The YouTube Comment Report

The largest empirical study of YouTube creator reply behavior, comment velocity, and viewer intent ever published.

Based on 9,877,019 public comments collected across 15,403 verified channels, with 879,383 sampled comments analyzed in depth using a multi-model AI ensemble (ChatGPT 5.6 Sol, Claude Opus 5, Gemini 3.7). CommentShark planned, paid for, and conducted the study independently on public YouTube data.

Comments
9.88M
9,877,019 public comments in the main count
Channels
15,403
Four channel-size groups
Videos
1.43M
1,432,155 public videos included
Comments analyzed by meaning
975K
975,222 comments; each had a known chance of being chosen
Jump to a chapter23 chapters
  1. 01 · Executive summary
  2. 02 · Study sample
  3. 03 · Unanswered questions
  4. 04 · Creator reply rates
  5. 05 · What creators answer
  6. 06 · What viewers say
  7. 07 · The tag map
  8. 08 · Tone and who is addressed
  9. 09 · What viewers ask
  10. 10 · Content categories
  11. 11 · Size or subject
  12. 12 · Short and long videos
  13. 13 · Business channels
  14. 14 · Languages
  15. 15 · Common comment text
  16. 16 · Duplicates and empty sections
  17. 17 · Comment timing
  18. 18 · Does replying matter
  19. 19 · Age and country
  20. 20 · Benchmarks
  21. 21 · Methodology
  22. 22 · Limitations
  23. 23 · Changelog

Executive summary

These are the main findings from the study. They describe the channels and comments in our sample, not every channel on YouTube.

81.2%

of mature questions in the study sample had no creator reply

Based on a weighted broad channel sample of 61,034 questions across 8,682 channels.

Read the finding

1.1%

median reply rate among eligible sampled channels

The middle reply rate across 9,310 broad-sample channels and 1,404,133 comments at least 7 days old.

Read the finding

10.2%

of sampled comments were real questions

Based on 90,336 sampled questions from 11,805 broad-sample channels.

Read the finding

68.1%

of comments arrived within 48 hours, median sampled channel

The middle result across 2,628 broad-sample channels with full video histories.

Read the finding

11%

repeated text, median channel in the study sample

The middle broad-sample channel. Repeated text does not tell us who or what wrote it.

Read the finding

was the most repeated comment text in the study

22,339 posts across 6,679 unrelated channels in the recent-comment sample.

Read the finding

68.5%

of sampled 1M+ channels replied to none of their mature comments

Counted per channel across 1,236 broad-sample channels with enough comment history. At 1,000 to 10,000 subscribers it was 30.1%.

Read the finding

5.5x

as many replies to money as to debate, study sample

In the broad study sample, buying and partnership comments were answered at 26.8% and debate at 4.9%. Commercial comments received replies most often of any type.

Read the finding

83.4%

of sampled comments were supportive or neutral

Fewer than 1 in 14 was hostile. Comment sections are calmer than their reputation, except in one category.

Read the finding

18%

hostile comments on sampled news channels

Against 3.4% on music channels. No other content group passed 6.1%.

Read the finding

11.7%

of mature sampled questions got a complete answer

18.8% got a reply of any kind. Of the replies that did arrive, 15% did not answer the question.

Read the finding

2.3 vs 0.6

comments per 1,000 views, median sampled channel

The middle broad-sample channel: longer videos compared with videos three minutes or shorter.

Read the finding

5.3% vs 10.5%

reply rate on timestamp comments vs all comments

Viewers who quote an exact moment in the video are among the most attentive, yet receive half the normal reply rate.

Read the finding

What a creator can act on

  1. Review comments inside the first 48 hours65.4% of a video's comments arrive by the end of day one and 81.7% within the week.
  2. Separate real questions from reactions before you startOnly 10.2% of comments are real questions, and a question mark finds the wrong ones both ways.
  3. Judge a reply by whether it answered, not whether it exists18.8% of mature questions got a reply, but only 11.7% got a complete answer.
  4. Compare against your own size and subject, not a platform averageThe unanswered rate runs from 59.1% to 99% depending on which cell a channel sits in.
  5. Do not skip timestamp commentsTimestamp comments are answered half as often (5.3% vs 10.5%), despite coming from attentive viewers quoting exact moments in the video.
  6. Do not assume replying causes growth or loyaltyThe measured lift is explained by the creator's own reply and their own like. This study cannot show an effect.

Each row links to the chapter it rests on. Nothing here is advice the study cannot support, which is why the last row is a caution rather than an action.

The YouTube comment study sample

We studied public YouTube channels that mostly use English and were found through search in August 2026. We split channels into four size groups and kept the two channel samples separate.

Source: final 2026 study database. Data collected August 19 to 22, 2026.

Channels in the study

Number of channels
  1. 1K–10KEmerging3,840
  2. 10K–100KGrowing4,409
  3. 100K–1MEstablished4,504
  4. 1M+Large2,650
Source: final 2026 study database. Counts include both channel samples. Reply and behavior rates stay separate for each sample.

Broad channel sample

The broad sample includes 11,805 channels found through YouTube search. We split them by channel size and eight content groups: Education & How-To, People & Vlogs, Entertainment & Comedy, Gaming, Music, Tech & Science, News & Commentary, and Lifestyle & Interests.

Each channel's group comes from the most common YouTube category in its recent uploads. It does not come from the search that led us to the channel.

Creator-business sample

The creator-business sample includes 3,598 channels in topics such as coaching, finance, health, property, and trades. We split them only by channel size. We did not set a target for each topic, so the topic mix is what the sample held.

These results describe only the creator-business channels we studied. We compare them with the broad sample only when it helps answer a clear question. We never combine the two samples.

Sample or collection methodCountWhat it covers
Broad channel sampleMain study group11,805 channelsMostly English channels found through search and split by channel size and eight content groups.
Creator-business sampleBusiness study group3,598 channelsBusiness and professional channels split by size and kept separate from the broad sample.
Recent commentsCurrent public conversations5.92M comment entriesRecent comments used to measure creator replies, comment structure, and what comments mean.
Full video historiesWhen comments arrived4.02M comment entriesAll available comments for videos that met the study rules and were posted 30 to 90 days before the study.

Some comments appear in both comment groups. That is why there are 9,944,142 study entries but 9,877,019 comments in the main count. We count each comment only once when a measure calls for it. We also keep track of which group each entry came from.

How many YouTube questions go unanswered

Viewers often ask questions in public comments, but creators rarely reply. The gap is large enough for channels to track unanswered questions.

Source: broad channel sample, recent comments, comments at least 7 days old, and channels whose captured comments reached back at least 14 days.

81.2%

Combined unanswered question rate for the broad sample

Based on a weighted broad channel sample of 61,034 questions across 8,682 channels.

From question asked to question answered

Share of questions at least 7 days old
  1. Questions asked100%61,034 sampled questions at least 7 days old, across 8,682 channels
  2. Got any reply18.8%A creator replied to the thread at all
  3. Got an answer15.9%The reply answered the question in full or in part
  4. Got a full answer11.7%The reply resolved the question on its own
Broad channel sample: 61,034 sampled questions at least 7 days old across 8,682 channels, adjusted for each comment's chance of being chosen. Each stage is measured against every question, not against the stage above it.

1 in 8

Roughly one question in eight gets a complete answer. Of the replies that do arrive, 15% do not answer the question at all.

11.7% of questions received a full answer, 4.2% a partial one. The gap between the 18.8% reply rate and the 11.7% full-answer rate is the cost of replies that acknowledge without resolving.

These numbers use the same broad-sample questions, all at least 7 days old. We adjusted the totals because comments had different chances of being chosen. The results describe only channels in the study. We did not calculate a margin of error that accounts for many comments coming from one channel.

A question needs time to receive an answer

We call a question unanswered only after it is at least 7 days old. We chose 7 days based on how long replies took in the study. Of the replies whose timing we could measure, 94.6% came within 7 days. In this study, 95% came within 7.6 days.

We also used only channels whose captured comments reached back at least 14 days. This stops new comments on very busy channels from looking old enough to count as ignored. We needed at least 20 older questions to report a rate for one channel.

What counts as a reply

A reply in this study means an observed, public text response from the channel owner. Private moderation actions such as creator hearts or pinned comments are not exposed in public comment threads and are not counted as text replies.

We sort questions by meaning

We do not decide based on a question mark alone. The fixed model puts each comment into one question type or marks it as not a real question. Indirect questions can count. Rhetorical questions and requests that do not seek an answer do not count on their own.

YouTube creator reply rates by channel size

Creators reply to a smaller share of comments as channels grow. This pattern is clear among channels that met the study rules. But the result for the largest channels has an important limit.

Source: broad channel sample, recent comments, the median across channels in each size group, and comments at least 7 days old.

Creator reply rate by channel size

Percent of comments. Scale: 0% to 10%
1K–10K10K–100K100K–1M1M+
Channel sizeCreator reply rate (%)Study base
1K–10K6.5%2,727 channels in this comparison; 268,859 comments at least 7 days old
10K–100K1.6%2,998 channels in this comparison; 403,824 comments at least 7 days old
100K–1M0%2,631 channels in this comparison; 495,067 comments at least 7 days old
1M+0%954 channels in this comparison; 236,383 comments at least 7 days old
This measures whether a creator replied, not what the reply said. Each channel needed at least 20 comments that were 7 days old, and its captured comments had to reach back at least 14 days. A median reply rate of 0% means at least half of the channels had no reply we could see. It does not mean every channel had a zero reply rate.

What a median reply rate of 0% means

Above 100,000 subscribers, the median channel-level reply rate was 0% for comments at least 7 days old. That is not a missing number. It means at least half of the channels in that group replied to none of those comments in the window we captured.

Bigger channels reply to a smaller share

The median reply rate was higher among channels with 1,000 to 9,999 subscribers than among channels with 100,000 subscribers or more. This shows a link in the sample. It does not prove that gaining subscribers makes creators reply less.

Reply speed is a different measure. We found the median reply time only for comments that got a reply we could see. A channel can reply to few comments but answer those few quickly.

Triage capacity and inbox volume

A median reply rate of 0% on larger channels reflects an inbox triage ceiling as much as creator choice: a channel receiving hundreds of comments a day cannot maintain a 5% reply rate by hand without reviewing thousands of posts a week. What diminishes as channels grow is manual bandwidth.

