AI analysis of YouTube comments
The answers to what viewers were missing and what to film next are in the comments. But at hundreds of comments nobody reads them systematically — so they never get used.
Straight up: this is a showcase build I made myself, not a client job. I show it so you can see what a system like this looks like from the inside and what stays with a person in it. My client jobs with a verifiable review are here: Luko Tools and a desktop personal finance application.
The problem: a content creator wants to know what viewers liked, what was missing, which questions keep coming up and what to film next. At hundreds of comments under a video, nobody has time to read them in a way that produces a decision.
The solution: the system pulls all the comments under a video, AI evaluates sentiment, groups recurring topics and questions, and returns an overview.
The main benefit is not sentiment but the grouping. Finding out that forty people are asking the same thing is concrete material for the next video or for expanding the description — reading one comment after another, that gets lost.
How it works
- Collecting comments. The system pulls all comments under the given video through the official API.
- Analysis. AI evaluates sentiment (positive, negative, neutral) and groups recurring topics and questions.
- The report. An overview is produced: what interested viewers, what they are asking about and which topics they suggest themselves.
A run takes minutes instead of hours of reading, so it can be triggered after every video — not just occasionally, when it gets its turn.
What the analysis gives you
Audience sentiment
The overall tone of reactions under the video, including whether negative comments are about the content or about something else.
Most common questions
Grouped by topic — ready-made material for an FAQ, for expanding the video description, or for the next episode.
Ideas from viewers
Concrete suggestions for further videos that otherwise get lost among the rest of the comments.
Criticism and weak spots
What viewers missed or what bothered them — the most useful and most frequently overlooked part.
Tools used
- YouTube Data API — pulling the comments
- A language model — sentiment evaluation and topic grouping
- An automation platform — orchestration and triggering
- A spreadsheet or a message — delivering the result where it will be read
What this analysis cannot do
Irony and sarcasm. Sentiment evaluation does not handle them reliably — and comments are full of them. Sentiment is therefore an indicative signal, not a measurement.
Small comment counts. At ten comments it is faster and more accurate to read them. The analysis makes sense from hundreds upward.
Representativeness. The people who comment are the ones something struck or annoyed. It is not a sample of the audience and should not be read as one.
The content decision. The analysis supplies material. What to film stays the creator's decision — following only what the comments ask for is the road to an average channel.
The outcome, and where else it works
Hours of reading comments become a few minutes, and content decisions rest on what viewers actually wrote rather than on the impression left by the first ten comments.
The same principle works anywhere there is a lot of unstructured feedback: product ratings in a shop, survey responses, support tickets or reviews on comparison sites. The input differs, the procedure is the same — pull, group, return an overview.
I do not quote a specific time saving as a measured figure — it depends on the number of comments and on how thoroughly they were read before. You can work out an estimate in the payback calculator.
Frequently asked questions
Can AI recognise irony in comments?
Not reliably. That is why sentiment evaluation is an indicative signal rather than a measurement — and comments under videos tend to be full of irony. The more useful part of the analysis is the grouping of topics and questions, which is markedly more reliable.
From how many comments is it worth it?
From the hundreds upward. At a few dozen comments it is faster and more accurate to read them. The value of the analysis grows with volume, because its main contribution is grouping what repeats.
Can it be used on other feedback sources?
Yes. The same procedure works for product ratings, survey responses, support tickets or reviews on comparison sites. What changes is how the data is obtained, not the analysis itself.
Are comments a representative sample of the audience?
No. The people who comment are the ones something struck or annoyed. The analysis shows what the vocal part of the audience thinks — useful, but not to be confused with the view of all viewers.
What does an analysis like this cost?
A deployment of this type starts at €390 for one source and a simple output. Running costs are mainly the language model by volume of text processed — at hundreds of comments those are negligible amounts.
Related services
Related articles
Related work
Sitting on feedback nobody has time to read?
Comments, ratings, surveys or tickets — tell me what there is and how much of it. I will tell you what can be extracted and whether it is worth it.
Get a free consultation →