If you have a high average path length or are running campaigns at the top end of the marketing funnel (social media, display, PR) then you should be looking at attribution closely and assessing the impact it is having on your results and therefore your decisions. In this post, I’ll breakdown the different attribution models available in Analytics to try and help you better measure results from your campaigns.
Why Use Attribution Modelling?
Accurate measurement is key to running a successful marketing campaign. It enables marketers to make informed decisions around budget and campaign adjustments without any guesswork.
You may be thinking that you are all set on this front; campaigns are set up, goals are tracking and eCommerce data is pulling through to Google Analytics. What you may not have considered is how sources are attributed to these conversions, this can make a significant and important impact on your results.
This should be of particular interest to you if a large number of your conversions come after the first visit. You can see this information clearly displayed in the path length report in Google Analytics. Take the example below, in this report, you can see that 42.4% of all conversions come after the first visit:
With this data set, it is clear to see why there would be an attribution problem. The standard attribution model for most Analytics tools is closely tied to last click conversions, this means that the last source a user visits your website from prior to converting will be attributed with that conversion. For the example above that means that up to 42.4% of the conversions on the site have interactions in them which never receive any credit for those conversions.
Click below to see information on each attribution model:
- Last Interaction
- Last Non-Direct Click
- Last Adwords Click
- First Interaction
- Linear
- Time Decay
- Position Based
1. Last Interaction (last click) attribution model
Apart from Google Analytics (which uses Last Non-Direct Click by default), the last interaction model applies as the default model in most Analytics packages.
In a lot of ways, this is the simplest model to explain which may well be why it is the preferred default in most packages. The final interaction/source that led to the conversion is attributed 100% of the value for that conversion.
Take the following conversion path which has 40 conversions where paid search, organic search and a direct visit were involved:
Using the last interaction model direct traffic would receive all 40 of these conversions. Looking at examples like this is clear to see the model has issues and often does not provide a useful view of conversions.
It is still useful however when you are running campaigns specifically designed to convert customers immediately. This could be the fast-moving consumer goods (FMCG) sector for example where purchases would not usually require a considerable period for the customer.
2. Last Non-Direct Click
The last non-direct click attribution model is used by default in Google Analytics to assign conversions. This model attributes the full conversion to the last source of a converting visitor that was not direct traffic.
Using the same example as the previous model we can see how this would be applied.
In this model rather than attributing the conversions to direct traffic, they would be attributed to organic search as this is the final interaction point when direct is disregarded.
Arguably this model is more helpful than the first when trying to measure campaign performance as it will skip direct traffic to give the conversion to the last campaign based visit. However, it would still ignore paid search in this example and by ignoring direct traffic does not treat sources in an equal way.
You can see in the comparison below the difference it would make account-wide looking at Last Non-Direct Click (Google Analytics default) vs Last Interaction:
As you can see in this example the difference can be dramatic. As you would expect when regular last interaction is applied the number of direct conversions rises sharply, at the expense of all other traffic sources.
3. Last Adwords Click
Last Adwords click is a model that anyone running Adwords campaigns should take note of. This model will always attribute the conversion to Adwords should it appear in the conversion path.
Using the previous example we can see how this model would apply.
In this instance Paid Search (assuming it was Adwords paid search) would receive all 40 of the conversions. It does not matter at which point Adwords is in the path, for example, below is another path whereby Adwords would still gain all 49 conversions using this model:
Although this model can be helpful in showing you the overall contribution of Adwords clicks, scientifically it makes little sense. There is some solid ground for skewing attribution toward a particular position (first, last, middle etc) but to skew this much toward a particular source does not provide a fair picture of results and performance.
A cynic might suggest this model only exists to promote the success of Google Adwords campaigns in Google Analytics data.
4. First Interaction
interaction is straightforward like the last interaction model in that it attributes the full value of a conversion to the first source in the conversion path.
Using a further example below we can demonstrate this easily
The first organic search visit in this path would be attributed to the full 187 conversions using the first interaction model.
Compared with the default Google Analytics model (last non-direct click) a first interaction approach can provide interesting results:
You may have expected the results to be more varied than they are. The reason the variation is not extreme is because still the majority of conversions take only one visit (so attribute the same conversion on both first and last interaction) and many other conversions will see the same source appear several times, take the example below where 318 conversions happened a result of a path which only included 4 separated organic visits.
It is important to note that this model is effective at showing conversions for campaigns that operate higher up the sales/marketing funnel. In particular display campaigns or social media campaigns set up with a sole purpose of generating brand awareness and exposure. Although small numbers you can see in the table above that when comparing the models’ conversions to display advertising rose from 7 to 17. This could well be the difference between wanting to continue a display campaign and stopping it entirely.
5. Linear
The linear attribution model gives each source in the path an equal share of the conversion triggered.
the above example each source; Paid, Organic and Direct would receive an equal share of the 40 conversions – 13.33 conversions per source.
This is the most ‘fair’ model available and the argument against its use would be that different points during a path do deserve a different level of attribution.
6. Time Decay
The time decay model addresses many of the issues with the linear model by attributing an ever-increasing % of the conversion the closer to the final click you get.
By default the Time Decay model has a 7-day half-life, in practice, this means that an interaction that occurs 7 days prior to conversion would receive ½ the credit of an interaction that occurred on the day of conversion. This continues to apply backwards so that 14 days prior to a conversion an interaction ¼ of the credit of one on the day of conversion.
If we assume that in the example above Paid search occurred 14 days prior to conversion, Organic search occurred 7 days prior and direct occurred on the same day then the time decay model would provide the following attribution:
- Paid Search: 14.3%
- Organic Search: 28.6%
- Direct: 57.2%
Or
- Paid Search: 5.7 conversions
- Organic Search: 11.4 conversions
- Direct: 22.9 conversions
This can be quite complicated to work out manually but that is what the tools are for!
In principle, this attribution model makes a lot of sense across a range of circumstances. It is logical to begin decaying attribution from first interaction as the users memory of that interaction will also decrease over time but won’t be completely lost. This model provides a reasonable method for showing that decrease without removing attribution completely from early interactions.
7. Position Based
The position-based model by default attributes 40% of the conversion to the first interaction, 40% to the last interaction and the remaining 20% is spread evenly across the middle interactions.
In the example above this would mean 40% of conversions would be attributed to organic search and direct, whilst email and display would receive 10% each. In numbers this would be:
- Organic Search: 4 conversions
- Direct Traffic: 4 conversions
- Email: 1 conversion
- Display: 1 conversion
This attribution method is particularly good for showing the value in both awareness-building campaigns and the final interaction campaign leading to conversion whilst not ignoring the interactions in between. It represents a happy medium between first and last interaction models, the two interactions which for many marketers are the most important.
Conclusion
There is no perfect attribution model. All attribution models have flaws and it’s important to know what those are when you are creating reports and measuring success/failure.
Each model has a different use case and it very much depends on what you are trying to measure and the campaigns you are running as to which one to try out. Models should not be chosen to try and skew results upwards but to ensure you have the correct information to make an informed decision.
Through the Google Analytics model comparison tool (Conversions->Attribution->Model Comparison Tool) you can compare various models across your dataset quite easily. This is a good place to begin when deciding whether you should be using a different model on a more permanent basis.
Once you become more familiar with the models it is also possible to create a custom model, assigning varying levels of attributions to different positions within the path. It is probably best to steer clear of this however unless you have a scientific basis for modifying the position values.
If you would like to learn more about attribution modelling or help with measuring your campaigns you can contact us here.