TikTok comment sentiment analysis, done properly

"Sentiment was 72% positive" is a sentence that appears in a lot of reports and survives very few follow-up questions. Positive about what? Measured how? TikTok comments make it harder still: they are short, full of slang and emoji, and often sarcastic. A skull emoji under a joke is praise. "I'm crying" is usually a compliment. Generic sentiment tools get these wrong often enough to matter.

The method below costs an afternoon and produces something better than a percentage: a breakdown you can defend, with the reasons attached.

1. Export the comments

Pull the comments from the videos you care about with the comment exporter, up to 10,000 top-level comments per video. Pick videos deliberately: a brand's own posts, a competitor's, or the top videos on a hashtag will each tell you something different, so keep them in separate sheets rather than averaging them together.

2. Decide what you are measuring

Overall positive or negative is rarely the question. Write down what you actually want to know, then define labels for it. For a product launch that might be:

• Wants it: asking where to buy, asking for a link.
• Likes it: praise without buying intent.
• Objection: price, quality, trust, "is this an ad".
• Off-topic: tags, jokes, comments about the creator.

Four or five labels is plenty. More than that and nobody, human or model, will apply them consistently.

3. Hand-code a sample

Add a label column and code 100 to 200 comments yourself, picked at random rather than from the top. In Excel or Google Sheets, a helper column of =RAND() sorted once gives you a random order. This is the slow part and the important one: it forces you to write down how you handle the awkward cases, and it becomes the answer key for everything after.

4. Scale it with a model

For the rest of the comments, an AI model is the practical option. Paste your label definitions and around twenty of your hand-coded examples into the prompt, including a few of the awkward ones, then send the comments in batches and ask for one label per comment id. Ask for the id back with each label, so you can join the results to your sheet with a lookup instead of trusting the order.

Check your company's rules before sending comments to a third-party model. The comments are public, but some teams still restrict what can be pasted into outside tools.

5. Check the model against your sample

Run the model on the comments you coded by hand and count how often it agrees with you. If it agrees on roughly nine in ten, you can report its numbers with that caveat. If it is lower, look at where it disagrees. The usual culprits are sarcasm and slang, and a couple of extra examples in the prompt often fix them. Do not skip this step: it is the difference between a measurement and a guess with a decimal point.

6. Report it with likes and quotes

Count labels two ways. By comment, each one counts once. Weighted by likes, a comment with 2,000 likes counts for more, because 2,000 people agreed with it. When the two disagree, that is a finding in itself: a loud minority, or a quiet majority. Then put two or three real comments under each label. A client remembers a quote long after they have forgotten the percentage.

For the wider research workflow this slots into, see how to research TikTok comments for a campaign. For getting the file into shape first, see how to export TikTok comments to Excel.