The largest-channel result has a limit

The broad sample had 2,251 channels with at least 1 million subscribers. Of those, 1,292 hit the 1,200-comment limit before their comments reached back 14 days. Only 954 met every rule for this chart. The channels left still tell us something, but they may not fairly represent the busiest channels.

68.5%

of channels with at least 1 million subscribers replied to none of their comments older than a week. Among channels with 1,000 to 10,000 subscribers, 30.1% replied to nothing.

Counted per channel, not per comment. Every channel here had at least 20 comments that were 7 days old, and captured comments reaching back 14 days.

Channel sizeChannels comparedReplied to nothingReplied to more than 1 in 10
1K–10K2,74830.1%45.3%
10K–100K3,12240.9%31.4%
100K–1M2,92253.4%16.7%
1M+1,23668.5%7.4%

The two right-hand columns do not add to 100%. The channels in between replied to some comments but fewer than one in ten. Together the columns show the shape a single middle number hides: replying is close to all or nothing, and the share of channels doing nothing climbs steadily with size.

49%

of the variation in observed reply outcomes is estimated to sit between channels, rather than among comments inside the same channel.

Estimated across 463,849 comments on 9,388 channels, with a channel-bootstrap range of 0.476 to 0.501. A second binary estimator gives 0.487. This describes clustering; it does not say that the channel caused any individual reply.

3.7 hours
Middle reply time
Measured on replies the study could time
31.2%
Replies within the hour
Of the replies that arrived at all
81.9%
Replies within a day
Replies come quickly or not at all
49.5 hours
Slowest 1 in 10 replies
255,207 timed creator replies

Reply speed and reply rate answer different questions. Creators in this sample are not slow to respond. They respond to very little, and what they do respond to they usually handle the same day.

When a reply arrives, if it arrives

Share of the replies the study could time
  1. Under 15 minutes15.5%
    n=39,521
  2. 15 to 60 minutes15.7%
    n=40,183
  3. 1 to 6 hours26.1%
    n=66,710
  4. 6 to 24 hours24.5%
    n=62,576
  5. 1 to 3 days10.7%
    n=27,346
  6. 3 to 7 days4.3%
    n=10,872
  7. More than 7 days3.1%
    n=7,999
Broad channel sample: 255,207 timed creator replies. The distribution has almost no long tail: only 3.1% of replies arrive more than a week after the comment.

Read this as a shape, not as a chance. It covers only the comments that were answered, so it says how fast those replies came, not how likely a reply was. The next figure asks the harder question.

The chance a comment has been answered by a given moment

Counting the comments that were never answered, not just the ones that were
1 hour6 hours24 hours3 days7 days14 days30 days
ElapsedAnswered by then
1 hour2.54%
6 hours4.77%
24 hours7.01%
3 days8.1%
7 days8.64%
14 days9%
30 days9.56%
All captured comments at least 7 days old in the broad sample: 1,684,513 comments, of which 153,509 were answered and 90.9% were still unanswered when we captured them. A comment that is still waiting is not a missing observation, so it is carried as unanswered-so-far rather than dropped.

73%

of the replies observed by day 30 had arrived in the first 24 hours. After the first week the curve is almost flat.

The chance of an answer climbs to 7.01% by day one and only reaches 9.56% by day 30. The study does not observe what happens after that point.

The chance of being answered within a week, by channel size

Shading runs within each column
Channel sizeCommentsAnswered within a dayAnswered within a week
1K–10K277,59915.18%18.24%
10K–100K456,3229.43%11.7%
100K–1M613,6744.45%5.6%
1M+336,9181.65%2.11%

Darker means a higher chance of an answer. Comments are counts, so they are not shaded.

Same method and population as the chart above, split by channel size. This is the figure to quote to a viewer: it is the chance their comment gets a reply within the stated time, not the rate at which a creator replies.

A comment on a channel with 1,000 to 10,000 subscribers has an 18.24% chance of an answer within a week. The same comment on a channel above a million has 2.11%, which is roughly one in fifty.

What creators reply to, and what they skip

Reply rates are not even across comment types. Sorting every comment at least 7 days old by what it was trying to do shows a clear order of priority.

Source: broad channel sample, recent comments, comments at least 7 days old from channels with enough comment history. Adjusted for each comment's chance of being chosen.

How often a creator replies, by what the comment is doing

Share of comments at least 7 days old that received a creator reply
  1. Buying or business request26.8%1% of comments, 2.6% of all replies
  2. Asking for information17.7%10.7% of comments, 18% of all replies
  3. Personal story or result14.1%6.6% of comments, 8.9% of all replies
  4. Request for the creator13.6%3.6% of comments, 4.7% of all replies
  5. Thanks or reaction12%39.2% of comments, 45.1% of all replies
  6. Feedback or correction10.6%6.6% of comments, 6.7% of all replies
  7. Talking with other viewers9.5%0.4% of comments, 0.4% of all replies
  8. Spam or self-promotion6.2%2.1% of comments, 1.2% of all replies
  9. Discussion or debate4.9%16.1% of comments, 7.5% of all replies
  10. Insult or attack4.6%4% of comments, 1.8% of all replies
  11. Other3.3%9.7% of comments, 3.1% of all replies

All comments at least 7 days old: 10.5%

Broad channel sample: 463,849 sampled comments at least 7 days old. Each comment has one main purpose. A reply means a creator answered in that thread, not that the reply was useful.

5.5x

A comment about buying or working together is 5.5 times as likely to receive a reply as a comment joining a debate, and 2.2 times as likely as a thank-you.

26.8% against 4.9% and 12%. These are rates, not volumes. Buying and business comments are only 1% of all comments.

Rates are not the same as attention

Thanks and reactions get a below-average reply rate, but they are 39.2% of all comments. That volume means 45.1% of every creator reply we saw went to a comment that was only saying thanks. Questions took 18% of replies.

So the picture is not that creators ignore questions on purpose. It is that the comments easiest to answer are also the most common, and a high volume of them may consume the attention available.

Debate and abuse get the fewest replies

Discussion and debate is the second largest group of comments at 16.1%, but it draws only 7.5% of replies. Insults and attacks draw 1.8%. Most creators in this sample simply do not engage with either.

This is a measure of what creators did, not of what they should do. A creator may reasonably decide that an argument in the comments is not worth entering.

What a creator reply actually says

Share of readable creator replies
  1. Substantive43.1%
  2. Brief acknowledgment26.3%
  3. Generic template25.9%
  4. Other3.9%
  5. Defensive or hostile0.8%
Broad channel sample: 80,208 sampled creator replies. A substantive reply adds information or a personal response. A brief acknowledgment shows the creator saw the comment. A generic template could answer almost any comment.

Slightly more than half of the replies we could read added nothing beyond acknowledgement. That matters for the question chapters: a reply is not the same thing as an answer.

17 vs 6

Words in the average creator reply to a question, against the average reply to a thank-you.

Creators triage twice: once on whether to answer at all, and again on how much to write. A correction gets the longest reply in the study at 18.7 words.

How much a creator writes back, and how long they take

Shading runs within each column
What the comment was doingWords in the replyAverage hours to replyReplies counted
Feedback or correction18.772.5h5,196
Asking for information17.280h14,684
Personal story or result17.158.1h6,999
Buying or business request16.2103.7h2,143
Discussion or debate15.855h5,559
Insult or attack13.167.9h1,415
Request for the creator11.451.7h3,577
Talking with other viewers8.238.1h300
Spam or self-promotion7.174.1h982
Thanks or reaction5.868.3h36,844
Other572.7h2,521

Darker means higher within that column. Averages, on the replies the study could read and time.

Broad channel sample: 80,220 sampled creator replies with a word count. Buying and business requests wait longest at 103.7 hours on average, which is the one place where the fastest-answered and the most-answered comment types come apart.

Do longer comments get answered more

Creator reply rate by comment length
  1. 1 to 3 words8.9%
    n=149,351
  2. 4 to 10 words10.3%
    n=158,830
  3. 11 to 25 words11.4%
    n=98,977
  4. 26 to 50 words12.9%
    n=36,025
  5. More than 50 words12.3%
    n=20,666
Broad channel sample: 463,849 sampled comments at least 7 days old. The effect is real but small, and it stops rising past 50 words.

Writing more helps a little

A comment of 26 to 50 words is answered 12.9% of the time against 8.9% for one of three words or fewer. The climb is steady and then stops: past 50 words the rate falls back to 12.3%.

Except when you cite a timestamp

2.4% of comments quote a moment in the video. They are answered 5.3% of the time against 10.6% for everything else, which is half the rate.

These are among the most attentive comments a channel receives, and they are the least likely to get a response. Based on 9,769 sampled comments citing a timestamp.

What YouTube viewers say in comments

For this study, each comment was put into one main group. A comment may react, ask for information, make a request, share an experience, give feedback, or join a discussion.

Source: broad channel sample and recent comments. The sample was adjusted for each comment's chance of being chosen.

What a YouTube comment is trying to do

Estimated share of comments
  1. Thanks or reactionPraise, thanks, or reaction36%
    297,038 selected comments
  2. Discussion or debateOpinion, analysis, or disagreement20.7%
    120,035 selected comments
  3. OtherNo other main group fit better10%
    73,750 selected comments
  4. Asking for informationInformation, explanation, or advice8.6%
    77,799 selected comments
  5. Feedback or correctionCritique, complaint, or correction6.3%
    46,359 selected comments
  6. Personal story or resultExperience, result, or disclosure5.7%
    47,608 selected comments
  7. Insult or attackInsult, trolling, or attack5.6%
    33,255 selected comments
  8. Request for the creatorRequest to make, cover, or change something3.8%
    25,651 selected comments
  9. Spam or self-promotionSpam, unrelated ads, or posts made only to get reactions2.1%
    15,846 selected comments
  10. Buying or business requestBuying, sponsorship, or partnership0.7%
    7,378 selected comments
  11. Talking with other viewersDirect interaction with other viewers0.5%
    3,091 selected comments
Broad channel sample: 747,810 comments chosen from 11,805 channels. After weighting, they represent 5,011,893 recent comments. Each comment gets one main purpose, so the shares add to 100%.

