Most comment sections get read for mood. You scroll the first twenty, reply to a few, register whether the tone is warm or hostile, and close the tab. That is the cheapest audience research available to anyone making videos, and it gets spent entirely on feelings.
The information is in the questions, and the questions are not at the top. YouTube orders comments by relevance, which weights likes and replies, and likes reward whatever made someone laugh in the two seconds it took to read. The most-liked comment under almost any video is a joke, a compliment, or a timestamp. None of those tell you what to make next. A question does: it is a viewer stating exactly what the video failed to explain, in the words they would have typed into search.
This covers how to find the questions — including the large share typed with no question mark — why one word repeated in a single rant is not a theme, what to discard first, and how a recurring question turns into a video.
The top comment is the worst comment to read
Sort order is the whole problem. A question is a poor candidate for likes by construction: it is specific to one person's setup, and half the people reading it already know the answer. So the comments worth the most to you sit near the bottom of the default order, and on a video with 800 comments nobody scrolls that far.
As a ranking signal comments are weak, and manufactured ones are recognisable. Their value is raw material for the next video, not a number to grow.
Find the questions that have no question mark
Searching for ? finds some of them. People type on phones, and punctuation is the first thing to go: "how do you get the audio to sync" is still a question.
What actually marks a question is the opening word: how, what, why, when, where, which, who, can, could, do, does, did, is, are, will, would, should, anyone, anybody. Look for them at the start of a comment or of a sentence inside one.
Those words also open narration:
- "How I found your channel was..." A story about them.
- "When I tried this on Windows it crashed." A bug report. Useful, but not a question.
A wh-word followed immediately by I, we, he, she, they, or you is someone telling you something, not asking. That single exclusion removes most of the false positives.
The comment extractor is built around this split. It reads a video's comments in the default relevance order, separates the questions into their own tab, sorts them by likes, and opens on that tab. The narration exclusion is applied to comments typed without punctuation; a narration that ends in a question mark still counts as one.
Run competitor videos through it as well as your own. It accepts any public watch or youtu.be URL, and under someone else's video the unanswered questions are the ones nobody has made a video about yet. That is the same demand signal the keyword research workflow sends you to comment sections for, gathered in one pass instead of by scrolling.
One word repeated in one rant is not a theme
Somebody writes four hundred words about lighting and uses the word eleven times, and a naive count reports that lighting is what your audience cares about. It is what one person cares about, at length.
Count each word once per comment instead, so a word scores by how many separate people used it. The extractor counts this way, discards any word appearing in only a single comment, and shows the top twenty with the number of comments each one turned up in.
Four rules for reading that list:
- Your own subject will top it. If the video is about lighting, "lighting" means nothing. The useful entries are the ones you did not put there.
- Look for nouns you never said. A competing product, an operating system, an error string, a piece of gear. Those arrived from the audience.
- Very short words never appear. Anything two characters or shorter is dropped, along with filler like "just" and "really", so "4k" will not show up. Search the exported CSV for those by hand.
- A frequency list is a pointer, not a finding. "Price" in nine comments could be nine people asking what something costs or nine saying it is too much. Read the comments before deciding.
Throw four things out before you read anything
A comment section is mostly not comments. Set these aside first:
- Anything with a link in it. Referral spam;
httpandwwware the giveaway. - Channel trading. "sub4sub", "subscribe to my channel", "check out my channel".
- Giveaway bait. Free in-game currency, free gift cards, and WhatsApp messages carrying a phone number.
- One-word filler. "first", "early", "nice", "lol", "thanks", and comments made entirely of emoji.
The extractor sets these aside rather than deleting them and reports the total as "Filtered out" — that number tells you what share of a comment section is noise. The CSV export still contains every comment, with a column marking the questions.
A comment section is a self-selected sample. The people who type are the enthusiasts and the stuck; the large majority who watched and left never appear. Use comments to find topics, never to estimate how many people want one. Demand is confirmed with search, not with replies.
Turn a repeated question into one video
Most questions name the gap precisely, and the gap is usually a video:
| What they wrote | What it is telling you | The video it becomes |
|---|---|---|
| "What mic is that?" | Your setup is more interesting than you assumed | A gear breakdown with prices |
| "This did not work on Windows" | The walkthrough silently assumed one platform | The Windows version, as its own upload |
| "What settings did you use?" | You showed the result and skipped the recipe | A settings walkthrough with numbers on screen |
| "I get error 0x80070005" | A specific failure that people are already searching | A fix video titled with the error string |
| "Which one should I buy?" | Your review compared and never decided | A verdict video that names one |
Five steps once a question has repeated:
- Write it down in their words. Do not tidy "audio keeps going out of sync" into "resolving timeline audio drift". The untidied version is what people type into search, and what the title should be built around.
- Check that it exists outside your comment section. Three people asking is a hint, not evidence of demand. Put the phrase through the keyword tool and YouTube autocomplete first.
- Give it one video, and answer it in the first sixty seconds. An FAQ compilation answering nine questions ranks for none of them, because no title can name nine things.
- Close the loop. Reply to the original comment with the link. That reply sits under a video that already has traffic.
- Check where it landed. Run the phrase through the rank tracker at one week and one month; Studio will not tell you your position for a specific query.
Run a twenty-minute pass once a month
Batch this rather than doing it upload by upload:
- Pick your three most recent videos, plus the two ranking above you for your main target phrase.
- Run each through the comment extractor. The default depth is 200 comments; signed in you can raise it to 500. Without an account you get the first 100 comments and two videos per day.
- Read the questions tab only, and export each CSV into one sheet: question, video, likes.
- Sort by how many separate videos a question turned up under. Anything asked under three of them is your next upload.
- Keep the sheet. After six months it is a written record of what your audience does not understand.
The short version
- The most-liked comment is a joke or a timestamp; the ones worth reading sit near the bottom of the default order.
- A large share of questions carry no question mark. Find them by the opening word, and exclude the wh-word-plus-pronoun narration that looks identical.
- Count each word once per comment, so a theme means several people rather than one person repeating a word.
- Discard links, channel trading, giveaway bait, and one-word filler before drawing any conclusion.
- One repeated question becomes one video, titled in the commenter's words and answered in the first sixty seconds.
- Comments find topics; search confirms demand. Do not skip the second step.
