YouTube
The algorithm is the product. And it's defaulting to the wrong answer.
YouTube's long-form viewership dropped 90% in three months after a single algorithm update. Users didn't stop wanting long-form content — YouTube stopped showing it to them. I researched the platform's structural recommendation failure across 71 survey respondents and 3 user interviews, mapped the behavioral gaps through competitive analysis against Netflix, TikTok, Spotify, and Prime Video, and designed a 4-layer AI-powered Preview Engine to bridge the gap between Shorts and long-form without cannibalising either.
Everyone is looking at the wrong problem
The standard framing is: users prefer Shorts now, so long-form is declining. That's true but it leads to the wrong solution. Most YouTube users don't only want Shorts — they want both. The same person scrolling Shorts on their commute would happily watch a 40-minute documentary on a Sunday evening. The algorithm can't tell when a user is in which mode. When it's not sure, it defaults to Shorts, because Shorts produce a faster and cleaner engagement signal. That default is the real problem. It's a product architecture failure, not a user preference problem.
What 71 survey responses and 3 interviews revealed
45% of users cite time constraints as the top barrier to long-form — yet the same users spend 95+ minutes a day on Shorts, equivalent to 5–6 long-form videos. They don't lack time. They lack confidence that a specific video will be worth it. 65% make their stay-or-leave decision within the first 30 seconds. Three user interviews spanning a college student, a homemaker, and an IT professional surfaced the same friction independently: no direct path from an engaging Short to its source video, generic trust signals that don't help, and a home feed that doesn't reflect their actual intent. The pain is structural and consistent across personas.
The AI Preview Engine: try before you commit
The solution is a 4-layer preview system that surfaces quality signals before a user clicks. Layer 1: a 15–30 second silent hover preview that auto-selects the highest-impact moment using retention peaks, replay heatmaps, and comment activity. Layer 2: a user-controlled 30–60 second extended preview with three selectable modes — Quick Gist, Insight Mode, and Drama Mode — tailored to intent. Layer 3: an AI-generated interactive timeline map with clickable chapter entry points. Layer 4: a personalization engine that adapts preview style per user based on watch behavior over time. YouTube already captures per-second engagement data for every video. The infrastructure exists. The viewer-facing feature doesn't.
The 4-layer architecture
15–30 second silent autoplay on thumbnail hover. AI selects the highest-impact moment from retention peak data, replay heatmaps, and comment activity timestamps.
User-controlled 30–60 second extended preview with three intent-matched modes: Quick Gist (summary), Insight Mode (key argument, for learners), and Drama Mode (best moments, for entertainment).
AI-detected chapter markers with clickable entry points to the most replayed, most commented, and most controversial timestamps. Users jump to the part that matters to them.
Preview style adapts dynamically per user based on rewatch behavior, drop-off patterns, and content preferences — the same video shows up differently depending on who's watching.