Tool · Customer feedback
AI review analysis: what customers praise and complain about
Paste your reviews, name the aspects you care about, and this AI review analysis tool tells you, review by review, which aspects were praised and which were complained about, with a confidence on each call and a list of the reviews worth reading yourself.
Quick answer · verified 2026-10-11
What does this AI review analysis tool do?
For each review it answers two things: is it positive overall, and for every aspect you named, does the review praise it, complain about it, or say nothing either way. The counts add up into a chart of what customers complain about most, and you can download every verdict as a CSV.
| Per run | up to 50 reviews |
|---|---|
| Aspects | your own, up to 8 |
| Complaints caught | 87% — tested on a public dataset |
| Cost | ≈1 credit a review |
Example: twelve earbud reviews, analyzed
The reviews below were written for this page, then run through the tool on 2026-10-11; this is the saved result, unedited. Load the same example above to run it again.
Reviews
12
Positive
42%
Negative
42%
Read yourself
2
Complaints and praise by aspect
Most complained about: Sound quality (2 of 12)
◀ complaintspraise ▶counted at confidence 0.5 or more; the rest are marked unsure below
Every review
1Sound is crisp and the bass is punchy without drowning the vocals. Easily the best I have owned at this price.
Positive 0.98Sound qualityPrice and value2Battery dies after about two hours. The box says eight. Returned them.
Negative 0.02Battery life3Comfortable enough to sleep in, and they stay put on runs. Sound is fine, nothing special.
Mixed 0.59Sound quality0.61Comfort and fit read yourself4The left earbud keeps dropping its Bluetooth connection every few minutes. Support sent me a reset guide that did not help.
Negative 0.03ConnectionCustomer service5Great value. I paid half what my old pair cost and honestly cannot hear the difference.
Positive 0.96Sound quality0.89Price and value6They hurt my ears after an hour, but the sound is fantastic and the case charges fast.
Mixed 0.50Sound qualityComfort and fit read yourself7Arrived a week late and the charging case was scratched. The earbuds themselves work.
Negative 0.24none of the aspects8Customer service replaced my faulty pair within three days, no questions asked. Very impressed.
Positive 0.93Customer service9Pairs instantly with my phone and laptop, switches between them without fuss.
Positive 0.94Connection10For the money I expected better. The sound is thin and the case feels cheap.
Negative 0.03Sound qualityPrice and value11Fine, I guess. They do what earbuds do.
Negative 0.20none of the aspects12My daughter loves them. I would have picked a different colour.
Positive 0.72none of the aspects
Three things to read in it. Sound quality draws both the most praise and real complaints, so it is the aspect to look at first. The review that hates the fit but loves the sound sits at 0.50 overall and is flagged, which is right: it is a mixed review, not a neutral one. And “Sound is fine, nothing special” was called a sound complaint at 0.61, a call a person could make either way. Low numbers are where to spend your own attention.
How does the AI review analysis work?
Each review goes to Jev, a decision model, as one request with one question per aspect plus one about the review as a whole. Each question has fixed answers, praise, complaint or no opinion, and Jev returns a probability for each instead of writing a summary. That makes every review comparable with every other: the same three answers, every time, and a number that says how sure the call was.
Calls under 0.5 confidence are marked unsure and kept out of the counts, so a vague review cannot inflate a chart. A review that is mixed overall, or has an unsure aspect, goes on the read-it-yourself list. Each review costs about 962 input tokens with six aspects, roughly one credit.
How accurate is AI review analysis? We tested it
Before this page went up we ran the method on 80 restaurant reviews from MAMS, a public set built so that every review names at least two aspects with different sentiments, the hard case. Two designs, same reviews:
| Measure | First design | This tool |
|---|---|---|
| Answers per aspect | 4, with “neutral” | 3 |
| Right, where the review names the aspect | 46% | 77% |
| Labelled complaints caught | — | 87% (47/54) |
| Flagged complaints that were complaints | — | 83% on a manual read |
| Labelled praise caught | — | 57% (20/35) |
The first design failed on “neutral” mentions, getting 11% of them, so the tool asks only what a business acts on. Against the dataset's labels only 59% of the complaints it flagged were labelled complaints. We read all 26 disagreements: 19 were plain complaints the labels missed, like “way overpriced” marked as no opinion on price, and 7 were real mistakes, three of them under 0.5 confidence. Praise is the weak side: it misses quiet compliments. The overall question scored 98% on 100 Amazon reviews in our decision model benchmark.
How should you choose the aspects?
- Four to eight, and none overlapping. In the test, “service” and “staff” as separate aspects split the same complaint in two; merged, they worked.
- Name them the way customers write. “Delivery” beats “fulfilment SLA”: the aspect name is the whole question the model reads.
- One thing each. “Price and value” is one idea; “price and packaging” is two, and a review can praise one while complaining about the other.
What can you do with the result?
Start with the aspect at the top of the complaints chart, read the flagged reviews, and download the CSV to sort or share. Run the same aspects each month and the counts become a trend. For thousands of reviews, the same questions run from code on the Jev API, or as a batch in the workspace.
AI review analysis: common questions
Can I try the review analysis without an account?
Yes. Three runs of up to ten reviews each work without signing in. After that an account has a monthly allowance of credits, and a review costs about one credit.
Does it work on Amazon, Google, app store or survey reviews?
On any review you can paste as text, one per line. It reads English best; other languages work less well, so check a few results by hand before you rely on them.
Is anything I paste stored?
No. Each review is sent to the model, the verdicts come back, and the text is not written to a database or a log. Only the number of reviews and aspects is recorded, for billing.