Restaurant Guest Sentiment Analysis: What Reviews Actually Tell You

Why star averages mislead

Two restaurants both averaging 4.2 stars can have completely different operational realities:

• Restaurant A: food is loved, service is variable — 80% food praise, 30% service complaint
• Restaurant B: service is loved, food is hit-or-miss — 60% service praise, 25% food complaint

Same 4.2 average. Completely different fires to fight. Sentiment analysis is what separates them.

The taxonomy that works for restaurants

Six categories cover ~95% of restaurant reviews. We tested narrower (3 categories — too coarse) and wider (12 categories — too fuzzy). Six is the sweet spot.

• Food — taste, freshness, temperature, presentation, portion size
• Service — attentiveness, wait time at table, staff attitude, training
• Price — value perception, hidden charges, portion-to-price match
• Ambience — noise, decor, music, lighting, vibe
• Cleanliness — visible hygiene, restrooms, table setup
• Wait — door-to-table, kitchen-to-plate, plate-to-bill

Each review gets scored independently per category, on a -1 to +1 scale, only for categories it actually touches. A review only mentioning food and service gets nulls in the other four categories — not zeros.

The early-warning signals operators miss

Three patterns that sentiment analysis catches before the star average shifts:

1. Aspect divergence

Food sentiment dropping by 0.3 in 14 days while service stays flat = kitchen issue. Service sentiment dropping with food stable = staffing or training issue. The star average won't move for another month because the other category is masking it.

2. New-complaint spikes

If "portion size" was mentioned in 2% of reviews last month and 15% this month, something specific changed — almost always a menu cost decision the kitchen made silently. Star average won't reflect this for weeks because portion-size unhappy reviewers still give 3-4 stars (they liked the taste). Sentiment-by-keyword catches it in week one.

3. Tone shift in praise

Praise mentioning "value" or "reasonable price" dropping while praise of "taste" and "service" stay flat = you've raised prices and the reviewer base noticed. Doesn't show in average ratings (they still love the food) but shows in conversion downstream.

What sentiment analysis can't tell you

• Why the change happened — only that it did. The chef, the GM, or you still need to investigate.
• What competitors are doing — that's a separate competitive intelligence layer, not sentiment.
• Whether to act on a single bad review — sentiment is for trends, not anecdotes. One outlier review is not a signal.
• What price point you should be at — sentiment shows reaction to current pricing, not optimal pricing.

Setting it up without overengineering

Three rules:

1. Pick a tool that gives you 6 aspect categories, not 20. More categories = more noise, harder to spot signal.
2. Set weekly alerts on any aspect moving >0.2 stars week over week. Don't check the dashboard daily; let it page you when something changes.
3. Look at sentiment-by-language separately if your reviewer base is multilingual. Tourists' "service complaints" often mean different things than locals' — pace expectations differ wildly.