Hierarchical Blended Metrics
Allow us to output incrementality metrics at the same granularity as the imported last touch data. Currently, we support 3 levels: channel, campaign, and site (which can be used for anything including ad set, publisher, sub-publisher, or creative). We'll be adding a 4th level dedicated to creative. Through our blending technique, further levels of granularity don’t decrease certainty in the model, which is the main problem with attempting to model deep granularity.
The second iteration of Blended Metrics, called Unified Blended Metrics will allow us to truly triangulate signal from last touch and modeled incrementality data.. These unified blended metrics will incorporate both forms of signal, weighting more heavily towards one or the other based on the confidence of modeled incrementality data. This is an incredibly exciting and actionable update for clients looking to use incrementality data to inform last-touch, and operate with one unified source of truth.
Hierarchical Blended Metrics allow us to output incrementality metrics at the same granularity as the imported last touch data. Currently, we support 3 levels: channel, campaign, and site (which can be used for anything including ad set, publisher, sub-publisher, or creative). We'll be adding a 4th level later dedicated to creative. Through our blending technique, further levels of granularity do not decrease certainty in the model, which is the main problem with attempting to model deep granularity with MMMs traditionally.
We train an MMM to evaluate incrementality and then use more granular last touch data to model deeper levels of granularity. This works because, at high enough volumes, last touch data is normally good at measuring proportions. For example, while last touch is very bad at measuring the true performance of a campaign, it is usually pretty good to determining the proportional performance of that campaign's two ad sets (e.g., last touch might attribute 70% of the revenue to the first ad set and 30% to the second).
New customers simply need to select a granularity other than Channel in the Billing page and import last touch data at that granularity. Existing customers can select a new, deeper granularity in the Billing page (keeping in mind that this may incur higher fees) and then re-import historical app events (last touch) data at that level of granularity. This will cause a retraining of the MMM, but it should result in the same or very similar incrementality at the channel level. The main difference in the new incrementality output will be the incrementality at the deeper levels like campaign and site. This historical re-import can be fully automated if using a data partner integration for app events data. Otherwise, the more granular last touch data must be uploaded to S3.
The second iteration of Blended Metrics, called Unified Blended Metrics will allow us to truly triangulate signal from last touch and modeled incrementality data.. These unified blended metrics will incorporate both forms of signal, weighting more heavily towards one or the other based on the confidence of modeled incrementality data. This is an incredibly exciting and actionable update for clients looking to use incrementality data to inform last-touch, and operate with one unified source of truth.
What It Is
Hierarchical Blended Metrics allow us to output incrementality metrics at the same granularity as the imported last touch data. Currently, we support 3 levels: channel, campaign, and site (which can be used for anything including ad set, publisher, sub-publisher, or creative). We'll be adding a 4th level later dedicated to creative. Through our blending technique, further levels of granularity do not decrease certainty in the model, which is the main problem with attempting to model deep granularity with MMMs traditionally.
How It Works
We train an MMM to evaluate incrementality and then use more granular last touch data to model deeper levels of granularity. This works because, at high enough volumes, last touch data is normally good at measuring proportions. For example, while last touch is very bad at measuring the true performance of a campaign, it is usually pretty good to determining the proportional performance of that campaign's two ad sets (e.g., last touch might attribute 70% of the revenue to the first ad set and 30% to the second).
How to Get Started
New customers simply need to select a granularity other than Channel in the Billing page and import last touch data at that granularity. Existing customers can select a new, deeper granularity in the Billing page (keeping in mind that this may incur higher fees) and then re-import historical app events (last touch) data at that level of granularity. This will cause a retraining of the MMM, but it should result in the same or very similar incrementality at the channel level. The main difference in the new incrementality output will be the incrementality at the deeper levels like campaign and site. This historical re-import can be fully automated if using a data partner integration for app events data. Otherwise, the more granular last touch data must be uploaded to S3.
Updated on: 23/05/2023
Thank you!