Take reviews from Amazon's own tools or copy public ones by hand, never scrape, then ask an AI model to group them into themes with counts and example phrases. Compare the themes with your bullets, product and images. The wins are feature gaps to fix, buyer wording to reuse in your own words, and expectations your listing sets wrongly.
Reviews are the cheapest research a seller has. Buyers tell you what they wanted, what disappointed them and what they used the product for, in their own language. The problem is volume: nobody reads 800 reviews carefully. An AI model reads them all, groups them and counts them in minutes. It also makes confident mistakes, so the method below keeps you in the loop at three points.
Where the reviews come from, and where they must not
There are three legitimate sources, and one to avoid.
- Voice of the Customer. Amazon replaced the older Customer Reviews dashboard with Voice of the Customer; trade press reporting the change put the full switch on 30 September 2025, and guides place it under the Performance menu in Seller Central. It rates each product's customer experience health and surfaces what unhappy buyers say about each ASIN. I could not find an Amazon page confirming a CSV export, so check whether your account offers one.
- Customer Review Insights. Amazon's own selling-partner site describes this tab inside Product Opportunity Explorer, under Growth in the main menu: search a keyword or ASIN, pick the product or niche, then open the Customer Review Insights tab. It groups recurring review themes into positive and negative snippets and shows how each relates to star rating, for products and niches beyond your own. It is the Amazon-sanctioned way to study competitors' reviews at scale.
- Reviews you copy by hand. For your own products and a small sample of public competitor reviews, copying text into a document for your own notes is the low-risk route. Amazon's Conditions of Use do not license commercial reuse of its content, so keep samples modest, use them only as research notes, never republish them, copy only the review text, and leave out reviewer names.
- Avoid: scraping. Amazon's Conditions of Use exclude "any use of data mining, robots, or similar data gathering and extraction tools" from the licence to use the site, so browser extensions and bots that harvest review pages breach the terms, and they tend to get blocked. Sellers who buy review datasets from unknown sources inherit the same problem and may not know where the data came from. Do not do either.
Also be clear about what you are not doing: you are not asking buyers for reviews, not changing reviews and not trying to get any removed. Those are separate matters with their own rules, covered in the buyer messaging tutorial.
Prepare the sample
Work on one product at a time. For your own item, copy all the 1 to 3 star reviews and a random 40 of the 4 and 5 star ones; the complaints are where the fixes are, but you need the praise for the language. For two or three competitor products, copy 30 reviews each, spread across star ratings, from the most recent pages so you see current versions of the products.
Make a spreadsheet with one column per field: star rating, date, "verified purchase" yes or no if shown, and the review text. Strip anything that identifies a person. Number the rows. Numbering matters later, because it lets you check any claim the AI makes about "review 17".
The Complete Excel Course with AI Integration
From formulas and pivot tables to Power Query, DAX, VBA and AI with Copilot, ChatGPT and Claude — hands-on from the first lesson.
By Sagnik Bhattacharya and Virginia Thorn, Primetime Emmy Award winner
Buy nowThe first pass: themes with counts and quotes
Say a car repair garage sells a portable tyre inflator and has 85 reviews of its own to study. The prompt below asks for a structure you can verify.
Below are 85 numbered customer reviews of a portable tyre inflator I sell on Amazon. Analyse only the text provided.
1. Identify up to 10 themes that come up more than once. For each theme give: a short name, whether it is mostly positive, negative or mixed, the review numbers that mention it, and one short quote (under 15 words) copied exactly from one of those reviews.
2. List things buyers say they wanted that the product does not do (feature requests or gaps), with review numbers.
3. List any claim reviewers repeat that contradicts what a listing would normally state (for example battery life, noise, accuracy), with review numbers.
4. Do not invent reviews or numbers. If you are unsure whether a review fits a theme, leave it out. State how many reviews you could not place in any theme.
REVIEWS:
1. [text]
2. [text]
...
