Treating clicks and impressions differently in attribution logic

[benjaminsavage]

A very common heuristic in attribution is:

• If there are clicks, attribute to the most recent click.
• If there are no clicks, attribute to the most recent impression.

This is a little different than "last-touch", and it would be good to be able to support this.

This would require annotating conversions with a bit of metadata to indicate if it was an impression or a click.

[csharrison]

At the risk of "premature generalization", we can do something like ARA (spec link) which tags each ad with a numeric priority. The attribution algorithm just selects the highest priority ad (with ties broken by recency).

Your proposed algorithm can be implemented with this approach, by setting clicks priority 1 and impressions priority 0. It can also support other algorithms like first-touch (priority is negative timestamp), or "probabilistic linear" (priority is random number).

I like this approach because it doesn't hardcode as much ads business logic into the spec. If I suddenly have a new class of ad engagement (e.g. user punched the monkey 10+ times, watched a video ad to completion, etc) the two-priority approach isn't expressive enough to capture it.