Attribution best practices

Picking the right attribution model for your organization depends on a number of considerations. This article explores a methodology and some general best practices.

Step 1: Exploratory Analysis


This analysis needs to happen before you pick an attribution model.

This phase consists initially of understanding customer behavior and defining conversion metrics. Based on the conversion metrics, tools like Analysis Workspace and pulling in data sources from multiple channels (such as Impressions data) can facilitate your understanding of

  • How many customers are touching different marketing channels before converting?
  • The proportion/distribution of these behaviors.

For example, if 50% of customers touch 3 channels before converting, is there any interaction among those 3 channels?
You could then do upper- and lower-funnel analysis to expand your understanding.

Upper-funnel analysis

Upper-funnel analysis analyses channels used to create brand or product awareness. For example, the goal of most TV ads is brand awareness. You might use the “Time decay” attribution model, since people will forget about your TV ad over time.

Lower-funnel analysis

In this analysis, the assumption is that people already know about your brand and you want them to convert. Use e-mail or push notifications or Facebook ads.

Step 2: Rule-base attribution

The purpose of this step is to validate your hypotheses.

Example 1

Let’s say your hypothesis is "My First-touch channel has more impact on conversion than my last-touch channel. You would then use the “Inverse J-shaped” attribution model to test this hypothesis. This model gives 60% of the credit to the first touch point.

Example 2

Your hypothesis might be: "In our industry (such as travel industry), the attribution window is 60 or 90 days, not 30 days, because customers do a lot of research before buying a product. You would then change your lookback window to 90 days.

Step 3: Use Algorithmic attribution

Because it is very hard to validate a large number of possible hypotheses and combinations, you can use algorithmic attribution to leave this work to built-in algorithms. If you have already found the perfect attribution model that answers all your questions and is a perfect fit, then you obviously don’t need to take this step.

Other considerations

  • You might need to use the services of a data scientist instead of relying on Analysis Workspace alone.

On this page