Attribution Models

An attribution model is a rule for dividing credit for an outcome across the exposures that came before it. It governs how credit is shared, not whether the outcome happened, so the same outcome measured under two rules gives two different numbers and neither of them is wrong.

What you can do with a attribution model

• tableEvidence Plan · Evidence
• tableEvidence Plan · Settings

Its attribution models— 6

• First TouchCredit rests wholly on the earliest exposure in the window. Single-touch, so it reads a journey as though nothing after the opening mattered.
• Last TouchCredit rests wholly on the exposure closest to the outcome. Single-touch, and the quiet default in most reporting, which is what makes closing channels look like the ones that did the work.
• LinearCredit is divided equally across every exposure in the window. Multi-touch, and the plainest way to stop one touch taking all of it, at the cost of holding a first impression and a closing click to be worth the same.
• Time DecayCredit is divided across every exposure, weighted towards those nearest the outcome. Multi-touch, and it assumes recency is worth more, which suits a short buying cycle and misreads a long one.
• Position BasedCredit is concentrated on the first and last exposures, with the remainder shared among what came between. Multi-touch, for journeys where opening and closing are held to matter most.
• Data DrivenCredit is assigned by a model fitted to observed outcomes rather than by a fixed rule. Multi-touch, and the weights it arrives at have to be demonstrable before they can be audited, which is the price of not choosing them in advance.

Why it matters

What goes wrong without it, as the commitments that license this kind say it, each standing on its sources.

• Make outcome quality, calculation method and attribution inspectableAn outcome number is only as good as the inputs, exclusions and window that nobody wrote down. Media measurement standards exist because outcome claims circulate detached from their method, and a figure with no stated method cannot be challenged, reproduced or compared against anyone else's. Keeping what a measure means apart from how it is computed is what makes either one reviewable.
  Outcomes and Data Quality Standards · Declaring the Grain · Measuring Visibility in the AI Era · Data-driven multi-touch attribution models · Causally motivated attribution for online advertising · A Comparison of Approaches to Advertising Measurement: Evidence from Big Field Experiments at Facebook · Why Three Exposures May Be Enough · Digital Audience-Based Measurement Standards