Reactions are only part of the story

Thanks and reactions are the largest group. But viewers also ask for facts, share stories, make requests, correct errors, and debate. A list sorted only by newest first treats a thank-you, a product question, and a correction as if all three need the same kind of reply.

Purpose is not the same as positive or negative

The model also marked each comment's point of view. In a separate score, it estimated how much the comment might help or hurt the creator. This is not a positive-or-negative score, and we do not report it that way.

A model sorted the comments by meaning. We checked how often it gave the same answer when run again. But we did not test it against a separate set labeled by people. The method section explains this limit.

How much use a comment is to the creator

Estimated share of comments
  1. Nothing much either way70.1%
  2. Useful to the creator20.2%
  3. Costly to the creator9.7%
Broad channel sample: 747,810 sampled comments. This is a value axis, not a mood one. A friendly comment that gives the creator nothing to work with sits in the middle band.

Seven comments in ten are neither use nor trouble

The model scored every comment for how much it helps or costs the creator, and 70.1% landed in the middle. Only 20.2% carried something useful and 9.7% carried a cost.

This is not a positive-versus-negative score. A warm thank-you with nothing else in it is neutral on this axis, because there is nothing for the creator to act on. That is exactly why it is worth measuring separately from tone.

36%
Thanks or reaction
The single largest group of comments
10.2%
Are real questions
Determined by meaning, not a question mark
20.2%
Are useful to the creator
Something to act on, answer, or learn from
0.7%
Mention buying or working together
Rare, and the most answered kind of comment

Nineteen things a YouTube comment can be

Alongside its main purpose, every comment carries any number of intent tags. A comment can be praise and a testimonial at the same time. This is the finest-grained view the study has of what viewers are doing.

Source: broad channel sample and recent comments. Tags are not exclusive, so they do not add to 100%. Reply rates use comments at least 7 days old from channels with enough comment history.

How common a comment is, against how often it gets answered

Each dot is one intent tag
How answered: out of every 100 of these, how many got a creator replyCreator reply rate
How common: out of every 100 comments, how many carry this tagShare of all comments
LabelShare of all commentsCreator reply rate
Praise17.16%17.5%
Question10.22%18.8%
Humor9.8%6.2%
Debate5.93%6.6%
Hostile5.88%4.7%
Emotional impact3.61%10%
Complaint3.38%8.8%
Topic request2.61%14.8%
Correction1.9%9.3%
Constructive feedback1.67%15.8%
Testimonial1.24%22.2%
Engagement bait1.21%7.2%
Disappointment1.09%8.9%
Long-time fan0.78%20.3%
Buying intent0.77%28.9%
Spam0.63%8.4%
Confused0.45%10.1%
Off topic0.42%2.7%
Partnership request0.14%16.6%
Broad channel sample: 747,810 sampled comments for prevalence, 463,849 comments at least 7 days old for reply rates. The two measures are close to unrelated, which is the point of the chart.

28.9%

Comments showing buying intent are answered more often than any other tag, and they are the second rarest thing in the comment section at 0.77%.

Humor is the mirror image: 9.8% of all comments, answered 6.2% of the time. The most common tags are not the most answered ones.

Every tag, by how common and how answered

Shading runs within each column
TagShare of commentsCreator reply rate
Praise17.16%17.5%
Question10.22%18.8%
Humor9.8%6.2%
Debate5.93%6.6%
Hostile5.88%4.7%
Emotional impact3.61%10%
Complaint3.38%8.8%
Topic request2.61%14.8%
Correction1.9%9.3%
Constructive feedback1.67%15.8%
Testimonial1.24%22.2%
Engagement bait1.21%7.2%
Disappointment1.09%8.9%
Long-time fan0.78%20.3%
Buying intent0.77%28.9%
Spam0.63%8.4%
Confused0.45%10.1%
Off topic0.42%2.7%
Partnership request0.14%16.6%

Darker means higher within that column. Compare the two columns: a dark cell in one is often pale in the other.

Broad channel sample. A comment can carry several tags, so the share column does not add to 100%. Sorted by how common the tag is.

The tags answered most often

Buying intent at 28.9%, testimonials at 22.2%, and long-time fans at 20.3% are the three most answered tags. All three are rare, and all three are the kinds of comment a creator has a reason to want.

Constructive feedback at 15.8% is answered nearly twice as often as a plain correction at 9.3%, which suggests creators respond to how criticism is framed as much as to what it says.

The tags answered least often

Off-topic comments are answered 2.7% of the time, the lowest of any tag. Hostile comments follow at 4.7%, then humor at 6.2% and debate at 6.6%.

One tag is worth watching on its own: 0.45% of comments say the viewer was confused. It is small, but it is a direct report that a video did not land, and only 10.1% of those get a reply.

How hostile are YouTube comments really

YouTube comment sections have a reputation for being nasty. In this sample they are mostly not. The model recorded a stance for every comment, and who each comment was speaking to.

Source: broad channel sample and recent comments, adjusted for each comment's chance of being chosen.

83.4%

of comments were supportive or neutral. Fewer than 1 in 14 was hostile.

Based on 747,810 sampled comments from 11,805 broad-sample channels. We can only see comments that stayed public, so this is a floor for what creators receive, not a full count.

The stance of a YouTube comment

Estimated share of comments
  1. Supportive46.2%
  2. Neutral or informational37.2%
  3. Critical but constructive7.9%
  4. Hostile or adversarial6.9%
  5. Mixed or unclear1.9%
Broad channel sample: 747,810 sampled comments representing 5,011,893 recent comments after weighting. Stance is about tone, not about whether the comment was useful.

Removed comments are not in this count

We cannot see comments that were deleted, held for review, or removed before we collected them. Hostility rates here describe what survived in public. The true rate a creator faces is higher, and we cannot say by how much.

Hostility is not spread evenly

News and commentary channels sit at 18% hostile comments. Music channels sit at 3.4%. The category chapter breaks this down. A single platform-wide hostility number hides a gap of more than five times between content types.

How hostile a comment is, by who it addresses

Share of comments in each group that were hostile
  1. To the creatorof these comments were hostile10.2%
    n=270,673
  2. To other viewersof these comments were hostile8.1%
    n=9,770
  3. About the video or topicof these comments were hostile5.9%
    n=378,138
  4. Unclearof these comments were hostile2.2%
    n=89,229
Broad channel sample: 747,810 sampled comments. Read each row on its own: of the comments addressed to the creator, 10.2% were hostile.

Hostility follows the person, not the video

A comment addressed to the creator is nearly twice as likely to be hostile as one about the video: 10.2% against 5.9%. Of every hostile comment in the sample, 47% was aimed at the creator personally.

That is the gap between a comment section that argues about a topic and one that argues with a person. Both exist on the same channel, and only one of them is addressed to someone who has to read it.

Who a YouTube comment is speaking to

Estimated share of comments
  1. About the video or topic55.9%
  2. To the creator31.6%
  3. Unclear11%
  4. To other viewers1.5%
Broad channel sample: 747,810 sampled comments. Each comment is assigned one audience.

Most comments are not addressed to the creator

Only 31.6% of comments speak to the creator. The majority talk about the video or the topic, as if in a room where the creator happens to be present. Just 1.5% speak to other viewers.

This changes how to read the unanswered question figures. Most comments are not waiting for a response. The ones that are waiting are the minority that this report tracks separately.

4.1%
Personal story
n=32,297
1.2%
Vulnerable disclosure
n=9,130
1%
Result or testimonial
n=10,840
0.8%
Loyalty or long-time viewer
n=5,710

Shares of the weighted broad sample. Personal signal categories are exclusive, so they add to the 7% total. A testimonial rate of 1% is small per comment and large per channel: it is roughly one in every hundred comments received.

What YouTube viewers ask creators

Viewer questions are not all the same. Some ask for help. Others ask about a product, a piece of gear, the creator, a future topic, or something that was unclear.

Source: broad channel sample and recent comments. Question types use all real questions in the weighted sample. Unanswered rates use older questions from channels with enough comment history.

Types of questions viewers ask

Estimated share of real questions
  1. Other84.4% unanswered24.9%
  2. Clarification88.3% unanswered15%
  3. How-to or troubleshooting75.1% unanswered14.9%
  4. Gear or source identification81.1% unanswered14.5%
  5. Opinion or discussion86.1% unanswered12.7%
  6. Personal or creator81.4% unanswered7%
  7. Product, price, or purchase72.1% unanswered6.1%
  8. Topic request77.1% unanswered5%
We used 90,336 sampled questions to estimate results for 512,312 questions in the broad sample. For unanswered rates, we used 61,034 questions at least 7 days old from 8,682 channels to estimate results for 173,930 questions. We did not count rhetorical questions or requests that only ask the creator to act.

Different questions need different answers

How-to and troubleshooting questions usually need an explanation. Product and price questions need a clear fact. Gear questions need a name or source. Personal questions may need a firm boundary instead of a detailed answer.

The study labels a reply as a full answer, partial answer, non-answer, or unclear answer. It checks the first creator reply that we saw. A later correction or fuller answer is not included in that label.

A reply does not always solve the question. The report measures both whether a creator replied and whether that reply answered the question.

42.6%

of comments containing a question mark are not asking anything, and 37.4% of real questions contain no question mark at all.

Counting punctuation would have found 11.1% of comments as questions against the 10.2% that really are, and it would have been wrong about which ones in both directions. This is why the study classifies by meaning.

Question marks against real questions

Share of all comments, on the same rows
  1. A question mark and a real question6.40%
    55,498
  2. A question mark but not asking anything4.74%
    29,639
  3. A real question with no question mark3.82%
    34,838
  4. Neither85.04%
    627,835
Broad channel sample: 747,810 sampled comments. Each comment falls in exactly one row. The two middle rows are what a punctuation rule gets wrong.

A punctuation rule would over-count by 42.6% on one side and miss 37.4% on the other. The two errors partly cancel in the total, which is exactly what makes the heuristic look acceptable until you check which comments it picked.

