Customer reviews → actions · 4,173 Amazon food reviews · 14 categories
Fix what costs stars. Sell in your customers' words.
I read every review sentence by sentence, not by its star rating, to answer three questions a brand team actually acts on: what to fix first, what to stop worrying about, and which words to sell in. Three charts, then five plays with an owner each.
01 — Where are the complaints?
Inside the good reviews. 1 in 6 sentences in a five-star review is a complaint
So what Mine your 4★ and 5★ reviews for defects. Any method that labels a sentence by its review's stars counts these as praise, and they're the most fixable complaints you have: from customers who still like you.
Share of sentences scored negative, by the review's star rating. Each sentence is scored on its own (RoBERTa), across 9,954 sentences.
Show the data
| Stars | Complaint sentences |
|---|---|
| 5★ | 17.5% |
| 4★ | 26.5% |
| 3★ | 55.4% |
| 2★ | 68.8% |
| 1★ | 79.2% |
02 — What should we fix first?
Rank by cost, not by volume: taste for reach, arrival condition for severity
So what Put taste consistency first: it's in 1 review in 5 and carries the most total drag. Then arrival condition: rarer, but each one costs the most. Don't chase price complaints just because they're loud.
Each bubble is a complaint theme. Across: how often it shows up. Down: how much lower a review with it rates (lines show 95% intervals). Bubble size and the dashed curves: total drag on the average rating. Review-level OLS with category and year fixed effects; an association, not a forecast.
Show the data
| Theme | Share of reviews | Stars when present | 95% CI | Drag on average |
|---|---|---|---|---|
| Taste & flavour | 21.9% | −0.63★ | −0.76★ to −0.51★ | −0.139★ |
| Condition & expectations | 5.5% | −0.84★ | −1.05★ to −0.63★ | −0.046★ |
| Ingredients & nutrition | 11.5% | −0.33★ | −0.48★ to −0.18★ | −0.038★ |
| Price & value | 4.0% | −0.54★ | −0.75★ to −0.33★ | −0.022★ |
| Packaging & portioning | 5.4% | −0.29★ | −0.48★ to −0.11★ | −0.016★ |
| Availability & sourcing | 3.4% | +0.22★ | +0.02★ to +0.43★ | +0.008★ |
03 — Where does each category lose stars?
Taste leads in 13 of 14. Protein snacks lose to logistics, dairy to ingredients
Stars lost to each theme, by category
lessmore stars lostlifts the rating
The plan for one category
Coffee · 634 reviews · 4.20★ average
634 reviews, averaging 4.20★. Taste carries the category — and the complaint vocabulary says exactly what goes wrong with it: the word that most distinguishes a low-rated coffee review from a high-rated one is “weak”.
Fix first
- Taste−0.16★
- Arrival−0.05★
- Ingredients−0.03★
Lead with
gooddrinkgreatmorningtriedsweetrichblend
Words that flag trouble
weakservingdescriptiongoingcoffee tastesmachineartificialtastes
“i know organic has a different taste, but this is bad.”2★ review · Taste
So what Same playbook almost everywhere, with two exceptions worth their own owner: protein snacks need a supply-chain fix (dates, seals, “not as described”), dairy a formulation one. Click any category for its plan.
Cells: estimated stars each theme takes off the category's average rating (complaint rate × cost per complaint). Blank where the theme isn't a measurable driver in that category. Categories sorted by average rating, lowest first.
Show the data
| Category | Reviews | Avg ★ | Taste | Arrival | Ingredients | Price | Packaging | Can't find it |
|---|---|---|---|---|---|---|---|---|
| Protein snacks | 137 | 3.88 | −0.171★ | −0.116★ | −0.043★ | — | — | — |
| Other | 249 | 3.90 | −0.076★ | −0.064★ | — | — | −0.024★ | — |
| Dairy | 74 | 4.06 | — | — | −0.067★ | — | — | — |
| Condiments | 236 | 4.13 | −0.196★ | — | −0.071★ | −0.034★ | — | — |
| Drink | 436 | 4.16 | −0.154★ | −0.037★ | −0.057★ | −0.035★ | −0.019★ | +0.008★ |
| Baby & pet | 431 | 4.17 | −0.112★ | — | −0.028★ | −0.021★ | −0.015★ | — |
| Coffee | 634 | 4.20 | −0.157★ | −0.054★ | −0.028★ | −0.025★ | −0.017★ | — |
| Candy | 285 | 4.27 | −0.178★ | −0.053★ | −0.032★ | — | −0.015★ | — |
| Baked goods | 91 | 4.28 | −0.146★ | — | — | — | — | — |
| Chocolate | 103 | 4.29 | −0.178★ | — | — | — | — | — |
| Snack | 558 | 4.34 | −0.135★ | −0.054★ | −0.042★ | −0.015★ | −0.018★ | +0.008★ |
| Seasoning & spice | 112 | 4.37 | −0.170★ | — | −0.068★ | — | — | — |
| Staples | 256 | 4.38 | −0.139★ | — | −0.053★ | — | −0.020★ | — |
| Tea | 334 | 4.38 | −0.163★ | — | −0.029★ | — | −0.017★ | +0.014★ |
04 — So what do we do on Monday?
Five plays, an owner for each, sized by the evidence
- 1
Make the back of the pack match the front
“the first ingredient in this formulation is mineral oil — olive oil is the fourth”
Brand · Regulatory
−0.06★206 reviews · 5 categories - 2
Publish brew strength and machine fit on coffee listings
“very weak … they recommend brewing at the 6 oz size, and even that is still weak”
Ecommerce · Content
−0.16★157 reviews · largest category - 3
Disclose country of origin on pet lines
“i am downgrading from 5 star to 1 star because they are made in china”
Sourcing · Brand
−0.11★112 reviews · strongest signal - 4
Reformulate two tea SKUs: peach and green
“tastes like it was flavored with peach-scented soap”
Product dev
−0.16★86 reviews · 2 SKUs - 5
Audit date rotation and seal integrity
“i returned these because the expiration date was the day i received them”
Supply chain
−0.12★59 reviews · 4 categories - +
Stop treating “I can't find it” as a complaint
“my favorite jerky was the world kitchens brand — but it's no longer available”
Channel · Supply
+0.22★rating above average · p = 0.03
Claims that cost you: what the copy promised, and the reply
- “Sugar free” · candy“bought this for my diabetic uncle … it ended up having tons of sugar”
- “Buttery” · snacks“not buttery as advertised; they just taste greasy with little flavor”
- Flavour naming · snacks“should have been called vinegar and sea salt — not sea salt and vinegar”
- Origin omitted · protein“nowhere in the product description is the country of origin listed”
So what Each play names who owns it, how many reviews it touches and the estimated star drag it carries. Impact is a triage order, not a promise that fixing it moves the rating by that much.