“Trending” feels permanent when you're scrolling it. It isn't. The board churns constantly, and most creators have no idea how short their window really is.
I wanted to treat the Trending page like a dataset instead of a vibe — pull every video on it, every day, across regions, and watch how positions appear, climb and evaporate. With enough snapshots, the lifecycle of a viral video stops being mysterious and starts being a curve you can plot.
I hit the YouTube Data API on a schedule, capturing the full Trending list for ten regions every few hours over several weeks. Each snapshot recorded position, views, likes, comments, category, channel size and publish time.
I de-duplicated videos across snapshots into a single lifecycle per video, then engineered features — title length, presence of numbers and emoji, thumbnail face detection, publish hour, category and channel size. A gradient-boosted model did the heavy lifting on “what predicts longevity,” with the raw curves kept close by as a sanity check.
Insight one: the lifecycle is front-loaded and brutal. Velocity in the first two hours decides almost everything.
“The thumbnail and first two hours matter more than the channel's entire subscriber count.”
Insight two: clout is overrated. The signals creators obsess over barely move the needle.
When I ranked features by how much they predicted longevity, early velocity and a few cheap title-and-thumbnail tricks dominated. Subscriber count — the thing everyone chases — landed near the bottom.
- 01Run sentiment on titles and thumbnails to test whether curiosity gaps really do outperform plain description.
- 02Compare lifecycle curves across regions — does a video trend longer in some countries than others?
- 03Ship a tiny “trending odds” tool that scores a draft title and thumbnail before you hit publish.