Comment sentiment analysis: reading what audiences wrote

The five classifications

Sentiment sorts comments into positive, negative and neutral, with campaign-relevant nuance: a joke, an emoji chain and a genuine product question are not the same kind of neutral. Topics cluster what people talk about, which is where you learn that audiences fixated on the price, the smell or the unboxing rather than the feature the brief emphasised.

Questions are extracted separately because they are commercially actionable: recurring questions reveal what your product page fails to explain. Risk signals, complaints, misinformation and brand-safety issues, are flagged for escalation rather than averaged away into a score. Purchase intent captures comments where someone says they bought, will buy or want the product, the closest thing organic comments offer to a conversion signal.

Scope honesty: what the corpus is and is not

Every comment analysis states its collection scope, because visibility is genuinely partial. Platform APIs restrict comment access for many account types, some platforms expose only a portion of threads, and discussion continues after any collection date. Pretending otherwise would make every downstream number wrong.

So the deliverable is framed precisely: an analysis of the collected corpus, with the number of comments collected, the publications covered and the collection window all stated. Within that scope the analysis is rigorous; beyond it we do not extrapolate. In practice the collected corpus across a multi-hundred-creator campaign is large enough to expose stable patterns, and those patterns, not decimal-point precision, are what drive decisions.

• Corpus size, covered publications and collection window disclosed
• No extrapolation beyond collected data
• Risk comments escalated individually, not hidden in averages
• Patterns reported with example comments as evidence

What brands do with the findings

Comment findings feed three loops. During the campaign, risk signals trigger fast responses and recurring questions get answered in follow-up content. Between waves, topic and sentiment patterns shape the next brief: what to lean into, what to stop saying, which format sparked real conversation. And beyond marketing, purchase-intent and question clusters routinely end up with product and e-commerce teams, because audiences tell creators things they never type into a survey.

On the ORLEN Paczka campaign, comment collection across 170+ publications produced a corpus of 1,652 collected comments whose analysis shipped with the final report, turning a reach story into evidence of how people actually received the service.

The same corpus also carries operational weight inside the platform: flagged complaints route to customer-service follow-up, creator-directed negativity is separated from brand-directed signals, and every classification stays reviewable by a human before it reaches a client-facing summary.