Making Sense of Attribution

This post picks up as Part 3 in a series on marketing attribution and analytics. In case you missed it, here’s Part 1 and Part 2, which both focus on the concept of lead source.

“All models are wrong, but some are useful” - George Box

What is an attribution model and why is it important?

An attribution model is a framework that determines how to assign credit for a transaction to various touchpoints in a customer journey.

You’re probably already using at least one model, even if you don’t realize it. 

Parts one and two of this series covered “source” and how the source of a contact is likely different from the source of a purchase or gift, but both data points are useful. The source of the contact is generally referred to as “First Touch”—the first time that person engaged with your organization. The source of a transaction is generally a “Last Touch” model—the last engagement point before the transaction took place. Both of these are different kinds of attribution models.

These models help us understand which marketing tactics are working and what aren’t. Ideally, we can make decisions about where to allocate resources based on these models—whether that’s investing more into a certain campaign or channel that’s performing well, or shutting down an initiative that’s lagging to try something new.

Multi-Touch

A more sophisticated approach is using a multi-touch model, meaning we credit multiple points of engagement with a desired outcome. These more sophisticated models can help us understand a “considered purchase”—a long buying cycle that involves analysis and comparison, often with multiple stakeholders.

In a multi-touch model, marketing touchpoints are each assigned credit as a portion of overall revenue. If we identify 10 touchpoints that influenced a $10,000 transaction, that $10,000 is broken up across each of the 10 touchpoints with a dollar value assigned to each.

So how do you assign that $10,000 across 10 touchpoints?

The simplest approach is to break $10,000 across the 10 touchpoints and assign $1,000 to each (“Even Touch”). However, there are many other approaches. You might choose to weight touchpoints close to the transaction more heavily and decrease the weight as you look further back in time (“Time Decay”). Another model is “U-Shaped” and puts equal weight on the first and last interaction, and less on everything in the middle. Hubspot has a nice breakdown of some different models if you want to get into the details, but the point is there are a lot of ways to assign credit.

Challenges with attribution models

Despite the promise of these models—and the benefits touted by app vendors—the reality is more complicated and these tools often don’t meet expectations.

There are a number of reasons why models might not show what we expect, or when we unpack the data we realize it doesn’t meet expectations. What are some of the reasons for this?

1. We don’t know which model accurately represents our customer experience. The Hubspot article above shows a variety of different models with some rough guidance around how to choose, but that’s no guarantee the model we pick is the right one.
   
   
2. The tech that we have may not be able to provide the reports that we want. Different systems provide different ways to model attribution; even if we know what we *want* it to look like, that doesn’t mean there’s a straightforward path to implementation.
   
   
3. Missing data. Even with a system that provides the model we want, it’s going to fall flat if we don’t provide the data it needs. We need to know what touchpoints took place, with whom, and what the conversion point was. At Pedal Lucid, this is a common problem we notice—from touchpoints that aren’t tracked in a CRM to Opportunities created at the last minute with no associated Contacts.
   
   
4. We don’t always know whether a touchpoint should “count.” Email is a good example. If we can track a link click from a specific email to a purchase, that’s pretty black and white. But everything else gets murky. Assigning credit to every email send isn’t practical. So what should we count? Opens are no longer a reliable metric. Clicks are more accurate, but what if we share content directly without requiring a click?

Even with the right model, right tech, right data, we can still struggle to get useful reporting. These are major challenges.

Making sense of it all

The good news is that we don’t need to rely on models being correct for them to be useful. We need to take the results with a grain of salt, but they can still provide significant value.

1. Break down journeys into smaller, more measurable conversions. These don’t necessarily need to live in the same system; see our previous Pretty Good Answers post for a deeper dive into this concept.
   
   
2. Use more than one model. Toggle between a few different options, look for differences and try to understand what those differences mean.
   
   
3. Focus on relative comparisons rather than absolutes. This is particularly relevant for revenue share and dollar amounts. Using an even touch model can help us with this by removing the “weight” variable entirely. It drives home the point that we know the raw numbers are wrong and makes us focus on comparative analysis that will be more useful.
   
   
4. Test to draw conclusions. Sophisticated models should be one of several tools you use, but A/B testing can help validate what you see in a model.
   
   
5. Build your own model if you can identify gaps in out-of-the-box features and can articulate how a model approach fills those gaps. This approach can help you work around known limitations in the tools or data you have. It may sound intimidating, but if you have a clear plan of how the model *should* work the tech lift might be lower than you think.

Attribution models aren’t magic but they can be powerful tools when used thoughtfully. Focus on asking the right questions, use multiple sources, focus on comparisons over absolutes, and don’t be afraid to customize your approach. The path might not be linear, but with a nuanced understanding of a model’s strengths and weaknesses we can use them to make better decisions.