TikTok comment analysis with AI is not about sentiment scores. It is about finding the thing your audience keeps saying that you have stopped hearing, because it arrives three words at a time under a video you posted two weeks ago. This guide covers how to get the comments out as a file — free, no signup, CSV, Excel or JSON — and what to actually ask of a model once you have them.
What a "hidden trend" is, in a comment section
Four things hide in comments, and only one of them is visible while scrolling.
- A repeating question. Forty people asking where the jacket is from is a product page you have not written.
- A repeating objection. "Too expensive", "does it work on curly hair", "shipping to Malaysia?" — the same doubt, phrased fifteen ways, blocking the same purchase.
- Migrating vocabulary. A word your audience starts using that you did not teach them. When the comments rename your product, that name belongs in your next caption.
- A demand you are not serving. Requests for a version, a size, a tutorial, a part two. This is the most commercially valuable pattern and the easiest to miss, because each request looks like one person being needy.
The TikTok app actively hides all four. Its comment ordering is engagement-weighted and personalised, replies are collapsed, and a section with 8,000 comments never shows you more than a few hundred without an hour of scrolling. You are not reading a dataset; you are reading whatever the ranking decided to show you today.
Where this guide comes from
We run the exporter, so we see the shape of comment data every day across TikTok, Instagram, YouTube and Facebook. Three things we observe, which matter before you point a model at anything:
First, the number under the video is not the number of rows you will get. TikTok's comment badge counts replies and comments that have since been deleted or hidden by a filter; the public list does not serve those. A gap of 10–20% between badge and rows is normal and is not the exporter missing data. If you analyse row counts as if they were the badge number, your percentages are wrong from the start.
Second, TikTok comment sections are the most multilingual we handle. A single viral video routinely carries Indonesian, Thai, Tagalog, Portuguese and English in the same thread, which is why every export ships a per-comment language column — themes cluster far better when you split by language first.
Third, replies are where the objections live. The top-level comment is a joke; the argument about price is eight replies down. That is why reply expansion is a checkbox rather than a default: it multiplies the row count, and for trend work it is usually worth turning on.
Everything quoted here about caps, columns and what costs money is the live product behaviour, not marketing copy. TikTok's own rules on what may be posted and how comments are moderated are in its Community Guidelines and Terms of Service.
Step one: get the comments out as a file
No analysis of any kind works against the app UI. Start with a file.
Paste the video URL into the TikTok comment exporter and run it. The first 100 comments of any post are free and need no account. Pick JSON if a model is going to read it — the comment text survives unmangled, including emoji and line breaks. Pick CSV or Excel if you want to sort and skim first, which is the right instinct for a single video.
Every row carries author, username, text, likes, replies, created_at, language, is_pinned, the comment id and the reply-to id. Those last two are what let you reconstruct a thread later; keep them even if they look like noise.
One habit worth adopting immediately: put the video slug in the filename. A folder of files called export(3).csv is worthless a fortnight later, and trend work is cross-post work by definition.
Step two: ask one question per pass
The failure mode of AI comment analysis is asking for "insights". You get a paragraph that could describe any video on the platform.
Ask instead for one thing, with a required output shape:
- "List every distinct question asked in these comments, with how many times each appears. Ignore compliments. Quote one real comment per question."
- "Group these comments into at most eight themes. For each, give the count, the summed likes, and three verbatim rows."
- "Which words or phrases appear here that do not appear in the video caption?"
Each of those returns something you can act on or disprove. "Analyse the sentiment of this comment section" returns a number nobody can check. The demand for verbatim quotes is the load-bearing part: it is how you catch the model inventing a theme, which it will do cheerfully on a thin file.
Built-in AI analysis in the app runs this shape of pass for you across a whole export and is available on the Pro passes. Exporting JSON and running it through your own assistant works equally well — the point is the questions, not whose model answers them.
Step three: weight by likes, not by count
A theme mentioned 30 times with four likes total is 30 people typing. A theme mentioned 12 times with 9,000 likes is the comment section agreeing loudly in public, and it is the more important of the two. Sort every theme by summed likes as well as by frequency and look at both columns before deciding anything.
Pinned comments deserve separate handling. The is_pinned flag marks comments the creator elevated, so they collect likes for reasons that have nothing to do with how widely the idea is held. Exclude them from counts, read them separately.
Step four: check the theme survives across videos
This is the step that separates a trend from a mood. Export five to twenty videos — yours, a competitor's, or a hashtag's top posts — and ask whether the theme you found appears in all of them or only one.
The qualitative-research rule of thumb applies well here: small samples find most of what is findable, and the returns flatten fast. Nielsen Norman Group's work on sample sizes makes the case for user testing, and the same logic holds for comment themes — the fifth video rarely surprises you if the first four agreed. What you are buying with more posts is confidence, not novelty.
What the model will get wrong
Three predictable errors, all worth pre-empting:
Sentiment on short slang. "this is criminal" and "im deceased" are compliments on TikTok. Any generic sentiment classifier reads them as negative. If you need sentiment, score it inside a theme — "how do people feel about the price" — never across a whole video.
Bot and spam clusters read as consensus. Dropshipping spam and follow-for-follow rows repeat perfectly, which is exactly what theme-finding rewards. Check whether a "theme" is thirty different accounts or the same five accounts thirty times; the username column answers this in seconds.
Confident summaries of thin data. On a 100-row free export a model will still produce eight themes with counts. Whether those counts mean anything is your call, not its. This is the single best reason to insist on verbatim quotes.
Turning a theme into the next video
A trend you do not act on is a nicer way of scrolling. The conversions we see from this work are boring and specific: the repeated question becomes the hook of the next video, the repeated objection becomes the thing addressed in the first three seconds, the migrated vocabulary goes into the caption and the on-screen text, and the unserved demand becomes a product decision or at minimum a part two. Creators mining their own back catalogue tend to find the same thing — the audience has been asking for one obvious video for months.
Agencies run the same loop against competitors instead, where the objection list is a gap analysis you did not have to commission. Either way the file is the deliverable; the analysis is just how you read it faster.
The permission question, answered once
Two distinct questions. On collection: in the US, the Ninth Circuit held in hiQ Labs v. LinkedIn that scraping publicly available data is not unauthorised access under the CFAA. On retention: a comment with a display name attached is personal data under the GDPR when a commenter is in the EU or UK. Neither of those forbids ordinary audience research. Both argue for keeping the file as long as the question needs it and no longer, and for not turning a comment section into an outreach list.
Nothing in this workflow touches private content, a login, or anything the app does not already show a logged-out visitor.
Download TikTok comments free →
Related reading: how to analyse YouTube comments, mining TikTok comments for UGC, analysing a competitor's TikTok campaign.