Data-driven attribution

Updated 2026-11-02

Definition

Data-driven attribution uses a statistical or machine learning model to estimate how much each touchpoint contributed to a sale. Rule-based models apply a fixed split. Data-driven models learn the split by comparing paths that converted with paths that didn't.

Common methods include Shapley values, which estimate each channel's average added value, and Markov chains, which estimate how much conversions would fall if a channel disappeared.

Example

This example uses made-up numbers. Suppose visitors who see both email and paid search convert at 4%, visitors who see only email convert at 2%, and visitors who see only paid search convert at 3%.

A data-driven model uses those gaps. Adding email to paid search adds 1 percentage point. Adding paid search to email adds 2 points. So on a $900 sale that touched both, the model might give paid search $600 and email $300, a 2:1 split. A linear model would give each channel $450.

Strengths and limits

Data-driven models can find patterns that a fixed rule misses. They also need a lot of conversions and paths that don't convert. With a small sample, the results swing around and can reflect noise. They're harder to explain, too. When someone asks why a campaign got $312, "the model decided" is a weak answer.

They still can't see untracked influence, and they still measure correlation.

How Headstream handles it

Headstream doesn't offer a data-driven model. It has four rule-based models: first-touch, last-touch, linear and U-shaped. The choice is deliberate: every credit can be traced from a Stripe event to the visits behind it, using a rule you can check by hand. The methodology page lists those rules, and reconciliation explains how the totals are checked.