How To Choose The Right Attribution Model
Recently I have investigated the question of attribution modeling for Blinkist. Blinkist is using last-touch attribution and would like to move to something sophisticated.
The issue with last-touch attribution is its ignorance of the customer journey. The model is not realistic in today’s world, because customers have many touchpoints. The touchpoints can be across channels and devices. Furthermore, the model is rule-based and retrospective.
Last-touch comes from the ’90s when e.g. computation/storage was expensive. Now we have the cloud and pay-as-you-go services. It makes no sense to use last-touch anymore, but we are still stuck in the ’90s!
After reviewing 15 attribution models I concluded that attribution modeling is complex. It comes from the fact that there is no “ground truth”. No way to a/b test your attribution models. Still, you can do two things to check your model’s performance.
- To measure incremental uplift or/and
- Test your models on other markets besides the main market
Even though you can’t a/b test your models there is something better than last-touch. Be aware of the “silver bullet” trap. There are no silver bullets with tenfold improvement ~ Fred Brooks. You should not expect to double your business because of the new attribution model.
Here are the three models that I found particularly interesting:
Custom attribution - A mix of rules and custom-weights: You assign points based on customer interactions. For example, if a customer bounces off, you can assign lower scores. To run this model you need to have a good understanding of the business domain. This model doesn’t need special data engineering skills. Suitable for small companies and startups with up to 50 employees.
Dynamic attribution - Markov chains, GBM + Shapley values: This model uses Markov chains to calculate channel performance. A so-called removal effect (Rv) will tell you the exact attribution of each campaign. Clear in explanation and low in computational resources. It requires data engineering skills. Suitable for companies between 50 and 100 employees.
Predictive attribution - Logistic regression, Ensembles, Neural nets: My favorite model because it calculates attribution based on the future outcome. In businesses with longer sales cycles (7+ days), it is the best performing model. Also, for businesses with higher CAC ($700+), it can drastically reduce the costs. The main drawback is its low explainability, which depends on which algorithm you use. Besides that, you need data engineering and machine learning skills. Suitable for companies with 100+ employees.
Remember, attribution modeling has one purpose: We want to find profitable campaigns and exclude the rest. Like segmentation in marketing: We don’t segment to know whom to target, we segment whom to exclude from our targeting.
Furthermore, attribution modeling is a management board topic. Don’t make the mistake and let marketing decide about it. Attribution modeling is a top priority for top management. Check and discuss the model’s performance every month.
Let me give you a final piece of advice. Always go with the simplest model first and add complexity as go. Want to know more? Check out my presentation here.