Assignment records which variant a user was allocated. Exposure records whether the user reached the point where that variant could affect them. Confusing the two can dilute an effect or introduce bias.
Use assignment for the main causal comparison
An intention-to-treat analysis preserves randomisation by comparing all eligible assigned users. It answers the operational question: what happens if we deploy this treatment policy to the eligible population?
Use exposure as a diagnostic, carefully
Exposure analysis can explain low treatment reach, but filtering after assignment may create incomparable groups. A user's journey to the exposure point may itself be affected by the variant. Define exposure before launch and test whether instrumentation fires equally.
Instrument the decision point
A robust event includes experiment ID, variant, assignment timestamp, exposure timestamp, user or account key, and the feature context. Keep server-side assignment logs where possible, and reconcile them against analytics events before estimating uplift.