Personalization
From recommendation engine to personalization system
Listen — audio summary in Nehal Vyas's voice
Transcript
Hi everyone, Nehal here. My latest blog post, "From recommendation engine to personalization system," argues that personalization is far more than just an ML model. True value and durable engagement come from the entire system around it. First, signals are your real moat. Personalization quality hinges on the data you collect – thoughtfully combining explicit and implicit signals while respecting privacy. Second, ranking is a product decision. Optimizing only for clicks leads to monotonous feeds, so we must encode product values like diversity, freshness, and long-term satisfaction directly into our algorithms. Finally, close the loop and make it legible. Systems improve by learning from outcomes, and users need simple controls – like "show me more like this" – turning a black box into a valuable collaboration that drives engagement. Understanding personalization as a system changes where you invest, creating durable engagement and retention. Read the full article at hinehal.com for more.
Teams often ask me to "add a recommendation engine," as if personalization were a component you drop in. In practice, the model is the smallest part. The lift comes from the system around it: what signals you collect, how you rank, how feedback closes the loop, and how legible the whole thing is to the person on the other side of the screen.
Signals are the real moat
Two companies can license the same model and get wildly different results, because personalization quality is bounded by signal quality. Explicit signals (ratings, saves, follows) are sparse but high-intent. Implicit signals (dwell, scroll depth, replays) are abundant but noisy. The craft is combining them into a picture of intent that respects privacy and degrades gracefully for new users.
Ranking is a product decision, not just an ML one
Relevance is necessary but not sufficient. A ranking that only optimizes predicted clicks collapses into a monotonous feed. I push teams to encode product values into ranking: diversity, freshness, and long-term satisfaction, not just the next tap. Those tradeoffs belong to the product manager as much as the data scientist.
Close the loop, and make it legible
Personalization systems that don't learn from outcomes stagnate. But the feedback loop only works if users understand they're steering it. Simple controls - "show me more like this," "not interested" - turn a black box into a collaboration. Legibility drives the very engagement the system needs to improve.
Treating personalization as a system rather than a model changes what you invest in: instrumentation, ranking policy, and user-facing controls. That is where durable engagement and retention actually come from.