# How common are YouTube outliers: method and data (Breakdown Theory, September 2026)

Free to reuse with credit to Breakdown Theory (breakdowntheory.com). Measured 25 September 2026.

## What was measured
- **Outlier score**: a video's views divided by the median views of the ten same-format uploads
  published before it on the same channel. The median is the middle value once those ten are sorted.
  5x or more is an outlier; 10x or more is a big outlier. This is the same calculation as the free
  YouTube outlier checker (breakdowntheory.com/tools/youtube-outlier-checker) and the channel analyzer
  (breakdowntheory.com/tools/youtube-channel-analyzer).
- **Formats never mix**: a Short (3 minutes or less, 180 seconds) is only compared with the channel's other
  Shorts, a long video only with its long videos. Live and upcoming streams are left out.
- **Judged videos**: a full ten-video baseline, at least 14 days old, and among the channel's newest 50
  uploads of that format. Only judged videos count in the rates. A channel needs 20 judged videos in a
  format to be counted for that format.
- **Channels with at least one outlier**: judged over each channel's newest 30 scored videos, so a channel
  with a longer history does not get more chances.

## How channels were found
60 YouTube searches (10 per niche: tech, finance, science, gaming, food, fitness), mixing broad and
long-tail queries, found 2,125 channels. 1,939 were eligible (a public, not hidden, subscriber count and 40 or more
uploads). Up to 20 were drawn at random per niche and size band (under 10k, 10k to 100k, 100k to 1M,
over 1M subscribers), 472 in total. Each channel's newest 150 uploads were read with the YouTube Data API.

## Uncertainty
95% ranges come from resampling channels 2,000 times, never individual videos, because videos from one
channel move together.

## What happened after an outlier
The median views of the channel's next five uploads (each at least 14 days old), divided by the median of
the 20 uploads before the outlier (`after_ratio_20` in videos.csv; the analyzer's ten-video version is `after_ratio_10`). The same measure after an ordinary video (0.8x to 1.25x) from the same
channels is the placebo: it captures plain channel growth.

## Titles
16 title features, 8 measured by code (numbers, question marks, capitals, first person, brackets, length)
and 8 judged by an AI title classifier (kind of video, whether the title holds back its answer, whether it
promises something extreme). Each is compared within the same channel (Mantel-Haenszel risk ratio,
stratified by channel) and only called a pattern when its 95% range excludes 1x and it points the same way
for channels under and over 100k subscribers.

## Limits
- Search favours channels that already rank. The direction of that bias on the rates is not known.
- Public view counts include paid promotion, which public data cannot identify.
- Impressions, click-through rate and retention are private and not used.
- Older outliers keep collecting views, so scores rise over time.

## Known oddities in the data
- One channel ("Personal Finance", finance) reports 0 subscribers through the YouTube Data API (a public count, not a hidden one) while
  having 32 judged long videos. It is kept exactly as the API returned it.

## Files
- `videos.csv`: one row per judged video (27,985 rows). Titles are not included. Columns:
  - `video_id`, `channel_id`, `niche`, `size_bucket`, `format` (long or short), `published`, `views`
  - `baseline_median`: the median views of the ten same-format uploads before the video
  - `outlier_score`: views ÷ baseline_median
  - `in_rates`: 1 if the channel has 20 or more judged videos in that format, so the video counts in the
    published rates (16,331 long videos and 10,386 Shorts)
  - `after_ratio_10`: for outliers only, the median of the next five uploads ÷ the ten-video baseline. This is
    the "after" figure the free channel analyzer shows
  - `after_ratio_20`: the study's "after" measure for long videos: the median of the next five uploads (each
    at least 14 days old) ÷ the median of the 20 uploads before. Empty when a video has fewer than 20 earlier
    uploads or fewer than three qualifying later ones
  - `role`: "outlier" (5x or more) or "ordinary" (0.8x to 1.25x, the placebo) for long videos in the rates
- `channels.csv`: one row per channel with 20 or more judged long videos (channel, niche, size_bucket,
  subscribers, judged_long_videos, outliers_5x, outliers_10x).

## Reproducing the headline numbers
- Share at 5x: long rows with `in_rates` = 1 and `outlier_score` ≥ 5, divided by all long rows with
  `in_rates` = 1 (8.3%). The same for Shorts (7.1%).
- What happened after: take rows with `role` = "outlier" and a value in `after_ratio_20`. The share under
  1.5 is "back to normal" (53.4%), 1.5 to under 3 is "lifted" (22.2%), 3 or more is "moved" (24.4%). Repeat
  with `role` = "ordinary" for the placebo (80.5%, 14.6%, 5.0%).
- The 95% ranges come from resampling channels 2,000 times and will vary slightly with a different seed.