Full Mock: Product Sense Round

This is a complete mock product sense round. Read the question, then set a timer for 30 minutes and answer it out loud (or write it out). Then compare your answer to the sample answer and scoring below. Be brutally honest in your self-assessment. The gap between your answer and the sample is your remaining preparation work.

The question: 'You are a PM at Spotify. Design an AI feature that helps users discover new podcasts.' This is a typical product sense question: broad enough to require scoping, AI-relevant, and set at a real company whose product you can research.

"Let me start with the user. I see three podcast listener segments on Spotify. Habitual listeners who have 3-5 podcasts they follow religiously and rarely try new ones. Exploratory listeners who actively browse and try new podcasts but struggle to find quality ones consistently. Lapsed listeners who tried podcasts but stopped because they could not find content that kept them engaged. I want to focus on habitual listeners because they represent the largest segment and have the highest lifetime value, but they are at risk of churning if their existing podcasts end or decline in quality. Getting them to follow even one additional podcast significantly increases retention."

"For the AI approach, I would propose a 'Podcast DNA' system. Analyze each podcast along multiple dimensions using ML: topic embedding (what the podcast is about), conversation style (interview, solo, debate, narrative), production quality, episode length, release cadence, and host personality. Then, for each user, build a preference profile from their listening behavior: which podcast DNA dimensions do they gravitate toward? Use this to recommend podcasts that match their preferences on dimensions they care about while introducing variation on dimensions they are flexible about."

"The model would be a two-stage system: candidate generation using the podcast DNA embeddings (find podcasts similar to what the user listens to), then re-ranking using a gradient-boosted model that incorporates contextual features like time of day, recent listening history, and the user's mood (inferred from their concurrent music listening). I chose this over a pure collaborative filtering approach because podcast preferences are more nuanced than music preferences, and the 'why' of a recommendation matters more for podcasts (users want to know why they should invest 45 minutes in a new show)."

[Interviewer note: Strong user segmentation with a non-obvious choice (habitual listeners rather than exploratory ones). The 'Podcast DNA' concept shows creative product thinking. The multi-dimensional analysis is technically sound and differentiated from generic collaborative filtering. The reasoning for choosing this approach over collaborative filtering is well-articulated. Score so far: 4/5 on User Focus, 4.5/5 on AI Appropriateness.]

"For the design, the primary surface would be a 'Your Next Podcast' card on the home screen. Rather than a generic grid of recommendations, each card shows: the podcast name and cover art, a one-sentence AI-generated summary of why it matches the user ('Similar depth to [favorite podcast] but explores tech policy instead of tech news'), a 3-minute preview clip automatically selected from the episode that best represents the show, and the podcast's DNA match score (displayed as a percentage, like '87% match'). The preview clip is critical: podcast discovery has a much higher trial cost than song discovery (45 minutes vs. 3 minutes), so we need to reduce the barrier to trial."

"For failure states: when the model has low confidence (new user or limited listening data), we show 'Trending in [genre you have shown interest in]' instead of personalized recommendations. When a user dismisses a recommendation, we show a one-tap feedback option: 'Not interested because: wrong topic / too long / not my style / other.' This feedback directly improves the model. When the AI-generated summary is low quality (a known failure mode with LLM summaries), we fall back to the podcast's own description."

"Evaluation: Model metrics: recall@10 (does the model surface podcasts the user would actually listen to?), diversity score (are we recommending across podcast dimensions, not just the same topic?), and explanation quality (rated by human evaluators on a 1-5 scale). Product metrics: trial rate (percentage of users who listen to a recommended podcast for at least 5 minutes), follow rate (percentage who subscribe after trial), and incremental listens (additional podcast listening hours per user per week). Success threshold: 20% increase in trial rate and 10% increase in follow rate over current podcast recommendations. A/B test for 4 weeks at 10% of users."

[Interviewer note: The design is excellent. The 3-minute preview clip is a creative solution to the podcast discovery problem. The 'why it matches' explanation adds transparency. Failure state design is complete. Evaluation includes both model and product metrics with specific thresholds. Score: 4/5 on Design Completeness, 4.5/5 on Evaluation Rigor.]

Scoring breakdown: User Focus: 4/5. Strong segmentation and non-obvious user choice. Could improve by discussing how the feature differs for the other segments. AI Appropriateness: 4.5/5. Podcast DNA is creative and technically sound. Two-stage architecture is appropriate. Alternative approach considered. Design Completeness: 4/5. Three states covered. Preview clip is innovative. Could improve by discussing accessibility (how visually impaired users experience the feature) and how the feature integrates with existing Spotify discovery surfaces. Evaluation Rigor: 4.5/5. Strong metric selection with clear thresholds. A/B test design is appropriate. Communication Clarity: 4/5. Well-structured, clear transitions. Could improve by explicitly stating the framework upfront.

Overall: 4.2/5 (Strong Hire). Self-assessment exercise: Compare your answer dimension by dimension. Where are your gaps? If you scored below 3 on any dimension, that is your priority for the next practice session.