Understanding What Reviewers Really Mean — Sentiment Analysis in v0.1.12

Release: v0.1.12

Raw review counts tell you a product is popular. Sentiment analysis tells you why — and more importantly, where that product is failing its users.

With v0.1.12, the platform now classifies every scraped review automatically before it feeds the scoring engine. Here's what that means in practice.

What Gets Classified

Every review collected from a proptech supplier directory goes through the new review text classifier. The classifier reads the full review text and returns three things:

A sentiment label — is this reviewer happy, indifferent, or frustrated?
Topic tags — what area of the product are they talking about? Pricing, UX, support, onboarding, integrations, or reliability?
Key quoted phrases — the exact words from the review that best capture the sentiment

All three are stored in product_review_analysis, linked back to the product and the original review.

Why This Matters for Opportunity Scoring

The scoring engine evaluates every product across four dimensions. Sentiment analysis now powers two of them directly:

Market Demand gets sharper. Instead of counting reviews as a raw proxy for demand, the engine now weighs sentiment. Lots of reviews with strong positive sentiment means a genuinely well-regarded product in an active market — a signal that replicating its core value proposition is worth the effort.

Competitive Gaps become specific. This is where the real value is. A negative review tagged pricing is a user telling you the existing product is too expensive or poorly structured. A cluster of support-tagged negative reviews means the incumbent isn't taking care of its customers. The classifier surfaces these patterns at scale — across hundreds of reviews — and the scoring engine translates them into a competitive gap score that tells you exactly where a new entrant could win.

Quoted Phrases in the Dossier

When you drill into a product's full dossier, the competitive weakness analysis section now includes direct quotes pulled from negative reviews. These aren't paraphrases — they're verbatim phrases extracted by the classifier. You can copy them straight into a positioning document or use them as the basis for feature requirements.

Two Classifier Modes

The classifier can run in LLM mode or rule-based mode. LLM mode handles ambiguous, nuanced language more accurately. Rule-based mode is faster and cheaper at scale. The platform selects the appropriate mode based on your configuration — both produce identical output fields in product_review_analysis.

Sentiment analysis is a foundational step toward making the opportunity scores genuinely defensible. The next releases will continue building on this layer.