Question typeShare of questionsNo creator replyGot a full answerQuestions counted
Product, price, or purchase6.1%72.1%18%5,240
How-to or troubleshooting14.9%75.1%15.4%13,219
Topic request5%77.1%12.8%2,474
Gear or source identification14.5%81.1%13.6%10,208
Personal or creator7%81.4%12.7%3,910
Other24.9%84.4%8.9%14,182
Opinion or discussion12.7%86.1%8.1%4,167
Clarification15%88.3%6.7%7,634

Sorted by how often the question type went unanswered. Share of questions uses all real questions in the weighted broad sample. The other columns use questions at least 7 days old from channels with enough comment history. Money questions do best and clarification questions do worst, which is the opposite of what effort would predict: a clarification is usually the cheapest question to answer.

YouTube comment sections by content category

Channel size is not the only thing that changes a comment section. What the channel is about changes it just as much. These are the eight content groups the study sampled.

Source: broad channel sample and recent comments. A channel's group comes from the most common YouTube category in its recent uploads, not from the search that found it.

Every measure, by content group

Shading runs within each column, so the darkest cell is that column's highest value
Content groupChannelsQuestionsHostileThanks or reactionSpeaks to creatorCreator reply rateQuestions unanswered
Tech & Science1,29617.6%4.4%26.4%36.3%11.6%79.6%
Education & How-To1,58713.5%4.2%38.2%42.6%16.2%77.3%
Gaming1,38412.1%4.5%30.4%28.9%10.2%82.5%
Lifestyle & Interests1,54610.4%5%38.2%34.3%14%75.9%
People & Vlogs1,5389.4%5.4%41.7%34.4%7.7%86.4%
Entertainment & Comedy1,8628.3%6.1%39.9%28.2%8.1%83.3%
News & Commentary1,1167.1%18%23.3%23.1%4.5%89.7%
Music1,4766.9%3.4%51.6%32.2%9%83.8%

Darker means higher within that column. Channels are counts, so they are not shaded.

Broad channel sample, 747,810 sampled comments. Reply and unanswered rates use comments at least 7 days old from channels with enough comment history. All other columns use every weighted comment in the group.

5.3x

News and commentary channels received hostile comments at 5.3 times the rate of music channels: 18% against 3.4%.

No other content group passed 6.1%. News is not slightly worse than the rest, it is a different environment.

Every genre, one at a time

Channel size moves the reply rate. Genre moves what is in the comment section at all. Pick a genre to see what its viewers write, what they ask, and what its creators write back.

The most answered genre, and the one viewers speak to most directly. A quarter of its questions are how-to problems, and 2.4% of comments report a result the viewer actually got.

13.5%
Questions
4.2%
Hostile
16.2%
Creator reply rate
77.3%
Questions unanswered
14.3
Words per comment
68.2%
Written in English
2.43%
Results or testimonials
0.98%
Buying or business

What the comments are doing

  1. Thanks or reaction38.2%
  2. Discussion or debate15%
  3. Asking for information11.3%
  4. Other9.2%
  5. Personal story or result8.1%

Share of every comment on a Education & How-To channel.

What the questions ask

  1. Other25.5%
  2. How-to or troubleshooting25%
  3. Clarification11.8%
  4. Gear or source11%
  5. Opinion or discussion7.8%

Share of the real questions asked on a Education & How-To channel.

What the replies say

  1. Substantive43%
  2. Generic template29.6%
  3. Brief acknowledgment23.9%
  4. Other3%
  5. Defensive or hostile0.5%

Share of the creator replies the study could read.

Unanswered questions by content group and channel size

Share of questions at least 7 days old with no creator reply
Content group1K-10K10K-100K100K-1M1M+
Education & How-To59.1%69.9%85.3%90.6%
Lifestyle & Interests61.5%67.4%83.5%96.2%
Tech & Science61.7%71.2%85.1%95.1%
Gaming59.8%78.6%90.2%96.8%
Music67.7%78.6%93.5%99%
Entertainment & Comedy71%78.9%86.2%95.9%
People & Vlogs80.5%83.7%87.8%98.5%
News & Commentary82%86.3%95.1%n/a

Darker means more questions went unanswered

Broad channel sample, 61,034 sampled questions at least 7 days old across 8,682 channels. Every published cell has at least 30 channels behind it. News at 1M+ is shown as n/a because too few channels met the rules.

Read the grid across, not down. Every content group gets quieter as channels grow, and the gap between the best and worst cell is 40 percentage points: 59.1% of questions unanswered for small education channels against 99% for the largest music channels.

The six-type hypothesis did not survive a stricter test

We reran the proposed six comment-section archetypes on 11,727 channels, using only each channel's mix of eleven comment functions. The clusters were somewhat repeatable, but they were not cleanly separated. Naming them as six real kinds would make the data look tidier than it is.

0.108
Separation
Silhouette; 1 is clean separation, values near 0 overlap
0.683
Typical stability
Median adjusted Rand index; the lower resample check was 0.439
0.028
Agreement with category
Adjusted Rand index; 0 is chance-level agreement

The useful result is the failed simplification: comment sections vary continuously, and YouTube's declared category explains little of that variation. The report does not publish the six provisional names as a taxonomy.

Does channel size or subject shape a comment section

The report has now cut the same comments two ways. Putting the two cuts side by side answers a question neither could answer alone: which of them is actually doing the work.

Source: broad channel sample and recent comments, adjusted for each comment's chance of being chosen. The same tags, measured across the two axes.

The same tags across channel size

Share of that size group's comments
Tag1K-10K10K-100K100K-1M1M+
Question10.16%11.23%10.52%9.46%
Humor8.17%8.86%9.24%11.16%
Hostile6.08%5.56%6.33%5.57%
Topic request1.78%2.55%2.75%2.7%
Testimonial1.19%1.31%1.46%1.01%
Buying intent0.89%0.88%0.86%0.61%
Long-time fan0.45%0.66%0.8%0.9%
Spam0.63%0.64%0.58%0.66%

Darker means higher. Shared scale with the grid below, so the two can be compared directly.

Broad channel sample: 747,810 sampled comments. Read across each row: the colour barely changes.

The same tags across content group

Share of that content group's comments
TagEducationTechGamingLifestylePeopleEntertainmentNewsMusic
Question13.48%17.62%12.13%10.39%9.36%8.29%7.14%6.88%
Humor7.18%8.28%12.48%9.4%10.38%14.2%6.96%6.93%
Hostile3.56%3.59%3.83%4.34%4.57%5.15%15.63%2.98%
Complaint3.29%5.86%3.75%3.96%2.47%2.78%3.82%1.66%
Topic request3.11%2.8%4.17%2.16%2.46%2.86%0.94%2.78%
Testimonial2.83%1.83%1.25%1.82%1.21%0.64%0.38%0.89%
Buying intent1.16%2.8%0.62%0.96%0.59%0.38%0.18%0.32%
Confused0.52%0.82%0.72%0.27%0.38%0.44%0.25%0.28%

Darker means higher. Same scale as the grid above.

Broad channel sample: 747,810 sampled comments. Read across each row: the colour changes a great deal, and one cell is darker than anything in the grid above.

15.6x vs 1.5x

Buying intent spans 0.18% to 2.8% across content groups, a range of 15.6 times. Across channel size it spans only 0.61% to 0.89%, a range of 1.5 times.

Hostility behaves the same way: 5.2 times across subject, 1.1 times across size. Both ratios divide the highest cell by the lowest on the same measure, which is why this report cuts by content group at all.

What size does change

Size moves the things creators control. The reply rate falls from 6.5% to 0% across the four bands. The share of channels replying to nothing rises from 30.1% to 68.5%. The unanswered question rate climbs in every single content group.

One audience measure does move with size: long-time fans rise from 0.45% of comments to 0.90%, and humor from 8.17% to 11.16%. Both are small next to what subject does.

Hostility does not rise with size

A common belief is that bigger channels attract nastier comment sections. In this sample they do not. Hostile comments run between 6.4% and 7.3% at every size, and the smallest channels are not the safest.

The smallest channels do get the warmest comments: 51.9% of their comments are supportive, against roughly 45% everywhere else. What falls away with size is warmth, not safety.

Comment stance by channel size

Shading runs within each column
Channel sizeSupportiveNeutral or informationalCritical but constructiveHostile
1K–10K51.9%33%6.3%7%
10K–100K46.1%37.5%8.1%6.4%
100K–1M44.9%37.3%8.7%7.3%
1M+46.1%37.8%7.4%6.6%

Darker means higher within that column.

Broad channel sample: 747,810 sampled comments. Rows do not add to 100% because mixed and unclear comments are not shown.

Comments on short and long YouTube videos

Shorter and longer videos have different comment patterns. This study uses three minutes as the dividing line.

Source: broad channel sample. Recent comments show comment types and replies. Public video data shows comment volume.

Videos three minutes or shorter get fewer comments per view

Across the size groups, the median channel-level rate was lower on videos lasting three minutes or less than on longer videos. We found the comments-per-1,000-views rate for each channel first, then took the median across channels. This keeps one viral video from setting the result for the whole sample.

This result does not describe every upload on YouTube. It covers videos from the sampled channels and the public information available during the study. Video length also changes with content type, channel size, upload plans, and video age.

Video length does not prove a video was a Short

Public YouTube data gives a video's length. It does not always give its shape or confirm that YouTube put it in the Shorts feed. We therefore say “videos three minutes or shorter,” not verified Shorts.

Measure3 minutes or lessLonger than 3 minutesHow it was counted
Comments per 1,000 views0.62.3Broad sample, median across channels
Questions9.4%11.1%Broad sample, weighted comments
Comments at least 7 days old with a creator reply9.3%11.6%Broad sample, weighted older comments
Reaction or appreciation38.8%33%Broad sample, weighted comments
Personal story or result4.3%7.3%Broad sample, weighted comments

76.3%

of videos published by channels with 1,000 to 10,000 subscribers are three minutes or shorter. Above 1 million subscribers it is 47.1%.

Based on 1,115,749 public videos with a known length in the broad sample. The smallest channels have already moved to short video; the largest have not.

Short video, empty comment sections, and channel size

Shading runs within each column
Channel sizeVideos3 minutes or shorterShort videos with no commentsLong videos with no commentsComments turned off
1K–10K188,75276.3%34.7%26.2%1.09%
10K–100K228,40562.5%20.3%10.4%0.87%
100K–1M346,37651.7%8.2%3.4%0.4%
1M+352,21647.1%2.6%1.4%0.21%

Darker means higher within that column. Videos are counts, so they are not shaded.

