How to choose your model
By default, ChannelMix provides you with six Rule-Based attribution models and the option to add Data-Driven attribution as a seventh model. But how do you know which model is best for your business? Here are some tips on choosing your attribution model.
Single Touch Attribution
Single touch attribution models are the easiest to understand and most common, but provide the least comprehensive modeling of your data. They are unable to model how channels work together to generate target events.
The first interaction in a customer’s path receives 100% credit. This model is great for understanding which channels generate the newest user engagement, but doesn’t provide much insight into which channels lead to target events.
In this model, the last interaction before a target event receives 100% credit. This is the default model used by the majority of marketers. It tells you which interaction led directly to a target event, but doesn’t provide any information about how other channels contribute to that event.
Multi Touch Attribution
Multi touch attribution models how channels interact with one another to produce target events. These models provide a much improved understanding of customer data compared to single touch attribution. But, these models are still rule-based models meaning we have to decide which parts of user paths are important and apply our intuition to the data.
Linear attribution distributes credit equally to all interactions along a customer's journey to a target. This model tells you how often each channel appears in converting paths relative to all other channels. This is useful to see if the same channel is always appearing in converting paths, but doesn’t necessarily mean the channel is the most important in driving target events.
This model is a natural evolution of the linear model where we start deciding some interactions are more important than others. The U-shaped model is used to give the majority of value to the first interaction (the interaction that initiated engagement) and the last interaction (the interaction leading to the target). A small amount of value is then dispersed equally across the rest of the interactions. This model is great if you like last touch attribution, but you also want to give credit to the interaction that got the user’s attention in the first place.
The U-shaped model behaves the same way independent of how many interactions are included in the path. The Smooth-U shaped model adds the assumption that the more interactions occur in the path to a target, the less any individual interaction matters. The first and last interactions still receive the most credit, but the difference between the credit given to the first and second touches decreases as the path length grows.
The bell model is the exact opposite of the Smooth-U Shaped model provided for contrast. The bell model uses a Bell curve or normal distribution to attribute the highest credit to the interaction taking place in the middle of the path and the least credit to the first and last interactions. Attribution value falls exponentially as you move from the middle interaction towards the beginning or end of the path. This model is included as a contrast to the Smooth-U Shaped model.
The time decay model gives progressively less credit to interactions the further removed from target event they occur. This model is great for brands with high brand-awareness. In that case, you don’t care how a customer initially became engaged with your brand. Instead, you have users continually interacting with your marketing channels and you want to know which interactions led from a continually engaging user to a converting customer.
Data-driven attribution provides a big upgrade when compared to rule-based attribution. With data-driven attribution, you have a model learning not only which interactions lead to the target, but which interactions lead to users never engaging with your brand again.
With this full set of converting and non-converting paths to train on, the data-driven model learns how your marketing channels interact with one another directly from your data. This is a huge upgrade because you are no longer responsible for creating a rule and applying it to data. The data determines the rule by itself.
Choosing your model
When choosing which model to use, the most important thing to consider is what you want to learn from the model. This is something that can change over time and lead to shifting interest in different models.
It may be the case that you started out using the time decay model because you really wanted to know which interactions occurred closest to the target event and you didn’t care how users began interacting with your marketing. But now you’re releasing a new product and you really want to drive up engagement. In this case, you might be interested in looking at the first touch model to really focus on new user engagement. Or, you can use the Smooth-U model which does give more credit to interactions closest to the target event, but also gives larger credit to the interactions driving initial engagement. The models you come back to repeatedly will change over time as your goals change.