Metrics question

How would you test the LinkedIn feature "People you may know" when you don't have any data to base your decision off of?

Practice this question out loud. An AI interviewer asks it, follows up like a real interviewer would, and scores your answer. Type or speak.

What this question tests

Ability to design an evaluation methodology for a recommendation feature before any usage data exists, using proxies and qualitative methods.

Clarify 'no data' as no historical usage data on this specific feature, though LinkedIn has plenty of graph data to power the recommendations.
Start with offline evaluation: hold out already-formed connections and test whether the algorithm would have recommended them, precision and recall against known ground truth.
Run a small qualitative test: show recommendations to a sample of users and ask them to rate relevance directly, since click history doesn't exist yet.
Launch a limited pilot at a small percent rollout to start generating real behavioral data before a full launch.
Define the metrics you'll graduate to once data exists: connection-request rate, acceptance rate, and downstream engagement from new connections.

Separates offline evaluation using existing graph data from online evaluation needing live data, showing you're not starting from zero.
Proposes a concrete qualitative relevance-rating method to substitute for missing click data early on.
Defines the metrics graph you'd move to once the pilot generates real data.

Assuming no data means no way to evaluate at all, rather than using existing graph data as an offline proxy.
Jumping straight to a full launch without a staged pilot to generate real behavioral data safely.

How would you avoid the recommendations feeling creepy or invasive during this test?
What would be your rollout plan from pilot to full launch?