Broad channel sample, public videos with a known length. A zero-comment video had public comment numbers and no comments. Videos with comments turned off are counted separately and left out of the zero-comment columns.

A short video is more likely to end up with an empty comment section than a long one at every channel size, and the gap is widest where it hurts most: on the smallest channels, 34.7% of short videos drew no comment at all against 26.2% of long ones.

Video length by content group

Shading runs within each column
Content groupVideos3 minutes or shorterAverage length of the longer videos
People & Vlogs121,92870.7%32.1 min
Entertainment & Comedy189,23358.5%31 min
Education & How-To142,32158.3%24.1 min
News & Commentary132,16455.3%31.9 min
Tech & Science116,59554.8%22.7 min
Lifestyle & Interests145,22353.8%25.1 min
Music138,21452.3%20.6 min
Gaming130,07149.3%62.6 min

Darker means higher within that column.

Broad channel sample, public videos with a known length. Gaming posts the fewest short videos and by far the longest long ones, at 62.6 minutes against 20.6 for music.

Two ways of counting, one direction

Pooling every view and every comment in the broad sample gives 0.61 comments per 1,000 views on longer videos against 0.29 on videos three minutes or shorter. Counting each channel first and taking the median across channels gives 2.3 against 0.6.

The two methods disagree on the size of the gap and agree on its direction. Pooling lets the biggest channels dominate; the median across channels better describes a typical creator. This report leads with the median for that reason.

These rows include 747,419 sampled comments from the broad channel sample with a known video length. After weighting, they represent 2,570,783 comments on videos three minutes or shorter and 2,439,291 comments on longer videos. The reply row uses comments at least 7 days old from channels with enough comment history.

Buying intent in business YouTube comments

Comments about buying or working with a creator are rare. We need a large combined sample to measure them. A few such comments are not enough to show what is normal for one channel.

Source: broad channel sample and creator-business sample shown separately. Results use recent comments and adjust for each comment's chance of being chosen.

This measure includes comments tagged with buying interest or a partnership request. A comment with both tags counts once. We combine results within each size group because these comments are too rare for a steady middle-channel rate.

The creator-business sample covers business topics. The broad sample covers many kinds of channels. The table compares them side by side. It does not combine them into one YouTube-wide average.

Channel sizeBroad sampleCreator-business sample
1K–10K1.01% (175,352 selected, 2,829 channels)3.11% (62,709 selected, 1,011 channels)
10K–100K1.01% (207,964 selected, 3,269 channels)2.44% (72,234 selected, 1,140 channels)
100K–1M0.99% (220,653 selected, 3,456 channels)1.55% (66,939 selected, 1,048 channels)
1M+0.76% (143,841 selected, 2,251 channels)1.04% (25,530 selected, 399 channels)

18.1%

of older comments on a business channel got a creator reply, against 10.4% in the broad sample. Business creators answer at close to twice the rate.

Based on 176,434 sampled comments at least 7 days old across the business sample. Their unanswered question rate is 73.3% against 81.2%, measured on 38,607 questions across 3,255 business channels.

The same measure in both samples

Shading runs within each column
What the comment is doingBroad sampleCreator-business sampleDifference
Thanks or reaction36%32.9%-3.1pp
Discussion or debate20.7%16.5%-4.2pp
Personal story or result5.7%13.7%+8pp
Asking for information8.6%13.2%+4.6pp
Feedback or correction6.3%7.7%+1.4pp
Other10%5.5%-4.5pp
Request for the creator3.8%3.5%-0.3pp
Insult or attack5.6%3.2%-2.4pp
Spam or self-promotion2.1%2%-0.1pp
Buying or business request0.7%1.5%+0.8pp
Talking with other viewers0.5%0.4%-0.1pp

Darker means higher within that column. The two samples are shown side by side and never combined.

Broad sample: 747,810 sampled comments. Creator-business sample: 227,412 sampled comments. Each comment has one main purpose, so each column adds to 100%.

Business audiences report results

Personal stories and results are 13.7% of comments on business channels against 5.7% in the broad sample, a difference of 8 percentage points and the largest gap between the two samples. Viewers of business channels tell the creator what happened when they tried the advice.

Testimonials specifically run between 3.49% and 4.45% of comments in every size group, against 1% in the broad sample.

And they argue less

Insults and attacks are 3.2% of business comments against 5.6% in the broad sample. Debate is lower too, at 16.5% against 20.7%. These comment sections are quieter, more instrumental, and more directly aimed at the creator.

The two samples were built differently, so this compares two groups of channels we studied. It does not prove that being a business channel causes any of it.

What business viewers ask

Share of real questions in the creator-business sample
  1. How-to or troubleshooting14.9% in the broad sample29.1%
    n=15,105
  2. Other21.2%
    n=9,596
  3. Clarification11.6%
    n=4,448
  4. Gear or source10.3%
    n=5,371
  5. Opinion or discussion8.7%
    n=3,132
  6. Personal or creator7.6%
    n=3,629
  7. Product, price, or purchase6.1% in the broad sample7.1%
    n=3,873
  8. Topic request4.4%
    n=1,753
Creator-business sample: 46,907 sampled questions. How-to and troubleshooting is 29.1% here against 14.9% in the broad sample, which is the clearest difference between the two.

Nearly a third of questions are how-to

Business viewers ask the creator to solve a problem twice as often as viewers in the broad sample. Product and price questions are also higher, at 7.1% against 6.1%.

The topic mix was not a target

The business sample was split by channel size only, never by topic. What it holds is therefore a finding rather than a quota.

What the creator-business sample turned out to be about

Share of the 3,598 business channels
  1. Education & How-To44.4%
  2. People & Vlogs25.1%
  3. Lifestyle & Interests12.6%
  4. Entertainment & Comedy9.8%
  5. Tech & Science4.6%
  6. News & Commentary1.3%
  7. Gaming1.1%
  8. Music1.1%
Creator-business sample. Topic was never a sampling target here, so this describes where business and professional channels actually sit on YouTube's own category axis. Gaming, Music and News together account for 3.5% of them.
Channel sizeChannelsQuestionsHostileResults or testimonialsSpeaks to creator
1K–10K1,01123.3%3.4%4.23%61.6%
10K–100K1,14019.4%3.9%4.45%54%
100K–1M1,04815.4%3.8%3.84%46%
1M+39912.7%4.8%3.49%39.9%

Both the question rate and the share of comments speaking to the creator fall steadily as business channels grow: 23.3% to 12.7% and 61.6% to 39.9%. The smallest business channels have the most direct comment sections in the entire study.

Buying interest means clear interest in price, access, availability, or making a purchase. Partnership requests include sponsorships, interviews, and other business offers to the creator. This does not measure sales. We adjusted for each comment's chance of being chosen. We did not calculate a confidence range that accounts for comments coming from the same channel.

What language YouTube comments are written in

The study only kept channels that mostly publish in English. Their comment sections are a different story.

Source: broad channel sample and recent comments, adjusted for each comment's chance of being chosen.

The language of a comment on a mostly-English channel

Estimated share of comments
  1. English69.6%
  2. Another language17%
  3. No words at all9.1%
  4. More than one language4.3%
Broad channel sample: 747,810 sampled comments. Every channel in this sample was screened as mostly English before capture, so this measures the audience, not the creator.

Screening a channel does not screen its audience

17% of comments are written in another language and a further 4.3% mix languages. A creator publishing in English is not running an English-only comment section.

A further 9.1% contain no words at all. They are emoji or symbols, which is why this report never treats comment volume as a proxy for how much was said.

Channel sizeEnglishAnother languageNo words at all
1K–10K59.6%22%14%
10K–100K67.2%18.6%9.3%
100K–1M72.5%15.9%7.7%
1M+70.1%16.2%9.3%

The smallest channels have the least English comment sections, at 59.6%. Rows do not add to 100% because comments mixing more than one language are counted separately.

Creator reply rate by comment language

Comments at least 7 days old
  1. More than one language13.1%
    n=20,394
  2. Another language12.2%
    n=82,541
  3. English10.4%
    n=301,790
  4. No words at all6.9%
    n=59,124
Broad channel sample: 463,849 sampled comments at least 7 days old. Comments in another language were answered slightly more often than English ones, not less. Comments with no words were answered least.

This is a caution about reading the report's headline rates as a story about English-speaking audiences. It is also a limit: the model assigned the language, and it was not checked against a separate set labelled by people.

The words and emoji YouTube repeats

Most YouTube comments are short, and a tiny shared vocabulary travels through thousands of unrelated comment sections. Here are the words and emoji at the top, from one heart to a surprisingly persistent council.

Source: both channel samples and 5,923,310 recent comment entries. Repeated text is grouped after ignoring capitalization and repeated whitespace.

15.4
Words in the average comment
25.3% are three words or fewer
73.3%
Comments with no likes
Only 2.6% reached ten likes or more
5.1%
Comments that are only emoji
No words at all
4,171,371
Distinct commenters
Across 5,011,893 recent broad-sample comments

How long a YouTube comment is

Share of recent broad-sample comments
  1. 1 to 3 words25.3%
    1,266,340
  2. 4 to 10 words35.6%
    1,782,026
  3. 11 to 25 words24.7%
    1,239,392
  4. 26 to 50 words9.3%
    464,333
  5. More than 50 words5.2%
    259,802
Broad channel sample: 5,011,893 recent comments with a word count. A quarter of everything posted is three words or fewer, and only one comment in twenty runs past fifty words.

Most comments are very short

Six comments in ten are ten words or fewer. That is part of why sorting a comment section by newest first works so badly: the short reactions arrive in the same stream as the few long, specific comments worth answering, and they look identical in a list.

Likes do not sort the list either

73.3% of comments never receive a single like. A creator sorting by top comments is choosing between a small number of comments that happened to arrive early, not between the comments that most need a reply.

422

pieces of normalized comment text each appear on more than 100 different channels. Together they account for 295,043 comments, or one comment in twenty.

Repeated text does not tell us who or what wrote it. It does tell us that a very small number of exact strings are doing a very large amount of the typing.

