Product experiments

Turn a product question into a measurable decision. A feature flag without trustworthy exposure data and a decision rule is release control, not an experiment.

Choose a mode

• Design: create an experiment brief and analysis plan.
• Implement: add assignment, exposure tracking, metrics, and safeguards.
• Rollout-only: release safely when causal inference is unnecessary.
• Validate: audit instrumentation and assignment before launch.
• Analyze: estimate effects and diagnose data-quality failures.
• Conclude: decide ship, iterate, continue, or rollback and record why.

Keep design vendor-independent. Use an existing analytics/flag provider when present; add a new provider only when selected or explicitly authorized.

Experiment brief

Before implementation, record: