Ranking Competing Work for the Team Behind Spotify's Music Recommendations

You are a product manager on the team that builds Spotify's recommendation engine: the models and ranking systems that decide which songs, playlists, podcasts and other audio each listener sees. The team is asked to do more than it can deliver. Explain how you would decide what the team works on and in what order, how you would handle competing requests from inside and outside the team, and how you would check that your priorities were the right ones. ```hint Score against one outcome Before comparing items, decide which outcome the recommendation engine should be accountable for, and remember that some of the team's most valuable work, such as data quality or experimentation tooling, ships nothing a listener can see. ``` ### Clarifying Questions - Which surfaces does the team own: the home feed, personalized playlists, radio and autoplay, search ranking, or podcast and audiobook recommendations as well? - Which outcome is the company asking the team to move this year: listening engagement, retention of subscribers, conversion from free to paid, or helping new artists get discovered? - Where does the incoming work come from: the team's own model roadmap, other product teams that consume its recommendations, content and marketing partners, or reliability and data issues? - Are any items fixed commitments, such as a privacy or regulatory deadline, that are not up for ranking? ### What a Strong Answer Covers - A goal and primary metric for the recommendation engine that every item is judged against, with guardrails that protect listeners, creators and serving cost. - A full inventory of the work a recommendation team carries, including model quality, new surfaces, data and platform work, experimentation and outside requests, not just features. - A scoring method that accounts for the uncertainty of ML work: reach, expected effect, confidence based on evidence, effort, and what an item unlocks. - A sequence and a capacity split that protect foundations and long-term bets, with checkpoints to stop bets that do not pay off. - How priorities are communicated, how requests are declined, and how the ranking is revisited as experiment results arrive. ### Follow-up Questions - A record label asks the team to push a major new release into more listeners' recommendations, but tests suggest it would raise skip rates. How do you decide, and who makes the call? - A new model wins clearly on offline metrics but shows no lift in listening time in an A/B test. Does it keep its place in the queue? - Leadership wants a new AI-generated playlist feature this quarter, which would take most of the team. What would you stop doing, and how would you explain the cost?