4

pieces of text account for 19% of every comment in the widely-repeated layer. Twenty-one of them account for 45%.

The pattern is broad and shallow: a widely-repeated text reaches a median of 180 channels and posts about 1.3 times on each. Before reading that as coordination, see what these texts are.

Comment superlatives

YouTube's comment starter pack

Four ways to read the leaderboard: raw volume, plain words, reach, and the difference between a universal reaction and a community refrain.

Most posted

22,339 posts

The same one-heart comment appeared across 6,679 channels.

Most posted word

Nice

4,543 posts

It also travelled furthest of any word, reaching 2,664 channels.

The reach climber

Wow

2,205 channels

18th by posts, but 10th when the list is ranked by channel reach.

Most concentrated

Amen

7.1 per channel

3,372 posts on 478 channels: a local refrain more than a universal one.

TextTimes postedChannels it appeared onPosts per channel
22,3396,6793.3
❤❤❤15,1644,7973.2
❤❤9,8463,5852.7
❤❤❤❤9,3463,2952.8
😂8,5713,3922.5
❤❤❤❤❤6,9872,7292.6
❤️6,7592,7172.5
😂😂😂6,0292,4192.5
Nice4,5432,6641.7
Hi4,5312,1532.1
😂😂😂😂4,3511,8972.3
Yes4,3181,4503.0
❤❤❤❤❤❤4,2562,0022.1
😂😂3,6561,7322.1
thank you3,4001,9841.7
Amen3,3724787.1
😢3,2361,8371.8
Wow3,2272,2051.5
😂😂😂😂😂3,1171,5242.0
❤❤❤❤❤❤❤3,1131,5952.0
😊2,7981,7761.6
🎉2,7671,8611.5
😮2,5001,7681.4
first2,3921,1782.0
Good2,3161,4621.6
JOIN THE COUNCIL2,3165194.5

The 25th position is a tie: “Good” and “JOIN THE COUNCIL” were each posted 2,316 times, so the table shows 26 entries. Capitalization and repeated whitespace are ignored when texts are grouped. The table shows one representative spelling for each group. These are shown verbatim because each was posted on hundreds of unrelated channels by many different people, which makes them a shared vocabulary rather than anyone's own words. No commenter or channel is named, and no comment outside this layer is quoted anywhere in the report.

What the displayed comment leaderboard is made of

Share · comments
  1. Hearts53.6% · 77,810
  2. Laughter17.7% · 25,724
  3. Other emoji7.8% · 11,301
  4. Words or phrases20.9% · 30,415
145,250 posts across the 26 displayed entries. This describes the leaderboard, not all 5.92 million recent comments.

Emoji own the leaderboard

17 of the 26 entries are emoji-only, accounting for 79.1% of their posts. Heart variants appear 3.0× as often as laughing-face variants. That is a result about these leading entries, not the full comment corpus.

The table carries 7 different heart run lengths plus a red-heart variant, and 5 laughing-face run lengths. People do not just choose an emoji; they choose how many.

❤❤❤ > ❤❤

54% more hearts: 15,164 posts versus 9,846.

😂😂😂 > 😂😂

65% more laughter: 6,029 posts versus 3,656.

How long a widely-repeated text is

Share of the 295,043 comments in that layer
  1. One word90.9%
    268,093
  2. Two or three words8.8%
    26,043
  3. Four to ten words0.3%
    907
  4. More than ten words0%
    0
Broad and business samples, recent comments. Length is described here rather than the text itself, which keeps this an aggregate and lets it stand alongside the rest of the report.

This layer is reactions, not a campaign

The repeated texts average 1.1 words. 90.9% of them are a single word and 56.7% are nothing but emoji. None of the 295,043 comments in this layer contains a link, and 0.3% are in capitals.

That matters for how the duplication number is read. Text repeating across hundreds of channels sounds like coordination, and some of it may be. But a layer made of one-word reactions is mostly what it looks like: many different people reaching for the same short response, which is not a campaign and not a bot.

It is also why this report never converts duplication into a bot estimate. The structure is measurable. The intent behind it is not.

14,659 of the 15,403 channels in the study received at least one of these texts, a median of 7 distinct ones each. That reach is close to universal, which is the part worth knowing. It still does not tell us who or what wrote them.

Duplicate comments, junk, and empty comment sections

Repeated text is one part of a wider low-signal layer: promotion, links, all-caps posts, and videos where no public conversation starts at all. This chapter separates those structures instead of turning them into one bot estimate.

Source: broad channel sample and recent comments for repeated text and spam identified by the model. A separate recent-video group measures videos with no comments or comments turned off.

11%

Same text on at least three unrelated channels, after ignoring capitalization and repeated whitespace

The middle broad-sample channel. Repeated text does not tell us who or what wrote it.

How far one piece of comment text travels

Shading runs within each column
Text appears onDistinct textsCommentsShare of all comments
1 channel only5,137,7095,211,19488%
2 channels40,13993,2461.6%
3 to 5 channels21,21193,6961.6%
6 to 20 channels9,638118,3782%
21 to 100 channels2,215111,7531.9%
More than 100 channels422295,0435%

Darker means higher within that column. Distinct texts are counts, so they are not shaded.

Both channel samples, recent comments: 5,923,310 comments reduced to 5,211,334 case- and whitespace-normalized texts. Read the last two rows against the first: 2,637 texts produce nearly 7% of every comment in the study.

88% of comments are written once and appear nowhere else, which is the reassuring half of this table. The tail is the other half, and it is extremely concentrated.

Channel sizeSpam or self-promotionHostileEmoji onlyAll capitalsContains a link
1K–10K1.75%7.03%8.98%1.66%0.11%
10K–100K1.91%6.45%5.40%1.82%0.12%
100K–1M1.78%7.32%4.35%1.75%0.13%
1M+2.53%6.56%4.75%2.10%0.10%

Viewer-to-viewer @mentions are almost non-existent: 0.23% of comments contain one, and they are answered 6.3% of the time against 9.1% for everything else. YouTube comment sections are not a place where the audience talks to itself by name.

Two patterns run in opposite directions. Self-promotion rises with channel size, from 1.75% to 2.53%. Emoji-only comments fall sharply, from 8.98% on the smallest channels to 4.35% on the largest. Links are rare everywhere, at roughly 1 comment in 800, because YouTube removes many of them before anyone outside the channel can see them.

Repeated text does not mean AI wrote it

This measure finds the same cleaned text on at least three unrelated channels. Repeated text may come from memes, copied praise, ads, templates, planned group activity, bots, or other sources. The data cannot show who or what wrote it.

Spam identified by the model and self-promotion are measured on their own. Both measures use only comments that stayed public long enough for us to collect them.

Zero comments and disabled comments are different

A zero-comment video had public comment numbers but no comments. A video with comments turned off did not have the public comment count used by the study. We leave those videos out of the zero-comment rate. The comments-off rate uses all videos we observed.

Empty recent comment sections are much more common among smaller channels. This finding covers the selected group of recent videos, not every video those channels have ever published.

When YouTube comments arrive

Most comments we saw arrived early in a video's public life. The pattern changes with channel size, and new comments still arrive after the first two days.

Source: full video histories for videos that met the study rules and were posted 30 to 90 days before the study.

How quickly comments arrive

Percent of comments. Scale: 0% to 100%
First 48 hours (%)First 7 days (%)
1K–10K10K–100K100K–1M1M+
Channel sizeFirst 48 hours (%)First 7 days (%)Study base
1K–10K72.5%92.3%137 channels in this comparison; 23,264 comments
10K–100K73%90.3%531 channels in this comparison; 137,796 comments
100K–1M71.1%88.6%1,073 channels in this comparison; 495,774 comments
1M+59.6%81%887 channels in this comparison; 820,631 comments
This chart uses the broad sample and full video histories. Each size-group result uses channels with at least 50 captured comments. The first 7 days include the first 48 hours.

Full video histories show comment timing

The recent-comments group starts with each channel's newest comments, so it cannot show a video's full history. For the full-history group, we chose up to five videos per channel that were 30 to 90 days old. We kept collecting until YouTube had no more comments to return.

All 27,424 completed video histories reached that end point. The group covers 6,371 channels across both samples. Those totals are a check on the full study. The chart uses 2,628 broad-sample channels. The videos had comments turned on, at least one public comment, and enough age to show their early life.

The chart covers active public comment sections

Videos with no comments and videos with comments turned off are not in this chart. The result shows when comments arrive on videos that met the study rules and had public comments. It does not describe every upload.

55.4%

of a video's comments arrive on the day it is published. By day 7 the video is receiving 41 times fewer comments than on day one.

Based on 3,484,795 comments on videos with a full public history in the broad sample. 81.7% arrive in the first week, and 3.5% arrive after day 30.

How a video's comments arrive over its first 45 days

Comments per day since publication
Day 0Day 5Day 10Day 15Day 20Day 25Day 30Day 35Day 40Day 44
Full video histories, broad channel sample: 3,484,795 comments on videos published 30 to 90 days before capture. Day 0 is the publication day. The curve stops at day 45 because the video cohort was picked by age, so later days are not equally observed for every video.

The shape is the finding. A comment section is not a steady stream, it is a spike and a long tail: 55.4% on day zero, 65.4% by the end of day one, 81.7% by the end of the first week, and a thin 18.3% spread across everything after that. A creator who answers comments once a week is answering into the tail, not the spike.

How much of the conversation has already happened

Running share of a video's comments, by day since publication
Day 0Day 22Day 44
55.4%
by end of day 0
65.4%
by end of day 1
71%
by end of day 2
81.7%
by end of day 6
88.7%
by end of day 13
96.5%
by end of day 29
Full video histories, broad channel sample: 3,484,795 comments. Same data as the chart above, read as a running total. The curve is steep for two days and then almost flat, which is the practical shape of a comment section's life.

This is the version to plan against. A creator who checks comments once a day has already missed most of the first-day rush. One who checks once a week arrives after 81.7% of the conversation is over.

