Quick summary
Summarize this blog with AI
Product sense interviews ask you to connect an ambiguous product question to a decision, a measurable outcome, and a credible investigation. These ten original practice prompts include answer outlines. They are not an employer’s official questions or a script to memorize.
For each answer, clarify the user and decision, define the metric’s numerator, denominator, and time window, identify failure modes, and propose evidence that would change your recommendation. State assumptions before selecting an experiment.
1. Advertising revenue fell. What do you investigate?
First establish the size, timing, affected markets, and whether billing and event collection are complete. Decompose revenue into eligible traffic, ad opportunities per user, fill rate, and realized revenue per filled impression using compatible definitions. A decline in one component can be hidden by growth in another.
Compare stable segments and check releases, inventory, advertiser demand, and seasonality. Distinguish a reporting defect from an economic change. Tell management what is confirmed, what remains a hypothesis, and which next check is most likely to separate the explanations.
2. How would you assess a secondhand marketplace?
Measure whether buyers find suitable items and sellers complete worthwhile transactions. Define a successful transaction, then inspect completion rate, time to sale, repeat participation, cancellations, and fraud reports. Listing count alone can rise while useful supply deteriorates.
Compare cohorts with similar age, geography, and category. A new group has had less time to transact than an established one. Report both total activity and an opportunity-based rate, such as completed transactions per eligible listing within a fixed window. Explain supply and demand constraints before proposing growth tactics.
3. Would threaded comments improve discussion?
Define the intended benefit: easier replies, better comprehension, or more useful exchanges. Consider eligible viewers who meaningfully participate or find a helpful answer, alongside harassment reports, moderation workload, latency, and abandonment. More comments or longer dwell time can reflect confusion or conflict.
Test a clear interface change, retain the same eligibility rules, and account for participants influencing one another within a thread. Discuss whether assignment by discussion or community is more appropriate than independent user assignment, and how that changes uncertainty.
4. What happens when parents join a network used by teenagers?
Separate a research question about existing behavior from an experiment on an optional product feature. You cannot randomly assign people to have different family relationships. Use only appropriately consented, permitted relationship information; do not infer family links from sensitive or covertly collected data.
Define the population and assess meaningful participation, unwanted contact, privacy controls, and reporting behavior. Observational differences between users with and without a parent on the network can reflect selection. If testing a consent-based feature, consider spillovers between connected people and use the product’s applicable privacy and research review process.
5. How would you measure and detect false news?
Define the label and review policy before counting. Estimate prevalence using a representative sample with appropriate sampling weights, rather than only reviewing posts already flagged by a detector. Distinguish the fraction of posts affected from exposure-weighted impact.
For a classifier, separate training from later evaluation data and avoid leakage from repeated stories. Report precision, recall, uncertainty, and performance across relevant languages or topics. To raise recall, examine false negatives and threshold tradeoffs while tracking false accusations and review capacity. A higher recall figure alone does not establish a better moderation policy.
6. How would you evaluate “people you may know”?
Start with a permitted candidate set and a simple baseline. Define a useful connection using acceptance and subsequent meaningful interaction over a stated window, with unwanted-contact reports and blocks as guardrails. Clicks are an intermediate signal.
Use time-separated offline evaluation without future relationship information. Account for new users and popularity bias. Online, consider network interference: a recommendation can affect both the recipient and the suggested person. Explain how you would limit exposure and investigate adverse outcomes.
7. A smaller feed layout raises US revenue but lowers Japanese revenue. Why?
Check assignment, instrumentation, confidence intervals, and whether the regional comparison was planned. Decompose changes in scroll depth, viewable impressions, fill, pricing, and user behavior. Do not infer a cultural explanation from a country label.
Inspect device mix, text layout, latency, inventory, and opportunity denominators. More items on screen need not mean more viewable ads. Treat exploratory segment findings as hypotheses to validate, and evaluate user experience alongside revenue before deciding on a regional rollout.
8. How many food-delivery notifications should users receive?
Clarify which messages are essential order updates and which are optional recommendations. Respect notification preferences. For optional messages, define incremental completed orders or useful engagement per eligible user, with opt-outs, complaints, and notification volume as guardrails.
Use an appropriate holdout to distinguish incremental behavior from orders that would happen anyway. Compare a small set of predeclared frequency policies and examine delayed effects. There is no defensible universal maximum derived from click-through rate alone.
9. An email account switcher increases accounts per user but not time spent. Is it successful?
Ask what problem the switcher is meant to solve. Faster access to the right account may reduce time while improving task completion. Measure successful switching, error rate, task time, and support requests among eligible multi-account users.
Check whether “accounts per user” reflects duplicate identities, newly linked accounts, or genuine adoption. Examine distributions and cohorts rather than only the mean. Flat total time is neither automatic failure nor success; connect the result to the intended user benefit.
10. How would you evaluate nearby restaurant recommendations?
Clarify the context: discovery, navigation, reservations, or delivery. Filter unavailable or unsuitable options using current, permitted data. Define useful outcomes such as a completed reservation or a user-confirmed successful visit when available; do not invent ground truth from a map click.
Compare a simple distance-and-availability baseline with personalization, and track coverage, latency, stale listings, and undesirable recommendations. Restaurants have inventory and opening-hour constraints that a people-recommendation problem does not share. Explain cold-start handling and how an online test would measure incremental benefit.
Practice by choosing one prompt and answering it in five minutes, then asking a partner to challenge your metric and causal assumptions. For a numerical follow-up, use the worked product analyst case.