Asking for review numbers is the key discipline. A model that has to name which reviews support a theme is less likely to invent one, and you can check any theme in a minute by opening those rows.
What the output might look like, and the checks it needs
Theme: Gauge accuracy (negative). Reviews 4, 9, 17, 22, 31, 40, 52, 66. Quote: "reads 3 psi higher than my gauge".
Theme: Cable length (mixed). Reviews 2, 11, 18, 23, 45, 70. Quote: "just reaches the rear wheel".
Theme: Noise (negative). Reviews 6, 13, 29, 44, 58, 61, 73, 81. Quote: "louder than a vacuum cleaner".
Theme: Auto shut-off (positive). Reviews 1, 8, 14, 20, 33, 47, 55. Quote: "set it and walk away".
Gaps: wants a carry case (reviews 5, 19, 36); wants kilopascal display (reviews 12, 38).
Not placed: 11 reviews.
This sample is illustrative. Now check it. First, count: does the number of review numbers per theme match your reading? Open the rows for one theme and see if every one belongs. Second, check the quotes: search your spreadsheet for the exact phrase. Models sometimes tidy a quote, and a changed quote is a fabricated one. Third, check proportions yourself: eight of 85 is about 9 per cent, which is a small group, not "most buyers". Fourth, look at the "not placed" number: 11 of 85 is 13 per cent, so ask the model what those were; sometimes a hidden theme sits there.
Turning themes into decisions: a table you fill in
| Theme | Share of reviews | Kind | What to do |
|---|---|---|---|
| Gauge accuracy | 8 of 85 (9%) | Product issue, possibly one batch | Test current stock against a reference gauge; ask the supplier; do not claim accuracy in the listing until verified |
| Noise | 8 of 85 (9%) | Expectation gap | Add a plain line about noise level to the bullets and images; a buyer who is told is less likely to complain |
| Cable length | 6 of 85 (7%) | Mixed | State the cable length in centimetres and a photo showing reach on a car |
| Auto shut-off | 7 of 85 (8%) | Strength | Move higher in bullets, describe what it does in your own words |
| Carry case wanted | 3 of 85 (4%) | Feature request | Cheap to add to the bundle? Note for the next order |
The "Kind" column is the useful one. A product issue needs a supplier conversation and perhaps a stock check. An expectation gap is solved by the listing. A strength belongs earlier in your bullets. A feature request goes on a list for the next order. Keeping them apart stops you "fixing" a noise complaint by redesigning the product when a sentence in the bullets would have done.
Rewriting bullets from the findings
Take the garage's current bullet: "Powerful portable tyre inflator for cars, bikes and balls". Buyers praise the shut-off, complain about noise and ask about cable reach. A rewrite prompt:
Rewrite bullet 1 and add two bullets for this product. Use only these facts: auto shut-off at the preset pressure; sound level around [verified figure] dB at 1 metre; cable length 3 m; battery or mains as stated on the spec sheet. Write for a buyer who has read the common complaints about noise and cable reach. Do not copy any reviewer's words, make no claim about accuracy, and keep each bullet under 200 characters. Also write a one-line summary of what changed and why.
A reasonable output: "Set the pressure and walk away: the inflator stops by itself at your target, so there is no watching the dial." Then: "Stated plainly: it is loud while it runs, about [figure] dB at one metre, so use it away from the house on early mornings." Then: "3 m cable reaches all four wheels on a family car without moving it." Each bullet addresses a theme with a fact you have verified. The AI added no claim about accuracy, because you told it not to, and the product might not deserve one. For the title and search-term side, see the bullets and search terms tutorial and the 75-character title tutorial.
Competitor reviews: finding the gap you can own
If Customer Review Insights covers the niche, start there, because it does the grouping on Amazon's side. Then paste 30 hand-copied reviews from each of two competing products to fill in the detail, and ask a different question: "What do buyers praise and complain about in each, and which complaint appears in both?" The complaint that appears in both is an opening. Say a cleaning company selling a steam mop finds that buyers of both rivals complain that the water tank is too small for a whole flat, while their own tank is larger. That is a comparison point for the bullets and the main image: "Fills once, cleans a two-bedroom flat" if it is true for your test, and only then.