What time of day comments arrive

Comments per hour, UTC
00:0003:0006:0009:0012:0015:0018:0021:0023:00
Hour (UTC)Comments
00:00146,211
01:00151,733
02:00147,502
03:00140,974
04:00146,494
05:00129,138
06:00124,308
07:00123,163
08:00142,148
09:00130,376
10:00125,249
11:00133,768
12:00149,046
13:00165,476
14:00192,380
15:00240,590
16:00253,450
17:00233,324
18:00210,076
19:00201,540
20:00185,709
21:00170,513
22:00161,554
23:00148,987
Full video histories: 3,953,709 comments with a recorded time. Hours are UTC because public data does not tell us where a commenter was, so this shows the shape of a global day rather than any one country's evening. The busiest hour carries just over twice the volume of the quietest.

Which days comments arrive

Comments per weekday, UTC
MonTueWedThuFriSatSun
DayComments
Mon571,067
Tue525,316
Wed528,501
Thu520,444
Fri551,001
Sat644,676
Sun612,704
Full video histories: 3,953,709 comments. Saturday is the busiest day and Thursday the quietest.

Comments arrive at the weekend

Saturday carried 24% more comments than Thursday. Sunday was second. The midweek dip is small but steady across the sample.

Speed is not the problem

When a creator does reply, the reply is fast. Half of the replies we could time arrived within 3.7 hours, and 31.2% arrived within the first hour. Creators in this sample are not slow. They are selective.

That is measured on 255,207 replies whose timing we could see in the broad sample. It says nothing about the comments that never got a reply.

Does replying to comments change anything

Public data cannot show what a reply caused, but it can test one narrow outcome with the reply placed before the outcome: whether that viewer writes another comment on the same channel.

Source: broad channel sample, recent comments, comments at least 7 days old from channels with enough comment history. 463,849 sampled comments, 63,048 of which received a creator reply.

MeasureNo creator replyCreator repliedWhat it looks like
Comments with no likes72%42.6%Looks like a large lift
Comments with 5 likes or more3.4%3.4%Identical
Comments with 25 likes or more0.6%0.4%Slightly lower
Average likes1.311.35Almost unchanged
Average replies in the thread0.101.27At least 1.0 of the gap is the creator's own reply

The first problem: the creator is in their own count

A replied-to comment averages 1.27 replies in its thread against 0.10 for a comment with no reply. That gap looks decisive until you notice that 99.4% of replied comments have a non-zero thread count, because YouTube counts the creator's own reply inside it.

Take that one reply back out and the gap shrinks from 1.27 to roughly 0.27. Most of the apparent conversation is the creator talking.

The second problem: the shape of the like lift is wrong

Replied-to comments are much less likely to have zero likes: 42.6% against 72%. But they are no more likely to reach 5 likes, and slightly less likely to reach 25. A real popularity effect would show up at the top of the range. This one only shows up at the very bottom.

That pattern is consistent with a creator liking the comment they answered, which would add exactly one like and move nothing else. The data cannot confirm the cause, only that the lift appears where a single like would put it.

0.94 vs 0.78

Average likes on questions with no creator reply against questions that got one. The replied-to questions did slightly worse.

Based on 61,034 sampled questions at least 7 days old. Whatever a reply does, it does not reliably make the original comment more popular.

One thing we can measure: does the commenter come back

Whether a viewer subscribed or bought something is invisible in public data. One outcome is not: whether the same person commented again on the same channel. The naive version of that comparison is badly confounded, because a channel that replies is different from one that does not in every other way too, so comparing across channels measures the channel rather than the reply.

We set a 24-hour landmark. A viewer counts as exposed only when the creator replied during those first 24 hours; the return window starts after that point and ends on day 14. Anyone who returned before the landmark is excluded before the groups are formed. We then compare within the same channel, comment-length band, and question flag—without using likes, which can change after a reply.

What survives when the channel is held fixed

Share who wrote another comment after the 24-hour landmark
  1. Replied within 24h, across channels8.55%The naive comparison, mixing responsive and silent channels together
  2. No reply within 24h, across channels4.27%A gap of +4.28 points; more than half disappears after adjustment
  3. Replied within 24h, adjusted6.64%Standardized within channel, length band and question flag
  4. No reply within 24h, adjusted4.68%A gap of +1.97 points, 95% range +1.73 to +2.21
Broad channel sample: 65,263 commenter-channel relationships in 3,317 overlap strata across 1,824 channels. Index comments had at least 14 days of observation; 29,054 people who returned before the 24-hour landmark were excluded before exposure assignment.

54%

of the naive return-rate gap disappears after the channel and two pre-reply comment features are held fixed. The gap falls from +4.28 points to +1.97 points.

The 95% channel-bootstrap range is +1.73 to +2.21 points. That is evidence of an association in this observed cohort, not proof that the reply caused the return.

Why this still is not proof

Creators choose which comments to answer, and that choice is not random. If they tend to answer people who already seemed engaged, some of the remaining +1.97 points may be that judgement rather than the reply.

The landmark fixes temporal order, and stratification removes the channel, comment-length band, and question flag from the comparison. It cannot remove unmeasured differences, especially whatever made the creator pick that comment.

What changed, and what did not

This report previously refused to put any number on replying. It can now put a number on one narrow outcome, which is progress, but more than half of the naive gap disappears under the adjusted comparison.

Nothing here supports the wider claim. Replying is not shown to cause views, subscribers, revenue, or loyalty. It is associated with a slightly higher chance that one person writes one more comment.

How channel age and country change a comment section

Two things the study recorded about every channel and has not used yet: how long the channel has existed, and which country it says it is in.

Source: broad channel sample and recent comments. Country is known for 9,429 of 11,805 channels (79.9%) and is the channel’s own declared country, not the commenter’s location.

Comment behaviour by how long the channel has existed

Shading runs within each column
Channel ageChannelsQuestionsHostileComments in EnglishCreator reply rate
Under 2 years1,7608.1%7%58%12.2%
2 to 5 years2,21110%6.3%62.5%10.9%
5 to 10 years3,39211.2%6.4%67.3%10.7%
Over 10 years4,44210.3%7.3%77.2%9.3%

Darker means higher within that column. Channels are counts, so they are not shaded.

Broad channel sample: 747,810 sampled comments, channel age measured to 20 August 2026. Reply rate uses comments at least 7 days old from channels with enough comment history.

The reply rate falls steadily with channel age, from 12.2% under two years to 9.3% past ten. So does the share of non-English comments: a channel over ten years old runs at 77.2% English against 58% for one under two. Newer channels are both more responsive and more international.

Comment behaviour by the channel's declared country

Shading runs within each column
CountryChannelsQuestionsHostileComments in EnglishWords per commentCreator reply rate
United States3,9949.9%7.7%88.6%17.89.6%
India2,06712.4%6.1%38%8.412.5%
United Kingdom64410.2%7.7%89.4%20.210.4%
Canada34610.8%5.6%89.6%19.211.6%
Indonesia2439.7%4.6%14.2%8.414.5%
Pakistan2219.5%9.4%32.1%10.810.5%
Australia19212.1%5.2%90.7%1814.4%
Brazil1467.5%6.4%11.4%11.716%

Darker means higher within that column. The eight countries with the most channels in the sample.

Broad channel sample. Every channel here was screened as mostly English before capture, so a country row describes English-publishing channels based there, not that country's YouTube as a whole.

9.6% vs 16%

Creator reply rates on channels declaring the United States against channels declaring Brazil. The US has the lowest reply rate of the eight largest countries in the sample.

US and UK channels also draw the most hostile comments, both at 7.7%, against 4.6% for Indonesia and 5.2% for Australia.

Comment length is a country-level habit

A comment on a UK channel averages 20.2 words. On an Indian or Indonesian channel it averages 8.4. That is a difference of more than two to one, larger than any gap this report finds between content groups or channel sizes.

Reading it as national character would be a mistake. Language, keyboard, phone versus desktop, and what people use YouTube for all sit inside this number, and public data separates none of them.

What this column is not

Country is what the channel declares in its own settings. It is not where the commenters are, and this study cannot see that at all. A US channel with a global audience appears in the US row.

It is also missing for one channel in five, and the screening step means every row here is a mostly-English channel. Indonesian channels in this sample run at 14.2% English comments, which is a good reminder that the screen applied to the creator, not the audience.

YouTube comment benchmarks

Compare a channel with others near its size, not with one average for the full sample. This table shows the middle broad-sample channel in each size group.

Source: broad channel sample. Reply data uses recent comments from channels with enough older comments. Video and timing data use the separate groups named below.

Channel sizeChannelsCreator reply rateReply delayRepeated textZero-comment videosComments in first 48 hours
1K–10K2,8296.5%4.5 hours14%44.7%72.5%
10K–100K3,2691.6%5.2 hours10%10%73%
100K–1M3,4560%5.2 hours9.1%0%71.1%
1M+2,2510%4.0 hours11.4%0%59.6%

The Channels column shows every broad-sample channel in that size group. Each other column uses only channels that meet the rules for that measure. Reply delay needs a reply we could see. Video measures use the public-video group. Timing needs at least 50 comments from full video histories.

Compare with channels most like yours

First choose the size group closest to your channel, using the table above. Then read the same measure for your content group in the content category chapter, which cuts unanswered questions across all 32 size-by-category groups.

Percentile spreads are published on this page for the reply rate only, in the chart below. We hide any cell fewer than 30 channels can support. Two measures in the same row may therefore rest on different numbers of channels.

How far apart channels of the same size are

Creator reply rate per channel, comments at least 7 days old
  1. 1K–10K6.5% median · p25 0 · p75 34.1 · p90 702,727 channels. The most responsive 1% replied to 97.9%.
  2. 10K–100K1.6% median · p25 0 · p75 18.5 · p90 56.52,998 channels. The most responsive 1% replied to 97.2%.
  3. 100K–1M0% median · p25 0 · p75 5.3 · p90 23.12,631 channels. The most responsive 1% replied to 90.9%.
  4. 1M+0% median · p25 0 · p75 0.8 · p90 8.2954 channels. The most responsive 1% replied to 56.3%.

Bar spans the 25th to 75th percentile channel, the dot is the median, the tick is the 90th

Broad channel sample, channels with at least 20 comments that were 7 days old and captured comments reaching back 14 days. The bar covers the middle half of channels in each size group.

This is the reason to compare against a percentile rather than an average. In the 1,000 to 10,000 group, the median channel-level reply rate was 6.5%, but the most responsive tenth replied to 70% or more. Both are normal. A single platform-wide reply rate describes almost no real channel.

