Product manager metrics interviews: frameworks that hold up

Build a metric tree from the product job

A metric tree explains how user behavior produces an outcome. It also gives you places to investigate when performance changes.

A metric can be useful without being the north star. Click-through rate may explain discovery, but it can reward curiosity without value. Revenue may matter to the business, but it can lag product quality or conceal who is being harmed.

• North-star outcome: the recurring action that best represents delivered value.
• Reach: eligible users who encounter the product.
• Activation: users who complete the first meaningful action.
• Depth or quality: whether the action actually solves the job.
• Retention: whether users return when the need recurs.
• Guardrails: safety, trust, latency, cost, fairness, or ecosystem harm.

Worked example: measuring saved job searches

Suppose a job marketplace launches saved searches that alert candidates when relevant roles appear.

Diagnose a metric drop in the right order

Use four passes: definition, shape, segmentation, explanation. Fixing before diagnosis is guesswork.

You do not need to write SQL in every PM interview. Knowing how you would define the population and compare funnel stages demonstrates analytical precision, even when an analyst would run the final query.

• Definition: did event logic, eligibility, deduplication, or the denominator change?
• Shape: when did the movement start, was it sudden or gradual, and is it statistically unusual?
• Segmentation: where is the effect concentrated by platform, version, market, cohort, source, or journey step?
• Explanation: which release, operational event, market shift, or behavior change matches that pattern?

WITH funnel AS (
  SELECT user_id,
    MAX(CASE WHEN event = 'alert_opened' THEN 1 ELSE 0 END) AS opened,
    MAX(CASE WHEN event = 'job_viewed' THEN 1 ELSE 0 END) AS viewed,
    MAX(CASE WHEN event = 'application_sent' THEN 1 ELSE 0 END) AS applied
  FROM product_events
  WHERE event_date >= CURRENT_DATE - INTERVAL '14 days'
  GROUP BY user_id
)
SELECT AVG(opened), AVG(viewed), AVG(applied)
FROM funnel;

Design an experiment with a decision rule

An experiment answer should cover hypothesis, unit, population, metrics, duration, risks, and the action you will take after the result.

Randomize at the level where treatment can spill over. A marketplace change may require geographic or seller-level randomization rather than user-level assignment. Run through relevant cycles such as weekdays and weekends, and define the minimum effect that is worth the operational cost.

• Name the causal hypothesis in one sentence.
• Choose a primary metric before seeing the result.
• Set guardrails and stopping criteria for harm.
• Inspect heterogeneous effects for major user groups.
• Explain what ship, iterate, or stop would each require.

Handle metric conflicts as product decisions

When one metric improves and another declines, translate both into user and business consequences.