Scenario

You work on Instagram Reels growth and want to test a new save affordance that makes the Instagram Save action more prominent on the reel viewer. The team believes the change will increase downstream retention in the AARRR funnel by making it easier for users to bookmark content they want to revisit, but they expect the immediate effect on saves to be small and noisy. Because Reels engagement is highly user-heterogeneous, you are considering CUPED using pre-experiment Reels behavior to reduce variance and shorten the readout.

Constraints

• Eligible traffic: 18M daily Reels viewers globally; you may expose at most 20% of traffic during the experiment
• Maximum experiment runtime: 14 days including ramp
• Primary guardrail: Reels watch time per viewer must not decline by more than 0.5%
• Messaging/share surfaces are unchanged, but virality via k-factor should not decline by more than 1%

Question

How would you design and analyze this experiment, including whether and how you would use CUPED or another variance-reduction method, what metric framework you would pre-register, and what ship decision rule you would use given the traffic and guardrail constraints?