Share of commenters who only ever comment once

Recent comments, by channel size
  1. 1K–10K22.7% comment more than once77.3%
    2,668 channels
  2. 10K–100K19.7% comment more than once80.3%
    3,229 channels
  3. 100K–1M16.6% comment more than once83.4%
    3,440 channels
  4. 1M+14.1% comment more than once85.9%
    2,246 channels
Broad channel sample, channels with at least 50 captured comments. The rest are people who commented more than once on the same channel inside the captured window.

No small group runs the comment section

Across 11,138 channels, the busiest 1% of a channel's commenters wrote 5.3% of its comments, the busiest 5% wrote 14.6%, and the busiest 10% wrote 22.2%.

That is far flatter than the concentration usually described in online communities. It is measured inside the captured window, so treat it as a floor: a regular who happened to comment once during that window looks like a one-off here.

Smaller channels have more regulars

On channels with 1,000 to 10,000 subscribers, 22.7% of comments came from people who commented more than once. Above 1 million subscribers that falls to 14.1%.

The captured window is short, so this undercounts regulars on any channel that posts rarely. Treat it as a floor. The direction holds across every size group.

Likes, comments, and channel size

Shading runs within each column
Channel sizeVideosLikes per 1,000 viewsComments per 100 likes
1K–10K188,26911.344.45
10K–100K228,42113.022.99
100K–1M346,39918.193.09
1M+352,43816.262.77

Darker means higher within that column. Videos are counts, so they are not shaded.

Broad channel sample, public videos with a view count. Pooled across all views and all likes in the group, so the largest channels carry the most weight here.

The two columns move in opposite directions. Bigger channels earn more likes per view, up to 18.19 per thousand, but convert far fewer of those likes into comments: 4.45 comments per 100 likes at 1,000 to 10,000 subscribers against 2.77 above a million. Liking scales with reach. Commenting does not.

Who writes a channel's comments

Shading runs within each column
Channel sizeChannelsShare from the busiest 1% of commentersShare from the busiest 10%Share from the single busiest commenter
1K–10K2,3355.9%23.1%4.54%
10K–100K3,1445.5%23.3%2.86%
100K–1M3,4175.2%22.3%1.89%
1M+2,2425.2%21.3%1.51%

Darker means higher within that column. Channels are counts, so they are not shaded.

Broad channel sample, channels with at least 50 distinct commenters in the captured window. Measured inside that window only, so every column is a floor.

The group figures barely move with channel size: the busiest tenth of commenters wrote between 21.3% and 23.3% of comments at every size. One column does move. On a channel with 1,000 to 10,000 subscribers, a single person wrote 4.54% of all its comments, three times the 1.51% seen above 1 million subscribers. Small channels have a most-frequent commenter who is genuinely identifiable. Large ones do not.

Channel sizeUseful replyGeneric templateReplies / channels
1K–10K39.9%28.2%30,678 / 1,831 ch.
10K–100K44.6%24.6%22,015 / 1,597 ch.
100K–1M49.4%21.8%9,071 / 1,008 ch.
1M+47%28.3%1,276 / 207 ch.

We weighted reply styles to correct how comments were chosen. A useful reply adds information or a personal response. A generic reply, such as a bare thank-you or emoji, could fit almost any comment. The last column shows sampled replies and channels. Three unreadable replies were left out of this broad-sample table.

Study methodology and analytical framework

This report was constructed from an immutable corpus of public YouTube data collected via the official YouTube Data API v3. All analytical datasets were stratified, classified, verified, and synthesized using advanced frontier LLMs and rigorous statistical methods without using private CommentShark customer data.

Dataset record: 15,403 verified channels and 9,877,019 public comments captured August 19–22, 2026. All findings are derived directly from the immutable study snapshot.

Channel discovery and sampling frame

Channels were discovered via systematic topic search queries across four stratified subscriber cohorts: 1K–10K, 10K–100K, 100K–1M, and 1M+ subscribers. Strict exclusion criteria filtered out made-for-kids channels, channels with concealed subscriber counts, channels without active upload histories, channels with fewer than 20 public comments, predominantly non-English channels, and permanent API retrieval errors.

Known CommentShark customer accounts and outreach prospects were excluded to preserve research independence. In the database, the broad representative sample is designated Frame A and the creator-business sample is designated Frame B.

Cross-sectional comment collection (Mode A)

For current comment section analysis, comments were ingested in reverse chronological order until reaching at least 14 days of comment history (up to 12 pages or approximately 1,200 top-level comments per channel).

To ensure valid observation windows for creator engagement, reply-rate analyses are strictly restricted to comments matured to at least 7 days of age on channels with at least 14 days of ingested history and a minimum of 20 matured comments. In total, 12,701 channels satisfied this maturation requirement.

Full video lifecycle histories (Mode B)

To investigate temporal comment velocity and arrival distributions, up to five videos per channel aged 30 to 90 days were sampled across diverse view tiers. Comment threads were fetched exhaustively until pagination terminated.

This longitudinal lifecycle dataset provides the empirical basis for Day 0 through Day 7 arrival curves and decay rates across active public uploads.

Stratified comment sampling & weighting

From each channel's cross-sectional history, up to 64 comments were sampled (balanced between 32 comments ≥14 days old and 32 newer comments) via deterministic hash-based ordering, completely blind to engagement metrics, likes, or comment length.

Inverse probability weighting was applied during aggregation to correct for within-channel sampling probabilities and ensure unbiased cohort estimators.

Frontier AI synthesis & semantic classification

The primary semantic classification corpus comprised 879,383 unique comment records (from 975,222 total sampled records). Initial high-throughput semantic extraction was performed using gpt-5.6-luna (research prompt v2, low reasoning effort, 42,291 batches at a total token cost of $44.70), annotating comment intent, communicative directedness, stance, personal narrative markers, language family, and question taxonomy.

To synthesize cross-chapter narratives, extract qualitative patterns, and cross-validate complex findings, the research pipeline utilized a multi-model ensemble of frontier LLMs—including ChatGPT 5.6 Sol, Claude Opus 5, and Gemini 3.7. These models conducted multi-perspective synthesis, thematic validation, and cross-verification of statistical results across the study's analytical chapters.

Discriminative validation & signal attribution

To confirm that semantic classifications capture meaningful signal rather than noise, we evaluated predictive utility against held-out channels. Predicting creator replies using structural metadata alone yielded a ROC AUC of 0.684. Incorporating semantic classifications increased discrimination to 0.757—an absolute gain of 0.072 (channel-bootstrap 95% CI 0.065 to 0.079). Semantic labels account for 28% of the combined model's discriminative power.

This five-fold evaluation used 463,849 matured comments across 9,388 channels, strictly isolating channels within single folds. Leakage variables (such as total thread reply count and creator likes) were strictly isolated and excluded from predictive features.

Quality assurance & reliability

All model outputs underwent automated schema validation. Test-retest reliability across duplicate sample runs demonstrated 88% concordance on creator-value classifications and 86% agreement on multi-label taxonomy assignments.

All qualitative metrics represent model-coded statistical classifications rather than manual human annotations.

OutcomeChannelsWhat it means
Entered the study15,403Met every rule and was captured
Outside the four size groups5,164Under 1,000 subscribers or a hidden count
Held back as surplus4,595Its group had already reached the target
Mainly another language993The study is English-only by design
Fewer than 20 public comments292Too little to measure anything
Made for kids272Comments are disabled by YouTube
No uploads5Nothing to capture
No viewer comments3Only creator posts
Permanent API failure3Could not be read at all

26,730 channels were evaluated in total and 15,403 entered the study. The largest single exclusion is channels outside the four size groups, reflecting the 1,000 subscriber floor. Surplus channels met all inclusion criteria but arrived after their cohort quota had been fulfilled, and were held in reserve.

Central tendency & aggregation

Most headline benchmarks reflect the median of eligible channel-level rates rather than pooled comment averages, preventing high-volume channels from dominating results.

Modeled statistical estimators

Beyond descriptive counts, five specialized analysis families employ formal statistical estimators: Kaplan-Meier reply survival horizons, intra-class correlation (ICC) for within-channel clustering, landmark matched-cohort return models, cluster stability evaluations, and held-out classifier AUC validation.

Uncertainty & bootstrap estimation

Descriptive tables report exact sample counts. Matched comparisons, ICC metrics, and AUC gains report cluster-robust 95% confidence intervals derived from channel-level bootstrapping (resampling all comments from a channel together).

Data suppression thresholds

Benchmark segments with fewer than 30 qualifying channels are suppressed to prevent unreliable comparisons.

Privacy & aggregation standard

All findings are reported in aggregate. The only verbatim text displayed comprises high-frequency phrases repeated across hundreds of unrelated channels. No individual commenter or channel identities are disclosed.

What this YouTube comment study cannot show

A large study can still have clear limits. This report sees only public activity from channels that fit the study rules. It cannot see private moderation, every YouTube channel, or cause and effect.

Who the sample covers

The sample covers public, mainly English channels found through YouTube search. It leaves out channels below 1,000 subscribers. The results describe this study sample, not every YouTube channel.

What moderation hides

We cannot see removed, deleted, private, or held-for-review comments. Public data also does not show pinned or hearted status. Spam and abuse rates are minimum estimates of what creators may receive.

Cause and effect

A link between two measures does not prove that one caused the other. This report does not claim that creator replies cause more views, loyalty, or growth.

No growth measure

This report measures a single point in time. It does not track a channel over time, so it shows no growth figure and no change from the same time last year.

Subscriber counts

YouTube rounds public subscriber counts. A small visible change may come from a display step, not an exact gain or loss.

Model-coded labels

Model labels can change between runs. We did not test the final model against a separate set labeled by people. Direct facts such as reply presence, time, and video length do not depend on model labels.

The YouTube Comment Report changelog

We record every update that changes a number, method, or conclusion. Small spelling and layout fixes do not need an entry.

  1. Report published.

Suggested citation

CommentShark. 2026. The YouTube Comment Report. CommentShark. https://www.commentshark.com/youtube-comment-report

Public summary tables may be reused under CC BY 4.0 with credit. When citing one number, include its chapter, channel sample, comment group, count, and access date.