Fake and bot comments on YouTube — what the signals actually prove.

The question behind this page is usually practical: is this thread worth believing? Maybe you are reading sentiment before a launch, judging a competitor’s engagement, or running a giveaway where the comments decide the result. All three need the same thing — not an accusation, just an honest read.

What a signal can prove

Start with what is not on offer: there is no single feature that proves a comment is a bot. Automated posting mimics human noise well, and humans produce plenty of noise themselves. What manufactured comments cannot hide is the pattern — they arrive together, repeat each other, and come from accounts with nothing behind them. One signal is a hint; several signals pointing the same way, across several accounts, is the actual finding.

This matters in both directions. If you accuse on a single tell, you will mostly accuse real people — a short comment and a strange username describe half of YouTube. The checks below are built to be wrong rarely, which means they need a minute of work rather than a glance.

The patterns worth checking

Five patterns, each with the check that turns it from a feeling into a fact. Run them on an exported thread where all comments sit on one screen — the signals are about order and repetition, and neither is visible while scrolling a page.

1. Bursts in the timelineA real discussion arrives in a ragged stream — minutes apart, at different hours, around the video's actual life. Manufactured comments often arrive in bursts: several appearing within the same minute or two, well after the video was posted.How to check: Export the thread and sort by published_at. Scan for clusters — several comments landing inside a minute, hours or months after upload, is the shape to look at first.
2. Repeated phrasing across accountsThe clearest single signal is language that repeats. Different names, different avatars, and the same sentence — or the same sentence with one word swapped. Real viewers occasionally echo each other; manufactured batches do it for pages at a time.How to check: In the exported file, search a distinctive phrase from one suspicious comment and see how many different accounts used it. Two is a coincidence; six is a batch.
3. Link-first commentsComments whose whole job is a link — a channel plug, a crypto pitch, a “check my profile” — are the most common manufactured comment, because the link is the point. A comment that adds nothing to the conversation and carries a URL belongs in this pile.How to check: Filter the export for http in the comment text and read those comments in order. Judge them as a group, not one by one.
4. Thin or patterned accountsReal accounts have trails: old comments, replies, a history of watching. Many manufactured ones have one comment, no photo, and a name following a template — a first name plus four digits, or a string that means nothing.How to check: Open the profile of a suspicious commenter and read their comment history. An account whose entire public activity is promo comments is the fact you are looking for.
5. Engagement that outruns substanceLikes can be bought, so a deeply generic comment — “Great video!” — sitting at 400 likes deserves a second look, especially when the comments around it say more and have fewer. The gap between what a comment says and how it is received is itself a signal.How to check: Compare like counts against content for the top of the section. If the pattern looks off, the like counts themselves are worth reading skeptically — see the most-liked page for how that order works.

Tells that mean nothing

These four show up in every “how to spot a bot” list, and every one of them describes ordinary people. Do not let any of them, alone or together, decide anything:

• Imperfect spelling and grammar — real people write that way, constantly, in every language.
• Short replies like “nice” — most of the internet is brief, and brevity is not a bot signature.
• Non-native English — clear meaning with odd phrasing is a person doing their best, not a machine.
• A single generic compliment — one “love this” proves nothing; look for the company it keeps.

A calm one-pass workflow

The whole check fits into one pass. Export the thread, open it in a spreadsheet, and sort by time — the column layout puts every comment’s moment and every account’s name in one view. Scan the times for bursts, then search one distinctive phrase from any suspicious comment and count the accounts behind it. Open the profiles of the accounts that recur.

Then decide what the finding is for. If it is your video, you have moderation tools: hold comments for review, block words, hide users — all available in Studio and on the watch page. If it is someone else’s, report the comments that break policy and move on. And if the thread is being used to settle something — a giveaway, a claim, a record — this is also the moment to check the earliest and the most-liked comments with the same honesty, since those are the two positions a manufactured batch aims for: the first comment and the most liked ones.

Checking a thread

One export and a few minutes of reading beats any gut feeling, in both directions: you will catch the batches a glance would miss, and clear the real people a glance would accuse.

• Read the thread with every reply expanded — The full public thread, so clusters and repeated phrases are visible side by side instead of clicks apart.
• Export it and sort by time — CSV or JSON with published_at and like_count per row — the two columns the check runs on.
• See what the thread actually says — When the goal is the themes rather than the fakes — a report on what the comments are about.