Every comment-generation system trains on like counts as though a like count measures the comment. It mostly measures when the comment arrived. Checking that on YouTube was impossible with public data: the two large comment corpora, YT-30M and YTCommentVerse, record like counts and no timestamps. This corpus supplies the missing field, and two analyses show what it makes answerable. The dataset is on Hugging Face under CC BY 4.0.

Abstract

Public YouTube comment corpora carry like counts without the time each comment was posted. That omission is load-bearing rather than incidental: a comment posted in a video’s first hour is seen by orders of magnitude more people than one posted a week later, so a like count read without a timestamp confounds what a comment said with when it arrived. Models trained on those corpora fit the head start with no way to see it. This corpus records what was missing: 1,345,353 top-level comments from 633 fully-enumerated videos across 34 channels, with post time, fetch time, reply count, video statistics, and the position YouTube’s own ranking gave each comment. Two analyses show what the timestamps buy. Decomposing what orders a comment’s likes, exposure accounts for 83.5% of the ordering skill available above chance, against 8.2% for what the comment says (16.5% if comment length counts as content rather than exposure); the conclusion holds under either convention and only the headline number moves. A 384-dimensional sentence encoder adds 0.0015 AUC over a bag of words, so the small content effect is not an artifact of a weak text model. Absolute predictability is low throughout, with the best model reaching AUC 0.671, so exposure beats content while both lose to variation nothing here measures. A reimplementation of the retrieval-based engagement estimator used to certify comment generators is then found unfit for the purpose on two counts: it orders real comments at Spearman 0.050 against 0.19 for a model knowing only post time and video size, and its headline ratio reverses direction depending on whether it is summarised with a mean or a median at n=480.

The data

1,345,353 top-level comments, 633 videos, 34 channels, spanning science, tech, gaming, commentary, music, geopolitics and cooking. Every video is enumerated completely rather than sampled, 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.

One property matters for anyone reusing it: 82.1% of comments have zero likes and the top 1% hold 94.9% of all likes. Any summary that averages raw like counts will report no effect where a large one exists.

Cite

@misc{liou2026commentexposure,
  title  = {YouTube Comment Exposure: 1,345,353 Comments with Post Timestamps},
  author = {Liou, Brian H.},
  year   = {2026},
  note   = {https://brianhliou.com/publications/youtube-comment-exposure/}
}