FIELD GUIDE 18 / Data & measurement
Attribution without false certainty: share credit across a messy journey
Attribution distributes credit for an outcome among recorded touchpoints. It is useful for reporting journeys, but credit is not the same as causation and the recorded journey is never the whole decision.
QUICK ANSWER
The idea in one minute.
What last-click, rules-based, modelled and experimental measurement can and cannot say. Attribution assigns credit; incrementality estimates cause Tracking observes only part of the journey Compare models and state assumptions
KEY IDEAS
What you will learn
- Attribution assigns credit; incrementality estimates cause
- Tracking observes only part of the journey
- Compare models and state assumptions
01
Distinguish credit from cause
Last-click gives all credit to the final recorded interaction; first-click favours discovery; linear and position-based rules distribute credit according to a formula. These models describe a chosen accounting rule. They do not prove what would have happened without each contact.
Incrementality asks the causal question by comparing outcomes with a counterfactual. Use experiments to estimate whether activity changed behaviour, then use attribution and journey data to understand where recorded interactions occurred. The methods answer related but different questions.
02
Know what tracking misses
People switch devices, clear identifiers, refuse tracking, see offline media, consult other people and return through untagged links. Platforms observe different portions of the same journey and may each claim credit for one outcome.
Privacy protections and data minimisation appropriately reduce individual-level visibility. Do not treat missing data as a technical defect to defeat. Use aggregated measurement, surveys, experiments and business totals to complement event streams.
03
Use models as estimates
Data-driven attribution estimates contribution from observed patterns under model assumptions. It can adapt to complex paths but remains sensitive to data quality, selection, eligibility and changes in the platform. A precise percentage is still an estimate.
Compare more than one view. If last-click, a modelled report, a geographic experiment and business trend suggest different channel roles, investigate the assumptions instead of averaging them into artificial agreement.
04
Make allocation decisions transparently
Publish the attribution window, eligible events, identity rules, deduplication, excluded traffic and model version. Show how reported conversions reconcile with orders and cancellations. A change in method can create a performance jump without changing customer behaviour.
Use ranges and scenario analysis for budget decisions. Protect strategically important discovery and brand activity from a model biased toward easily observed final interactions, while still demanding evidence that those activities add value.
SOURCE DESK
Research and further reading
These links lead to public guidance, open textbooks or freely accessible research. This guide explains the ideas in original language; open the sources to examine context, definitions and limitations.
- About attribution models ↗Google Ads Help explanation of last-click and data-driven attribution.
- Advertising attribution ↗Google Analytics Help overview of attribution paths, windows and reporting.
- Introduction to Experimental Design ↗OpenStax Statistics on random assignment, treatment and control.
- Privacy-enhancing technologies ↗UK Information Commissioner overview of techniques that reduce data use and exposure.