Comment-generation systems train on like counts as though a like count measures the comment. It mostly measures when the comment arrived. The two large public YouTube comment corpora, YT-30M and YTCommentVerse, record like counts and no timestamps, so that could not be checked on public data. This corpus supplies the missing field, in three languages, and adds a panel that watched the top comments form on new videos. The dataset is on Hugging Face under CC BY 4.0. Two write-ups cover it in plain terms: the first version’s result and the panel.

Abstract

Public YouTube comment corpora carry like counts without the time each comment was posted. A like count read without a timestamp mixes what a comment said with when it arrived, and models trained on those corpora fit the head start with no way to see it.

This corpus records what was missing, now at two scales and in three languages. A bulk corpus holds 7,152,754 top-level comments from 6,218 fully enumerated videos on 311 English, Japanese and Chinese-language channels, each with post time, fetch time and the position YouTube’s ranking gave it. A panel followed 500 new videos twice a day for 21 days and stored the ranked first page at every look: 20,454 page fetches and 1,462,289 ranked rows. The first version’s 1,345,353-comment English snapshot is kept.

What orders a comment’s likes replicates in all three languages: exposure accounts for 86–95% of the above-chance ordering (version 1, English: 83.5%), and text alone stays below AUC 0.60 on the registered rungs everywhere except Chinese (0.602).

The panel, pre-registered before its second observation, shows who reaches the top and when. 84% of the day-7 top 20 were posted within a day of upload. The page keeps moving after that: median lock-in comes at 76 h against a predicted limit of 72, and consecutive looks half a day apart keep about 0.80 top-10 overlap from day 4 to day 21.

The panel also tests the feedback loop version 1 stated as fact. Per hour, a comment gains 1.85× as many likes at rank 1–10 as at rank 51 or below (CI 0.96–3.22), short of the predicted 3×, and the median comment’s own ratio is 0.90. Three of eight pre-registered predictions hold as written.

The data

Three parts, each on Hugging Face with a manifest of row counts and file hashes:

part what it holds videos channels
Bulk corpus 7,152,754 top-level comments: 5.42 million English, 1.13 million Japanese, 606,046 Chinese 6,218 311
Panel 1,462,289 ranked rows from 20,454 fetches of the first page, twice a day to day 21 500  
Version 1 snapshot 1,345,353 English comments 633 34

Every video is enumerated completely, because commentThreads.list returns newest-first and any page cap would drop the oldest comments, which are the high-exposure ones. Author channel IDs are salted hashes; the dataset card is explicit that this is pseudonymity rather than anonymity, since comment text stays public and searchable.

How skewed the likes are depends on the language. In English, 78.5% of comments in the bulk corpus have zero likes and the top 1% hold 92.4% of all likes; in Japanese the figures are 44.1% and 60.6%. Any summary that averages raw like counts will report no effect where a large one exists.

The collector, the pre-registered plan with its dated amendments, the analysis code and the paper’s source are on the v2 branch on GitHub under MIT. Version 1 of the paper (August 2026) stays available; version 2 lists what it changed and why.

Cite

@misc{liou2026commentexposure,
  title  = {YouTube Comment Exposure: Timestamped Comments and a 21-Day Panel of the Ranking},
  author = {Liou, Brian H.},
  year   = {2026},
  note   = {Version 2.2. https://brianhliou.com/publications/youtube-comment-exposure/}
}