Context
This was a semester-long UX project exploring how AI-driven personalization could be applied to an existing product. We chose Netflix as our subject, a platform most people already know intimately, which made both the problem-finding and the usability testing more grounded in real behavior.
The Problem
We started with a broad opportunity:
After benchmarking how other platforms use AI (Spotify's recommendation engine, ChatGPT's personalization, Alexa's proactive assistance) and auditing Netflix's current state, we refined it into a specific design direction:
Research
We mapped Netflix's current state across three lenses:
- Current AI use: how Netflix already personalizes thumbnails, recommendations, and ad targeting based on watch history
- Pain points: an overwhelming content library with weak discovery tools, minimal interactivity, and no way to see what similar viewers thought of a title
- Opportunities & constraints: the opportunity to connect viewers watching the same content in real time, weighed against real constraints: data privacy regulations, multi-language user bases, age restrictions on chat features, and the cost-to-revenue case for building it at all
We built a user persona (Alex, a millennial web developer) grounded in real Netflix usage data, subscriber demographics, password-sharing patterns, and regional revenue distribution, to keep the design decisions tied to an actual audience rather than an assumed one.
Prototype
We designed a working prototype around the refined concept: a feature letting viewers see and chat with others currently watching the same title, alongside a social ratings layer for shared reviews.
Usability Testing
We ran moderated usability tests with 6 participants spanning a wide age range, using a structured script with pre-test, in-task, and post-test questions to capture both behavior and reasoning, not just reactions.
Findings
The results didn't confirm our original hypothesis, and that was the most valuable part of the study.
By contrast, testers responded well to the ratings/reviews feature, which let them see what others thought of a title without the social pressure or privacy concerns that came with live chat.
Recommendation
Based on this, we recommended Netflix deprioritize the chat-based social feature entirely and instead invest in expanding the ratings and review experience, potentially layering in AI-assisted summarization of reviews rather than a live social/chat layer. This was a direct reversal of our initial design direction, made because the data said so, not because the original idea was wrong to try.
Reflection
The real lesson here wasn't the chat feature failing, it was that we built a complete concept, tested it honestly, and were willing to let the data override our own assumption rather than defend it. That's the process I'd want to repeat: test early enough that a wrong hypothesis is still cheap to abandon.