Ever get the feeling your product’s recommendation engine is a brilliant expert in what your users did yesterday, but clueless about what they need right now? You’re not alone. In a world saturated with digital products, predicting the next click, the next show, or the next purchase is the name of the game. But this relentless focus on “what’s next” often misses a more profound and powerful question: “Why are they here today?”
A recent research paper from Netflix, unpacking their “FM-Intent” model, offers a compelling new playbook. It’s a sophisticated approach that moves beyond simple prediction to actively infer the goal of a user’s session. This article translates the dense data science into a strategic briefing, breaking down how this shift from predicting actions to understanding motivation can unlock a new level of product innovation and user engagement.
So, what is “user intent” if no one tells you?
(Hint: the clues are already there)
The first hurdle is a big one: users don’t typically announce their goals when they open an app. Netflix’s model gets around this by looking for proxies: implicit signals hidden in plain sight within a user’s session behavior. It’s less about mind-reading and more about brilliant detective work.
This matters for product leaders because it means you can start decoding user intent without adding cumbersome surveys or explicit feedback forms. The data is likely already in your logs. The FM-Intent model listens for clues like:
Action types: Is the user hitting “play” on a series they’re halfway through, or are they scrolling through search results? One suggests a “continue watching” mission, while the other signals a “discovery” mindset.
In-session genre focus: Is the user clicking on three comedies in a row? Their intent for this session might be a desire for light-hearted content.
Content age: Are they Browse brand-new releases or digging up old favorites? This distinguishes a “lean-forward” discovery mode from a “lean-back” comfort watch.
The takeaway for your team: Start thinking like a detective. What behavioral clues in your product could signal a user’s underlying goal for that specific session?
The “why before the what” model: a simple trick for smarter recommendations
Traditional models often try to predict the next item and the user’s intent as two separate, parallel tasks. The genius of Netflix’s approach is its hierarchy: it predicts the “why” before it predicts the “what”.
Here’s the mini-playbook:
First, predict intent: The model analyzes a user’s history and current session behavior to make an educated guess about their primary goal (e.g., “find a new sci-fi movie,” “continue my show,” “rewatch a classic”).
Then, inform recommendations: This predicted intent is then fed as a powerful new input into the next-item recommendation model. The recommendations are now directly guided and refined by this understanding of the user’s mission.
The result? The research cited a 7.4% improvement in next-item prediction accuracy. In the world of recommendation systems, that’s a massive leap. It’s proof that understanding the “why” doesn’t just feel better: it performs better.
From raw data to richer analytics: discovering your “intent clusters”
Perhaps the most powerful outcome of this model isn’t just better recommendations, but a fundamentally deeper understanding of your users. The model produces what are known as “intent embeddings,” which can be used to cluster users into meaningful, behavior-driven groups.
Instead of generic segments like “power users,” product teams can identify dynamic, session-based mindsets. The Netflix research identified distinct clusters such as:
Discoverers: Actively seeking out new and diverse content.
Continue Watchers: Focused on finishing a series they’ve already started.
Rewatchers: Returning to comforting old favorites.
This is gold for analytics. It allows product and content teams to move beyond tracking raw views and start analyzing the motivations driving those views. You can finally answer questions like, “What percentage of our weekend traffic is driven by a ‘discovery’ mindset versus a ‘comfort’ mindset?”
Putting it all to work: from a smarter row to a smarter UI
This is where strategy turns into reality. Predicting session intent doesn’t just have to live in the background; it can become a tool for creating a truly dynamic and responsive user experience.
Imagine a user logs in. The model confidently predicts their intent is to “continue watching.” What can you do?
Dynamically rerank the UI: Instead of showing the standard homepage, you could automatically boost the “Continue Watching” row to the very top, minimizing friction.
Tailor the entire page: If the intent is clearly “discovery,” maybe the interface highlights “New Releases” or “Trending Now” more prominently. If the intent is “find a movie for the family,” perhaps genre filters become the hero of the UI for that session.
This approach enhances core recommendations, improves search relevance, and gives your team a powerful new lever for personalizing the entire product experience on the fly.
The shift from simply predicting the next action to inferring the goal behind it is more than a technical upgrade; it’s a strategic evolution. It allows teams to respond directly to a user’s immediate needs, transforming the product from a static library of options into a dynamic, intelligent partner.
So, as you head into your next workshop session, here’s the question to debate with your team: What are the fundamental whys driving your users, and how would everything change if you could start building for those, not just for their next click?
Share this post