PM Recommendation Systems
(2026 Edition)
In a recommendation system, ML owns the model but the PM owns the objective function: what to optimise, which signals count, and the diversity and freshness guardrails that keep candidate generation and ranking from collapsing into filter bubbles. That means setting the cold-start experience for new users and items and owning evaluation — making sure offline metrics actually predict online business outcomes.
By Naman Goyal · Product manager · Builder of PM Streak · Updated July 3, 2026
5 system layers and 5 PM-owned levers in a recsys product.
Build Recsys PM Skills — Free →5 System Layers
Candidate generation — narrowing from millions to hundreds
Ranking — ordering the candidates by predicted engagement
Diversity and de-duplication — avoiding filter bubbles and repetition
Business constraints — margin, freshness, content policy
Post-ranking polish — presentation, thumbnails, labels
5 PM Levers
Define the objective function — what are we optimising?
Choose which signals are in scope — explicit ratings? implicit behaviour?
Set guardrails — diversity, freshness, policy compliance
Design the cold-start experience — new users, new items
Own evaluation — offline metrics must predict online business outcomes
FAQ
Do PMs need to understand ML to build recsys products?
You don't need to write models, but you must understand the basic shape — candidate generation vs ranking, offline vs online evals, training data feedback loops, and why short-term engagement optimisation can destroy long-term retention. PMs who can't engage with ML tradeoffs get steamrolled by the ML team.
Keep learning