Many brands still measure engagement by counting views, clicks, session length, likes and metrics like ER, but these numbers themselves cannot be relied upon—if your post or your in-app chat collects all negative comments, this post or chat will still demonstrate the highest ER. Sentiment and semantic analysis, applied to the conversations users are already having inside a product or comments to the brand’s digital outlets, is becoming the tool that makes the level and the polarity of engagement more obvious.
Sentiment analysis has moved well past “positive or negative”
Early sentiment analysis was too straightforward as an instrument: it classified pieces of text as positive, negative, or neutral. Today’s approaches use transformer-based models. Research comparing sentiment classification methods on large social datasets found BERT-based models reaching over 91% accuracy on real-world text, meaningfully ahead of older architectures, particularly on the messy, informal, sarcasm-laden language that actually shows up in user-generated content.
More importantly, the unit of analysis has changed. Aspect-based sentiment analysis identifies which specific thing the sentiment is attached to. The confused onboarding can be revealed with sentiment analysis and can show what exactly helped to solve the issue, analysing different parts of a product. That distinction is what turns sentiment analysis from a vanity metric into something a product or marketing team can actually act on.
Why in-app conversation is a richer signal than reviews or surveys
Surveys and reviews share a structural weakness because they’re prompted, retrospective, and filtered through whatever question you decide to ask. A user leaving a review is answering “how do you feel about this,” after the fact, in a format they know is being read by the company.
In-app conversation, like community chat, comments on live content, and discussion threads, is a different kind of dataset entirely. It’s unprompted, it happens in the moment something is actually occurring, and it’s addressed to other users, not to the company. That makes it a less curated, more honest signal of what people actually think, and — this is the part that matters more — of what they’re about to do. A cluster of confused questions about a specific feature, in real time, is an earlier and more specific warning sign than a drop in a satisfaction score three weeks later.
From sentiment to semantic: behavioural insights that have been already tracked
The most interesting recent work in this space uses sentiment as one input into broader behavioural and trend detection, and semantic analysis becomes a new trend. A 2026 methods paper on early trend detection for small and mid-sized businesses combined transformer-based embeddings, hybrid sentiment and semantic analysis, and topic clustering into a single pipeline explicitly designed to surface shifts in consumer behaviour before they show up in churn.
That’s the real shift underway: sentiment analysis stopped being a brand-monitoring tool bolted onto a marketing dashboard, and started being treated as a genuine input into product and business decisions closer to a leading indicator than a lagging one.
Semantics applied to a live product
- Identifying reasons for friction before they become churn. A spike in negative-sentiment messages clustered around a specific feature or flow is an early signal that works more effectively than a retention dip, which typically shows up weeks after the underlying frustration started.
- Identifying what’s actually driving engagement, not just that it happened. A trading platform, for instance, can look at which instruments or market events generate the most discussion and the strongest sentiment shifts among users, not just which page got the most traffic, but what people were actually saying about it while it happened.
- Prioritising support and moderation resources. Aspect-based sentiment lets a team distinguish “users are frustrated with X” from “users are frustrated in general,” which changes where a team actually spends its next sprint.
The part that has to be handled carefully
None of this works, or should be attempted, without being explicit about what’s being analysed and why. The useful version of this is aggregate and pattern-level — identifying that sentiment around a specific topic shifted, or that a theme is trending across many conversations — not surveillance of individual users’ private conversations. Companies doing this well are transparent in their privacy policies about conversational data being used for aggregate product insight, keep the analysis at a pattern level rather than an individual one, and treat it as an extension of product analytics, not a new category of user monitoring that wasn’t previously disclosed.
The takeaway
The organisations getting real value out of sentiment analysis in 2026 have started treating it as a product signal, sitting upstream of the metrics everyone already tracks. Views and session counts tell you that something happened. Semantic and sentiment analysis applied to what users are actually saying to each other while it happens. So these results allow you to operate and prognose precisely and meet your product with the actual audience needs, not just guesses.
