Review sentiment analysis.
How to score the sentiment in product reviews, why an overall score is usually the wrong unit, and how to get to the aspect-level answer a product team can act on.
- Live reviews, not a sample
- Free result before any email
- No account or install
The short answer
Review sentiment analysis scores the opinion expressed in review text as positive, negative or neutral. Overall polarity is the common starting point but rarely the useful unit, because a single review often praises one aspect and criticises another. Aspect-based sentiment analysis splits the review into what it is about — battery, fit, packaging, delivery — and scores each separately. For a consumer brand, the analysis is only actionable once each scored aspect is tied to the specific product it concerns.
Why an overall score misleads.
Take a genuine pattern from Amazon review data: a product holding a 4.9-star average, with the phrase "stopped working" appearing across a dozen reviews. Overall sentiment reads as overwhelmingly positive, and it is. The defect is real anyway. It sits in the minority of reviews that an average is designed to smooth away.
This is the central problem. Averages answer "how do people feel", which a brand usually already knows. Product and quality teams need "what is going wrong, on which product, starting when" — a different question, and one that a single number cannot answer.
The five steps.
1. Collect the reviews
Pull the review text, star rating, date and verified-purchase flag. The rating and the text disagree often enough that you want both. Review data can be exported with a scraper or read directly by an analyzer.
Compare Amazon review scrapers →
2. Clean, but not too much
Strip markup and boilerplate. Resist heavy normalisation — misspellings, capitals and punctuation carry signal in reviews, and aggressive stemming destroys the phrases you are trying to find.
3. Score polarity
Assign positive, negative or neutral. Lexicon methods are transparent and cheap; transformer models handle context and negation far better. Either way, hold out a sample you have labelled yourself so you know the error rate rather than assuming it.
4. Split by aspect
This is the step that makes the output useful. Extract what each passage is about, then score sentiment per aspect. "Great sound, terrible battery" should produce two findings, not one neutral average.
Aspect-based sentiment analysis, defined →
5. Tie it to a product and a date
An aspect score with no SKU and no timestamp cannot be acted on. Linked to a product identifier and a date, the same score tells you which item changed and roughly when — which is the difference between an observation and an investigation.
Where it usually goes wrong.
Averaging across products. Brand-level sentiment hides the one SKU dragging it down.
Trusting the score over the text. Read the most-helpful critical reviews. They are short, specific, and disproportionately what prospective buyers see.
Ignoring velocity. Twenty negative reviews spread over a year is background. Twenty in a week is an event.
Stopping at measurement. Sentiment that nobody routes to a product or quality owner changes nothing.
Doing this continuously.
Running the analysis once answers a question. Running it continuously across every product and every channel is a different job — ingesting reviews, tickets and returns as they arrive, linking each to the right SKU, and raising the ones that moved.
See how Indellia does this at SKU level → · Compare review analysis tools →
Frequently asked questions.
How do you do sentiment analysis on reviews?
Collect the reviews, clean the text, score each one as positive, negative or neutral, then group by what the review is actually about. The scoring is the easy part — modern models handle it well. The hard part is aspect-level grouping: a review can praise the design and condemn the battery in one sentence, and a single overall score loses that.
What accuracy can I expect?
Overall polarity scoring is reliable on clear reviews and degrades on sarcasm, mixed opinions and domain-specific language. Aspect-level sentiment is harder still. Treat any single score as an indicator, and validate against a sample you have read yourself before acting on a trend.
Do I need machine learning to start?
No. For one product, a hosted analyzer returns aspect-level results immediately. Build your own only when you need it embedded in a pipeline or tuned to vocabulary a general model does not know.
What is the difference between sentiment and rating?
A star rating is the reviewer's summary judgement. Sentiment is what the text actually says. They diverge more often than people expect — a four-star review can contain a serious defect report, which is exactly the signal a quality team needs and a rating average hides.