How this page is built.

We read the top 25 of YouTube's most-popular chart for the United States and take the first five that pass a fixed filter. A video is skipped if it is a live broadcast, if comments are turned off, if it has fewer than 200 comments, if it is under a minute, if its language is set to anything other than English, or if it has already had its day here.

Nothing is chosen editorially and nothing is chosen for how it will read. When fewer than five survive, the edition runs short and says so.

For each video we fetch up to a thousand top-level comments. Sentiment, recurring themes, emoji counts, comment timing, and the automation estimate are all computed in ordinary code on our own servers, not by a language model. That is a deliberate choice: it is free, it is fast, and it gives the same answer twice for the same input.

A model is used for exactly two things per video, because neither is possible in code. It watches the first twelve minutes of the video to say what the video is. And it picks the funniest comment from a shortlist of about 25 that code has already narrowed down by likes, length, and cleanliness. Everything else it is shown is a summary of numbers we computed first.

Sentiment is scored with a weighted word list written for comment sections specifically, including slang a general-purpose list gets backwards: “sick”, “insane” and “goat” are praise here, “mid” and “cooked” are not. Negation flips a score, intensifiers scale it, shouting amplifies it, and emoji count.

It cannot detect sarcasm. It only reads English. On a single comment it is noisy and sometimes simply wrong. Across a thousand comments the direction is meaningful, which is the only level the page reports it at.

The automation figure is an estimate of how much of a comment section did not look like a person typing. It weighs observable signals: identical text posted by several different accounts, known scam and engagement-farm phrasing, links and phone numbers, hidden characters used to dodge filters, very new accounts, accounts with no uploads, and comments posted within minutes of upload. No single signal is enough on its own, because plenty of real people have new accounts and auto-assigned handles.

We never label a named account a bot. Calling a specific person automated is a factual claim about a human being, made by a heuristic with no way to verify it. So per-comment scores are never stored, never published, and never attached to a name. Only the aggregate is reported, and any comment that scores as possibly automated is excluded from being quoted at all.

Comments are analysed in memory and thrown away. What we keep is our own analysis, which contains no comment text and no author names, plus at most six featured comments per video, held so the page can quote them.

YouTube's API terms require stored data to be refreshed or deleted within 30 days. Every day the pipeline re-checks any stored comment or statistic older than 25 days against YouTube. If a comment has been deleted there, its text is deleted here and the page says so in its place. Our own commentary about it stays, because that part is ours.

Video titles, thumbnails, statistics and comment text come from the YouTube Data API and belong to their creators and authors. Every comment and video on this site links back to its original. Sentiment, themes and automation figures are PAID LLC's own estimates and are not YouTube metrics.

PAID LLC is not affiliated with or endorsed by YouTube or Google.

If your comment is quoted and you would rather it were not, or if something on this page is inaccurate, write to us and it will be removed or corrected. Deleting the comment on YouTube also removes it here within a day.