redirects. While tracking solutions are all similar for the most part, there are a few things to keep in mind when choosing or implementing one:

• Track as much as you can behind first-party domains. This increases cookie persistence and makes pathway stitching more reliable, leading to more accurate models.
• Many solution providers utilize probabilistic identification of customers. This means that, instead of identifying an individual directly by cookie, other information such as IP address and hardware specifications are used to derive a high-quality guess as to who an individual is. This approach augments customer journeys by tying together multiple devices (or multiple sessions from the same device if cookies have been reset). A potential disadvantage lies in the word probabilistic. There is a good chance that many of the consumer journeys from probabilistic identification will actually be composites of multiple individuals, which muddies the data significantly and can lead to false results. Probabilistic stitching is not a bad thing in itself, but its inherent limitations and drawbacks should be considered before employing it.
• Data QA, preferably in an automated manner in real time, is extremely important. Tracking will occasionally break for a segment of live campaigns. This can happen for a number of reasons – bugs, human error – and is practically unavoidable. Most tracking issues can be resolved quickly once they are identified, mitigating data loss and minimizing the impact on MTA models. When issues are not discovered immediately, MTA models become degraded and their outputs will be unreliable. We strongly recommend automated QA between multiple sources of truth to monitor for any aberrant measurement and avoid outages in your MTA solution.
• Data sparsity is always a concern. When determining the granularity of MTA analysis you would like, keep in mind that adding more touchpoints means there will be less data to base decisions on for each one. In other words, if you have 100 campaigns in one channel, each campaign will, on average, have 1% of the amount of data that the channel does in aggregate. More fine-grained analyses result in more directional model outputs.

Algorithms

Once an effective tracking and measurement solution is in place and sufficient historical data has been collected, work can begin on selecting and validating an algorithmic approach to attributing value across touchpoints. There are many different mathematical approaches to choose from – this is far from an exhaustive list. However, the following four algorithms are common and illustrate different ways to approach the MTA problem.

Markov

Markov chains are a very common framework across many scientific disciplines for analyzing systems that have a probability of existing in, and transitioning between, multiple states. A Markov chain is simply a set of states and the associated probabilities of transitioning between any two states, which are represented by “nodes.” There will be a change in the probability of reaching an ultimate state if any single node in the chain is removed. In marketing, each touchpoint can be considered a state. Simulations can be run whereby touchpoints are removed from the chain to determine the theoretical impact on conversions. The more conversions lost by removing a touchpoint, the more value is attributed to that touchpoint. The Markov approach is convenient because it makes relatively few assumptions, therefore it is easy to build models, and exists as an open-source package in R.

Shapley Value

Imagine a game where multiple players cooperate to gain value. Each player contributes a certain amount to the total output of the team and should receive an award at the end of the game commensurate with how much value the player has contributed. This is exactly the scenario that the Shapley value method of attributing value imagines and it originates from the game theory branch of mathematics. In marketing, each touchpoint is a player in the game, with multiple touchpoints cooperating to drive each conversion. Each touchpoint is attributed the value that it drove across all conversions. This approach is commonly used in marketing – in fact, it is what Google Analytics uses for its data-driven attribution methodology. One of the nice things about the Shapley value method is that there are no parameters to tune or set, making it ideal for automated, set-and-forget modeling.

Logistic Regression

Logistic regression attempts to predict an outcome from a number of input variables. For example, it can be used to predict a team’s score given statistics about player performance during a game. The application to marketing is very straightforward: by being able to accurately predict whether a customer will convert (or what the value of the conversion will be) from a combination of touchpoints, individual touchpoints can be evaluated for their importance in predicting the conversion, which is the value they should be attributed. The drawback of logistic regression is that, as with other machine learning models, there are various parameters that need to be selected and tested. That procedure can be time consuming and computationally expensive and should be performed by someone with some experience in statistics.

Survival Models

Survival analysis is generally described as predicting system collapse: mechanical failure, a sports team losing, and so on. More broadly, it can be used to predict when a change in a system will occur. Our goal would be to predict the “survival” of a journey until a conversion, given the touchpoints constituting the journey. By changing the touchpoints that make up a journey and analyzing its survival probability, values can be assigned based on how much less likely the journey is to “survive” being exposed to a new touchpoint. This type of model is very attractive mathematically. There are numerous secondary analyses, apart from the attribution, that can be performed as side-benefits of the models, and the models can be made very complex to account for phenomena that are not generally accounted for in any of the previously described algorithms. However, with the additional flexibility and complexity comes a much lengthier and more difficult implementation process, requiring a high degree of expertise.

Which one is best?

Ultimately, choosing the right model depends on your data and the resources you have available. There is no harm in testing multiple algorithms and comparing and contrasting the outputs of each. Some approaches require more data than others. Some approaches require more computational power than others. Having a reliable model that is 95% accurate is far more valuable in most cases than having a model that is 96% accurate but is only functional half of the time. Finally, it is important to actually test the outputs! If you re-allocate some portion of your marketing budget based on the MTA outputs, do you see an improvement in ROAS? Good, the model works. Otherwise, you should try a different approach.

Our Approach

One use case of Softcrylic’s Tapestry first-party tracking capabilities is powering MTA models, among other customer journey-centric analyses. We stitch customer journeys together across channels and devices with custom tracking implementations designed to deliver the insights our clients require at varying levels of granularity. Along with the tracking implementation, our AutomateOn automated data QA solution monitors discrepancies between data sources and flags any potential tracking issues.

We are algorithm-agnostic when it comes to MTA. All of the aforementioned algorithms have been demonstrated to be effective in attributing value across marketing touchpoints and we have the expertise needed to build and tune any of them.

MTA is a huge step forward from heuristic attribution approaches and usually requires a very large investment of resources to implement. Establishing an MTA solution can be a complicated and long process, but by collecting clean data, choosing an appropriate attribution model and testing to verify the outputs, MTA can have a profound impact on your marketing efficiency.