Going viral has a shelf life. It is about 38 hours.

Overview

"Trending" feels permanent when you are scrolling it. It is not. 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.

The data

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.

Approach

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.

What the data says

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.

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.

What is next

1. 01Run sentiment on titles and thumbnails to test whether curiosity gaps really do outperform plain description.
2. 02Compare lifecycle curves across regions to see whether a video trends longer in some countries than others.
3. 03Ship a small trending-odds tool that scores a draft title and thumbnail before you hit publish.