Attribution asks how credit should be assigned across observed interactions. Incrementality asks what changed because of an intervention compared with what would have happened otherwise.
Both can be useful. Treating them as the same question can turn a reporting result into a causal claim the evidence does not support.
What attribution provides
An attribution method organizes credit under defined rules or models. It helps describe observed journeys and compare activity within that framework. Google’s Analytics attribution overview explains the platform’s concepts.
The result depends on available data, identity, windows, and the method. A credited sale is not proof that the final observed interaction independently created the entire purchase.
What incrementality requires
An incremental effect depends on a credible comparison with the counterfactual: the outcome without the intervention. A well-designed experiment can help estimate that difference. Observational methods require their own assumptions and limitations.
You cannot directly observe both versions of the same customer journey at the same moment. The design must address that problem rather than hide it behind a dashboard label.
A hypothetical example
Suppose many existing customers search for a company’s name before buying again. An attribution report may assign credit to a brand ad they clicked. The separate question is how purchasing would differ if that ad were absent under an appropriate comparison.
This example does not establish that brand ads have no value. It shows why credit and causality need different evidence.
Use each for its proper decision
| Question | More relevant approach |
|---|---|
| Which observed interactions receive credit? | Attribution reporting |
| Where do reported journeys differ? | Consistent journey analysis |
| Did an intervention create additional outcomes? | A credible causal design |
| Should the next budget increment be funded? | Economics plus the strongest available evidence |
Report uncertainty explicitly
Explain the method, scope, and what the result cannot establish. Avoid presenting a modeled estimate as a directly observed fact. When evidence is limited, make the next test smaller and more informative.
The experiment design guide and brand-search analysis apply this distinction. A useful measurement system lets the team ask a sharper question before choosing a more elaborate model.
Co-founder and CEO of Stackmatix, startup advisor, and former Head of Sales at MightyHive. Connect with Matt on LinkedIn ↗ · More about Matt →