Be fair in the sample. Do not collect only the worst reviews, and do not paste competitor material into any tool that trains on inputs without checking the data setting. Public reviews are not secret, but your own notes about which product you are studying might be.
Another use: questions in the Q&A and returns
Reviews are not the only text. The questions buyers ask on the listing, and the return reasons in your reports, tell you the same story. A furniture maker found that the same question, "does this need two people to assemble?", appeared 14 times in questions, and a handful of reviews mentioned assembly effort. The answer belonged in a bullet and an image. The same prompt format works: number the questions, ask for themes, and check the counts.
If you track returns, group return reasons from your report the same way and compare them with the review themes. Where both agree, the problem is real; where only reviews mention it, it may be a vocal minority. Tying this together weekly is what the weekly reports tutorial describes.
What buyers were trying to do: a second pass
Complaints tell you what failed. A second pass tells you what the product was hired for. Ask the model to list, for each review that describes a use, the situation in a few words: "topping up before a long drive", "bike tyres for a child", "caravan, no mains socket". Then count them. A pest-control firm selling a reusable mouse trap might expect kitchens and find that a third of the buyers mention garages and sheds. That changes the lead image, the first bullet and the search terms, with no change to the product.
From the reviews below, extract every stated use or situation (where, who, why). Group them. Give the review numbers for each group. Ignore reviews that do not describe a use. Do not infer a use that is not written down.
As with themes, open the rows behind the top group before you act. If the model grouped "garage" and "workshop" and "shed" together, that is a sensible merge; if it merged "garage" with "car" because both appear near each other, that is a mistake.
A before and after for one listing
Here is the whole change for the inflator, side by side (illustrative).
| Element | Before | After |
|---|---|---|
| Bullet 1 | Powerful portable tyre inflator for cars, bikes and balls | Set the target pressure and walk away: it stops by itself, so no watching the dial |
| Bullet 2 | Compact and lightweight | 3 m cable reaches all four wheels of a family car without moving it |
| Bullet 3 | Digital display | Display in psi, bar and kPa; read it with the car door open or shut |
| Image 3 | Product on white | Product beside a rear wheel, cable stretched, showing reach |
| Image 5 | None | Plain sound-level note, only once a measured figure exists |
The kilopascal line came from the two reviews that asked for it. Whether the product really displays kPa must be checked on the unit, not assumed from reviews. If it does not, the finding goes in the feature request list, not the bullets.
Turning themes into an image brief
Reviews also tell you which pictures are missing. The noise theme asks for a sound note, the reach theme for a photo with the cable out, the assembly question for a diagram. A prompt turns the decision table into a brief for whoever makes your images:
From this decision table, write a brief for five secondary listing images. For each: the buyer worry it answers (theme and review numbers), what the picture should show, and the overlay text in under 8 words. Only use facts I list under VERIFIED FACTS. If a picture needs a fact I have not verified, write FACT NEEDED instead of text.
Image rules are tighter than bullets on some points, so check the output against the infographics tutorial before commissioning anything.
Worked through for a small furniture maker
A furniture maker sells a folding desk and has 120 reviews. The AI finds six themes, and one is clearly larger than the rest: 19 reviews say the desk wobbles on carpet, and 14 of those are three stars or lower. The maker checks a sample of those 19 rows: 15 do mention carpet, 4 are about uneven floors generally. The honest number is then 15 of 120, or about 12.5 per cent. The fix is a product change (adjustable feet, a small extra cost per unit) and a bullet that states the surface the desk suits. Without the check, the headline would have been "16 per cent wobble" and the response a redesign. The difference between those two numbers is the whole value of reading the rows.
Where the AI goes wrong on this job
- It merges different things. "Battery" can mean run time, charge time or a dead unit on arrival. Ask it to split a theme if the review numbers read differently.
