· PM Editorial · Product Sense · 6 min read
Improve Spotify Discover Weekly: Complete Product Sense Answer
A structured, interview-ready answer to 'How would you improve Spotify's Discover Weekly?' covering recommendation quality, cold start, diversity, and listening habits.
“How would you improve Discover Weekly?” is one of the most common product sense prompts at Spotify, and it also shows up as a proxy question at Netflix, YouTube Music, and TikTok interviews because the underlying challenge — personalized recommendation at scale — is universal. This post walks through a complete, structured answer you can adapt in a live interview.
Why This Question Gets Asked
Discover Weekly is Spotify’s flagship personalization product: a 30-song playlist refreshed every Monday, built from a mix of collaborative filtering, natural language processing on music blogs, and audio analysis. Interviewers use this prompt to test four things simultaneously: can you clarify an ambiguous goal, can you segment users meaningfully, can you generate a recommendation-system-aware solution, and can you reason about tradeoffs between relevance and serendipity. It’s a product sense question wearing a machine learning costume — you don’t need to know the model architecture, but you do need to think like someone who understands feedback loops.
Step 1: Clarify the Goal
Before proposing anything, pin down what “improve” means. In an interview, say this out loud:
“Before I dive in, I want to clarify what ‘improve’ means here — are we optimizing for engagement with the playlist itself, for long-term retention, for subscription conversion, or for artist discovery diversity? These can pull in different directions, so I’ll assume we’re optimizing for sustained listening engagement and long-term retention, since that’s Spotify’s core business lever, and I’ll flag tradeoffs with diversity along the way.”
This single move separates strong candidates from average ones. Most candidates jump straight to feature ideas; strong candidates first negotiate the objective function.
Step 2: Segment the Users
Discover Weekly serves very different jobs depending on who’s listening. A useful segmentation:
- Power discoverers — save 5+ songs per week, follow new artists, use Discover Weekly as their primary discovery channel.
- Passive listeners — let the playlist autoplay in the background, rarely save songs, but keep listening (low active signal, high implicit signal).
- Genre loyalists — listen almost exclusively within 1-2 genres and get frustrated when Discover Weekly ventures outside that lane.
- New users (cold start) — have under 90 days of listening history, so the model has thin signal to work with.
Each segment needs a different lever. Power discoverers benefit from more aggressive novelty; genre loyalists need tighter relevance guardrails; cold-start users need a fundamentally different recommendation strategy (more reliance on demographic and context signals, less reliance on collaborative filtering).
Step 3: Diagnose the Core Problems
Three structural issues limit Discover Weekly today:
- Cold start is unsolved for new and casual users. A user with 20 total plays gives collaborative filtering almost nothing to work with, so recommendations default to generically popular tracks — which then get skipped, further starving the model of good signal.
- The feedback loop only sees success, not failure modes clearly. A skip could mean “wrong song,” “wrong mood right now,” or “already know this song, don’t need it in Discover.” Spotify currently can’t distinguish these well, so the model treats them identically.
- Diversity and relevance are in tension without an explicit dial. Users who want familiar-adjacent music and users who want genuinely novel discovery both use the same feature, and there’s no way to signal which mode you’re in this week.
Step 4: Propose Solutions
Solve cold start with an explicit onboarding signal. Add a lightweight, optional “taste calibration” flow for new users — 10 song pairs, pick-one-or-skip, taking under 60 seconds. This gives the model explicit preference data on day one instead of waiting 60-90 days for implicit signal to accumulate. Netflix does something similar with its rating carousel at signup.
Introduce context-aware skip reasons, sparingly. Rather than a heavy survey, add a single optional micro-interaction: after a skip within the first 5 seconds, show a barely-there “not my mood / not for me / already know it” tap. Keep it fully skippable — most users will ignore it, but the subset who engage generate high-value labeled data that disambiguates the skip signal.
Give users an explicit discovery dial. A simple “more familiar” / “more adventurous” toggle for the week’s playlist lets genre loyalists and power discoverers self-select their tradeoff instead of the model guessing wrong for one or the other.
Widen the diversity floor with a soft quota. Guarantee that at least 3-5 of the 30 tracks come from artists the user has never streamed, with audio-feature similarity to their taste profile as a bridge — familiar enough to not feel random, novel enough to actually expand taste.
Step 5: Define Success Metrics
Any solution proposal needs paired metrics so you can show you understand tradeoffs, not just upside.
| Lever | Primary metric | Guardrail metric | Risk if unchecked |
|---|---|---|---|
| Cold start calibration | Day-30 retention for new users | Onboarding completion rate | Adding friction to signup |
| Skip-reason micro-interaction | Skip-signal disambiguation rate | Overall skip rate (should not rise) | Feature fatigue if overused |
| Discovery dial | Save rate by dial setting | Playlist completion rate | Users defaulting to “familiar” and killing discovery goal |
| Diversity floor | New-artist follow rate | Weekly playlist save rate | Novel tracks dragging down completion rate |
Step 6: Prioritize and Sequence
If asked “what would you ship first,” the honest answer is the cold-start calibration flow, because it has the best evidence base (Netflix, Duolingo, and Hinge all use onboarding preference elicitation successfully), the clearest win condition (new-user retention is a known weak point), and the lowest cross-segment risk (it only touches new users, so it can’t regress the experience for existing power users). The discovery dial is the second priority because it directly resolves the genre-loyalist-vs-power-discoverer tension with a UI answer instead of a modeling answer, which ships faster than any model retrain.
Common Mistakes Candidates Make
Most candidates fail this question not by lacking ideas but by skipping structure. Watch for these traps:
- Jumping straight into “I’d use more machine learning” without specifying what data or objective changes.
- Treating all users as one segment and proposing a single fix for everyone.
- Proposing changes with no proposed metric to validate them.
- Forgetting that Discover Weekly’s business goal (retention, subscriptions) can conflict with its stated goal (discovery) — and not naming that tension explicitly.
How This Connects to the Broader Interview Loop
Recommendation-system product sense questions (Discover Weekly, Netflix homepage rows, YouTube “Up Next,” TikTok For You page) all reward the same underlying skill: naming the tension between relevance and exploration, then proposing a concrete, testable lever rather than a vague “more personalization” answer. If you can walk through clarify, segment, diagnose, propose, and measure for one of these, you can walk through it for any of them.
For a broader set of worked examples across product sense, metrics, and estimation questions, see The 100x Product Manager Interview Playbook (Amazon: https://www.amazon.com/dp/B0DBC1FQWH?tag=sirjohnnymai-20), which includes full frameworks for recommendation-system, marketplace, and growth product questions with the same structured approach used above.