Comparison of multi-channel attribution models
This article describes the causes of inaccurate assessment of advertising channels with popular attribution models in Google Analytics and formulates the principles of a more reliable algorithms. The material will be useful to analysts and marketers seeking to develop objective evaluation model of advertising channels.
The most popular model of evaluation of advertising channels on the last touchpoint (Last Click) or the last indirect contact (Last Non-Direct Click), which is used by default in Google Analytics.
It might seem that what could be more logical than assigning 100% of the value to the channel, which was the last before the transaction.
However, this model does not answer the important questions:
- Why decrease in the budget for Channel A has led to a drop in the channel B revenue?
- Why the increase in the budget for channel B has led to an increase of channel C CPA?
Obviously, the channels influence each other and their real contribution is different from the attributed on the Last Click model.
So, why is this happening?
Let’s have a look at the Path Length: Conversions /Multi-Channel Funnels /Path Length
Most likely, your report has a similar pattern, and more than half of the income is generated by the orders committed after a user visited your website for several times.
This means that a part of visits in the sequence that led to a purchase remains unestimated because the latest source receives the full value. To estimate the extend of the problem, it is necessary to measure the proportion of sessions that remains without evaluation using the Last Click model.
You can estimate the share of these sessions by making two segments in Google Analytics.
The first one will include all the sessions in which an order was made:
And the second — all sessions of users who made an order:
In the first segment, all sessions with conversion are counted. And in the second are all sessions of users who have placed an order, including sessions without a conversion. Obviously the second one is bigger. It is clear even on a small project, that using the Last Click attribution model only 26% of sessions that have contributed to the resulting transactions, obtain value.
There’s no need to wonder why you are taking the wrong decisions based on the biased channel estimates when you are using Last Click attribution model.
Well, let’s see what else we can use apart from Last Click.
In Google Analytics, we can get a report on the Assisted Conversions and see the income, that was brought by the channel, even though it was not last in the chain:
Better, isn’t it? But for what reason the total obtained value is not equal to the sum of the values of the associated and direct conversions? And have all associated conversions made the same contribution to the conversion of the visits chain? And If not, how to manage all these?
Unfortunately, there is no answer for these questions in Google Analytics. It would not be easy to convince your CEO in the trustworthiness of this report.
Non Last-Click Attribution Models
For those who feel that Last Click attribution model is not enough, Google Analytics offers a wide range of other models:
The problem with this approach is that if you apply different attribution models to the channels, the sum of attribution values will be bigger than the total value received.
CFO wouldn’t be happy. Therefore, you should choose a single model for all channels.
But First Interaction and Position Based models have the same drawback as the Last Click one — they ignore the contribution of the majority of sessions that influenced the sales funnel progress on customer’s way to an order.
Linear, Based Position and Time Decay Models, despite the fact that they distribute the value of all sessions, do so only on the basis of the source position in the chain. To clarify, let’s consider another model:
In Google Analytics 360, there’s an attribution model, which aims to solve the above-described limitations of other models — Data Driven Model. It distributes the value of all sessions in the chain on the basis of a correlation between the presence of a source in a chain and conversions in a chain. This approach is way more advanced and unbiased than all of the above.
|Last Click||Is clear for everyone||Doesn’t take into account 80% of session that led to an order|
|Position based||Is available in Google Analytics||Attributes more than we have earned|
|Associated conversions||Considers mutual impact||Attributes more than we have earned|
|Data Driven||Considers each session||Black box based only on Google Analytics data|
Attribution depending on user’s passage of the sales funnel
Is there a solution? Yes, and you will like it. Imagine that you have access to all information about the actions of each and every user in all sessions. You know, in which session user have seen the product that he purchased subsequently. Through what source the user decided to add the item to the cart, and what source motivated him to place the order. In this case, each session can be estimated based on the value of committed actions — strictly according to its contribution to the user’s passage of the sales funnel. What else could be better? Calculations in Google BigQuery! Just imagine that in Google BigQuery you have access to all data for all the necessary calculations, and you can easily check what value has each of the sessions received and why.
And the good news does not end there: you don’t need Google Analytics 360 for implementation and you also don’t need to involve developers! In OWOX BI, we automated the collection of unsampled data in Google BigQuery and calculations according to the funnel based attribution model.
Sorry, but you have no more excuses to use the Last Click model. Here’s what you should do now:
- Check what share of orders is made in the first session in your project.
- Setup calculations in Google BigQuery using Funnel Based Attribution.
- Learn the real value of your advertising channel.
- Make decisions and optimize advertising costs in favor of channels that really benefit your business.