Context
Rippling is considering a new employee referral campaign inside the Rippling product: admins can invite peer admins at other companies to book a demo and, if the referred company becomes a customer, both sides receive a payroll credit. The Growth team wants to test whether a more prominent referral flow in the Rippling admin dashboard increases qualified referred leads without degrading core product usage.
Hypothesis Seed
The proposed treatment adds a persistent referral card on the Rippling home dashboard plus a one-click share flow with prefilled invite text. The team believes this will increase the rate of qualified referred-company demo requests, but the experiment is tricky because one treated company can influence untreated companies through cross-company referrals, violating SUTVA.
Constraints
- Eligible traffic: 24,000 active customer companies per week
- Average company size in the experiment: 180 employees, but the referral feature is only visible to company admins
- Baseline weekly qualified referral rate: 3.0% of eligible companies generate at least one qualified referred-company demo request within 14 days
- Maximum decision window: 4 weeks
- False positives are expensive because finance must fund credits and sales capacity is limited; false negatives are acceptable if the design is cleaner
- The team can randomize by company, sales region, or referral-network cluster, but engineering prefers a simple design
Deliverables
- Define the primary metric, 2-4 guardrails, and an explicit MDE for this experiment.
- Design the experiment to account for network interference: choose the unit of randomization, allocation, duration, and any clustering/stratification.
- Calculate the required sample size with actual numbers and determine whether the 4-week traffic budget is sufficient.
- Pre-register the analysis plan: statistical test, handling of interference, peeking policy, multiple-comparison policy, and SRM checks.
- State a clear ship / don’t-ship / iterate rule that respects guardrails, including what you would do if the primary metric improves but interference makes interpretation ambiguous.
