Building the Personalization Foundation
Tubi
Overview
At Tubi, I led the design of personalization experiences that transformed how millions of viewers discovered movies and TV shows across mobile and connected TV platforms.
With over 100,000 titles available, Tubi offered something for everyone—but finding something worth watching remained difficult. New users encountered largely generic recommendations, spending too much time browsing before discovering relevant content.
Over two years, I helped establish a personalization foundation that evolved from capturing preferences during onboarding to continuously learning from viewer interactions throughout the product.
The goal was to make content discovery more relevant, reduce the time between opening Tubi and watching something, and improve long-term engagement.
My role
As Lead Product Designer, I worked with product management, UX research, data science, and engineering to shape Tubi's personalization experience over more than two years.
I helped define the product direction, evaluated competing approaches, designed preference-capture experiences across multiple touchpoints, and used controlled experiments to guide product decisions.
My work spanned onboarding, content browsing, title details, playback feedback, and connected TV experiences—connecting user-facing interactions with the signals needed to improve recommendations.
Key Impact
Over the course of the broader personalization initiative, Tubi improved activation, reduced discovery friction, and lowered new-user churn.
Aha Moment Conversion
New-user churn
The Problem
Tubi had an enormous content library, but discovery wasn't meaningfully personalized for new viewers. Users encountered essentially the same experience regardless of their interests, making relevant movies and shows difficult to find.
The behavioral data revealed three interconnected problems: activation conversion had plateaued, viewers spent more than seven minutes finding something to watch, and 56% of new users churned.
For an ad-supported streaming service, these challenges extended beyond usability. Every additional moment spent searching instead of watching limited opportunities for engagement and monetization.
The challenge wasn't a lack of content. It was a lack of understanding about what each viewer wanted to watch.
Product Strategy: From Generic to Personalized Discovery
Our initial opportunity was to help Tubi understand viewer preferences earlier, before sufficient viewing history existed to generate meaningful recommendations.
We considered two approaches: enriching recommendations with third-party signals such as IMDb and Rotten Tomatoes ratings, or collecting preferences directly from viewers.
We prioritized first-party preference collection. Although it required greater participation from users, it provided a foundation for developing Tubi's own understanding of viewer interests without depending on licensed third-party data.
I helped translate that direction into a phased personalization strategy.
Phase 1 — Establish relevance early. Capture explicit preferences during onboarding to improve the initial discovery experience.
Phase 2 — Learn continuously. Expand preference collection beyond onboarding so recommendations could evolve with viewing behavior and changing interests.
The strategic shift was from treating personalization as an onboarding feature to treating it as a continuous product capability.
Phase 1: Building the Foundation
Our first initiative focused on collecting genre preferences when viewers opened Tubi for the first time.
Working with data science and product, we identified 26 genre categories that could provide initial signals for Tubi's recommendation system.
I designed three approaches to collecting preferences: compact pills optimized for speed, a list optimized for focus, and an image-based grid optimized for recognition and engagement.
We tested all three experiences through controlled experiments.
The image-based grid performed best, achieving a 72% onboarding completion rate compared with 66% for pills and 68% for the list.
However, the results surfaced another opportunity. Most viewers selected only three genres, limiting the richness of the signals available to the recommendation system.
I refined the grid experience to encourage broader preference selection, including improved browsing through additional genre choices.
The subsequent iteration increased the typical number of selected genres from three to five—a 66% increase in preference signals collected at onboarding.
Phase 2: Evolving into Continuous Learning
While onboarding helped personalize the initial experience, it introduced a limitation: viewer preferences could become outdated as interests changed.
Instead of continuing to add questions to onboarding, I helped reframe the problem.
How could every meaningful interaction become an opportunity for Tubi to understand viewers better?
We worked with data science and research to identify two categories of signals: explicit signals such as genres, likes, dislikes, and favorite actors; and implicit signals such as watch completion, session duration, repeat viewing, and engagement with content.
This expanded the design scope from a single onboarding experience to multiple touchpoints across discovery and playback.
The goal was to gather useful signals without making personalization feel like additional work for the viewer.
Key Design Decisions
Genre Selection vs. Title-Level Preferences
Genre selection was a useful starting point, but individual movies and shows could reveal more nuanced preferences, including cast, release period, and tone.
I explored a swipe-based onboarding experience where users could react directly to recognizable titles. The approach turned preference collection into a more engaging interaction while generating richer signals.
The iOS experience achieved a 79% completion rate while enabling viewers to provide feedback on ten or more titles.
Collecting Feedback vs. Preserving Browsing Flow
We explored preference collection within the home browsing experience, including dedicated feedback rows.
However, inserting additional interactions into content browsing risked distracting viewers from their primary goal: finding something to watch.
I shifted the emphasis toward contextual feedback opportunities, including visible like and dislike actions on title detail pages, where viewers could express preferences with less disruption.
Feedback Depth vs. Interaction Friction
Player exit offered another valuable opportunity to collect feedback after viewers had experienced content.
I explored three interaction patterns: a bottom sheet, a persistent toast, and a timed toast.
The timed toast achieved a 68% response rate, outperforming the persistent toast at 61% and bottom sheet at 44%.
This reinforced the value of lightweight feedback interactions that respected the viewing experience.
Designing the Personalization Experience
The resulting experience connected preference collection with moments already familiar to viewers.
Onboarding — Establish initial interests
Viewers selected genres or reacted to titles to provide the recommendation system with an initial understanding of their preferences.
Home & Discovery — Reinforce relevance
Personalized content recommendations used these signals to help viewers discover titles aligned with their interests.
Title Details — Capture explicit feedback
Like and dislike interactions created opportunities for viewers to refine their preferences while evaluating individual movies and shows.
Player Exit — Learn from consumption
Contextual feedback patterns captured reactions after playback, complementing behavioral signals such as completion and viewing duration.
Together, these experiences helped establish a continuous feedback loop between viewer preferences, viewing behavior, and content discovery.
The approach also extended to connected TV, where interaction patterns needed to accommodate remote-based navigation and different viewing contexts.
Outcomes & Business Impact
Across approximately two years of personalization work, Tubi's measured results demonstrated improvements in activation, discovery efficiency, and retention.
Activation improved. The 2D7 Aha Moment conversion metric increased by 16%, indicating improved progression toward early engagement.
Discovery became faster. Time to content declined from more than seven minutes to approximately 3.5 minutes, reducing the effort required to find something to watch.
New-user retention improved. Churn declined from 56% to 39%, representing a 17-percentage-point reduction.
At the feature level, onboarding completion reached 79% in the later title-rating experience, while the strongest player-exit feedback pattern achieved a 68% response rate.
Beyond individual metrics, the work established reusable preference-collection patterns across onboarding, browsing, content details, and playback.
These improvements contributed to Tubi's broader growth efforts during a period when monthly active users expanded from 19 million to 88 million.
Reflection & Learnings
This project changed how I approached personalization as both a product designer and a strategic partner.
Our initial focus was improving onboarding. But the deeper opportunity was building a system that could continuously understand viewers as their interests evolved.
The most important lesson was that collecting more data isn't inherently better. Successful personalization depends on capturing meaningful signals at the right moments while preserving the experience users came for.
Over time, I learned to connect product strategy, behavioral data, experimentation, and interaction design into a system that could evolve beyond any single feature.
The biggest shift was from asking viewers what they liked once to building an experience that could keep learning from them.