Document-Level vs Topic-Level Scoring
Document-level sentiment (one score per review) is easy and almost useless. The same 4-star review can carry +0.9 sentiment on staff, -0.4 on wait time, and +0.2 on price. The averaged document score smooths the actionable signal flat. Topic-level scoring decomposes each review into the topics it touches and scores them independently. That is where the operational signal lives.
Industry Sentiment Baselines (2026)
Average document-level sentiment varied by industry. The numbers below are mean polarity scores on a -1 to +1 scale across our 480,000-review cohort.
- Independent hospitality: +0.62 average sentiment
- Healthcare and dental: +0.58
- Restaurants: +0.55
- Home services: +0.52
- Legal services: +0.48
- Small e-commerce: +0.46
- Fitness and wellness: +0.61
Top Negative Topics by Industry
The most-mentioned negative topics shifted by industry. Knowing which topic is dragging your score is the difference between fixing the right thing and chasing a moving average.
- Hospitality: cleanliness (-0.71), check-in time (-0.58), noise (-0.49)
- Healthcare: wait time (-0.74), billing clarity (-0.62), front-desk warmth (-0.41)
- Restaurants: wait time (-0.67), portion vs price (-0.54), staff attentiveness (-0.48)
- Home services: punctuality (-0.69), follow-up communication (-0.58), pricing transparency (-0.51)
- E-commerce: shipping speed (-0.62), packaging condition (-0.55), product fit/expectation (-0.49)
Sentiment as a Churn Signal
For subscription businesses, the sentiment of a customer's most recent review predicted next-period churn at AUC 0.74 in our test set. The strongest single feature was a -0.5 or lower score on a 'value' or 'support' topic, which lifted churn probability 3.6x relative to baseline.
Response Prioritisation
Topic-level scoring lets you triage responses by impact. Reviews scoring highly negative on a high-traffic topic (i.e. one mentioned by 20%+ of recent reviewers) deserve faster response than equally negative reviews on rare topics. Profiles that responded within 24 hours to high-impact-topic reviews recovered 41% of dissatisfied customers in a follow-up survey, versus 12% for slow or generic responses.
Triage by topic, not by stars. A 3-star review touching a high-traffic topic outranks a 1-star review touching a rare one for response priority.
Document-level sentiment scores are smoothed almost flat by topic averaging. The actionable signal lives at the topic level, where a single 4-star review can hide a -0.7 score on the one thing that matters.
Sentiment Drift Over Time
Tracking month-over-month topic sentiment surfaces issues before they show up in star averages. Hospitality profiles that experienced a -0.15 drift on cleanliness sentiment over a quarter saw their star average drop within the next 60 days in 78% of cases. Sentiment is a leading indicator; star average is a lagging one.
Aspect Extraction Accuracy in 2026
Modern aspect-based sentiment models (ABSA) hit F1 scores of 0.81 to 0.87 in our hold-out test, depending on industry. Accuracy was highest on cleanliness, wait time, and pricing aspects (0.85+) and lowest on intangible aspects like 'vibe' or 'overall feel' (0.62-0.68).
What Not To Do With Sentiment Data
Three common misuses appeared often in the cohort. First, treating positive sentiment growth as a goal in itself; the goal is to fix what drags negative sentiment. Second, responding to neutral or positive sentiment with a generic thank-you; this consumes attention without moving the needle. Third, over-weighting sentiment from very recent reviews when the underlying topic mix has changed; compare like-for-like topics, not raw averages.
