Review sentiment

A 4-star rating tells you the customer was satisfied. Sentiment analysis tells you why. Reading review text manually doesn't scale: a business with 500 reviews can't realistically extract patterns by hand. Sentiment extraction breaks the corpus into positive, neutral, and negative comments, then identifies recurring topics within each sentiment group.

This surfaces the actual drivers of ratings. A restaurant might have a 4.2-star average, but sentiment analysis shows that 67% of negative comments mention parking, not food quality. The owner fixes parking availability and watch word-of-mouth improve. The complementary insight — 89% of positive reviews praise staff friendliness — tells the owner what competitive moat to protect during hiring and training.

Sentiment models score review text on a spectrum from strongly negative to strongly positive. The model ingests the full review body, not just the star rating, because text often contains nuance the rating doesn't capture. A 3-star review saying "Great service but overpriced" is mixed sentiment; the model identifies both the positive (service) and negative (price) poles.

Once the sentiment is scored, the second layer is topic extraction. The system identifies which features or aspects generated the sentiment. Comments about "staff" cluster separately from comments about "location" or "wait times." This produces a breakdown: "parking - 34 negative mentions", "customer service - 12 positive", "menu variety - 8 negative". The specificity is what drives action. Knowing 34 customers complained about parking is more actionable than knowing 200 reviews were negative overall.

A disconnect between sentiment and stars reveals important signal. A business with 4.5-star average might have highly polarized sentiment: 40% strongly positive reviews, 35% strongly negative, 25% mixed. That polarization matters because it suggests the business excels for some customer segments and fails for others. Perhaps the restaurant serves excellent food but inconsistent service, or a plumber charges premium prices but delivers premium results.

Reverse mismatches occur too. A 3-star review is factually neutral, but the text is heavily negative with a single redeeming comment. Sentiment models catch this nuance and weight it correctly. Over time, sentiment trends reveal whether the business is trending toward satisfaction or away from it, independent of whether star counts are rising or falling. A business losing stars while sentiment improves suggests customer expectations are rising faster than satisfaction — a warning sign before public reputation damage occurs.

Response rate is a secondary ranking signal, but responding to everything is inefficient. Sentiment analysis enables triage. The highest-leverage response targets negative sentiment from high-volume themes: if 45 reviews cite slow service, responding to the 3 most recent ones signals you're aware of the systemic issue but doesn't prove you fixed it.

Smarter approach: identify the top 3 sentiment themes driving negative reviews, draft response templates addressing each, then attach them to new negative reviews matching those themes. This is what the Review Velocity API produces — review count per month, sentiment breakdown, and top themes. Connect it as an MCP server and ask an agent to draft themed responses; the agent prioritizes reviews by recency and theme volume, not by star count.