- It counts generously. A review that mentions noise once in passing is counted the same as one that is entirely about it. Ask for a "main complaint" column to separate them.
- It reads sentiment from stars. A three-star review may be glowing about one feature. Make it read the text.
- It cannot see time. A problem fixed in a new batch last year still shows up in old reviews. Add dates and ask for themes in the last six months separately.
- It over-trusts review language. Buyers say "waterproof" for splash resistant. Do not repeat a buyer's claim as yours.
- Long pastes get cut. A very long review list can be truncated without warning. Split into batches of about 50 and merge the counts by hand, or ask the model to say how many reviews it received.
A monthly routine, with a time budget
- Pull new reviews and questions for your top five products (20 minutes).
- Run the theme prompt on each batch (10 minutes plus checking).
- Check three themes by hand against the review rows (15 minutes).
- Update the decision table and mark what is a product fix, a listing fix or a note for next order (15 minutes).
- Rewrite one or two bullets or images with the verified facts, and record the date of the change (30 minutes).
- A month later, compare the listing's conversion and the new reviews' themes with the old ones.
About an hour and a half a month for five products. The point of recording the change date is that you can later see whether noise complaints fell after you added the line about noise. Without it, you cannot learn from your own changes.
How you know it worked
Watch for three signals over eight to twelve weeks. The theme you addressed appears less often in new reviews. Questions about it fall. And the listing's conversion rate, shown in your business reports, holds or rises. Be cautious reading conversion, because price, season and advertising change it too. If you changed several things at once, you cannot say which one helped, and that is fine as long as you do not claim to know.
Comparing this quarter with last quarter
A single analysis is a photograph; two make a film. Keep the numbered, dated sample and the output of each run in the same folder. Next quarter, run the same prompt on the new reviews only, then ask a short follow-up: "Here are last quarter's themes and counts, and this quarter's. Which themes grew, shrank or are new? Give the review numbers behind each new one." The comparison is where you learn whether a bullet change, a supplier fix or a packaging change did anything.
Take a cleaning company that added a line about the refill pack's smell after a run of complaints. If the smell theme drops from 11 of 90 reviews to 3 of 80 the next quarter, the line probably helped, though a changed formula might also explain it. If it stays at 11, the line did not set expectations and the product itself may need attention. Keep the interpretation modest: two quarters of small samples are suggestive, not proof.
One more habit pays back: write down, before each run, what you expect to find. If the model returns a theme you did not expect, you have learned something. If it returns exactly what you predicted, check whether your prompt led it there.
Rules to keep it clean
Use only the review text you have a legitimate right to read, do not store reviewer names, do not quote reviews in your listing, and do not use any finding to contact reviewers or influence a review. Keep a short note of where each sample came from and the date. If you are unsure whether a source is allowed, leave it out. The aim is better product and a more honest listing; the AI is a fast reader, and you are the one who decides what the findings mean. If a pattern suggests a safety fault, such as an inflator overheating, stop and treat it as a product safety matter, not a marketing one.
Questions about analysing Amazon reviews with AI
Is it allowed to scrape Amazon reviews?
Do not scrape. Amazon's Conditions of Use exclude data mining, robots and similar extraction tools from the licence to use the site. Use Amazon's own seller tools, including Customer Review Insights for competitor niches, and copy a modest number of public reviews by hand only as research notes.
How many reviews do I need for a useful analysis?
A pattern needs repeats. Around 50 to 100 reviews of one product usually shows the main themes, and fewer than 20 shows anecdotes. Treat counts from small samples as leads to check, not as percentages to quote.
Can AI tell me which reviews are fake?
No. It can flag oddities such as repeated phrasing, but it cannot verify who wrote a review or whether a purchase happened. Report suspected policy breaches to Amazon through its own route and do not base decisions on a model's guess.
Should I quote review text in my listing?
No. Use the meaning, not the text. Rewrite a recurring benefit in your own words and only state things that are true of the product. Copying buyers' words, or paraphrasing a claim you cannot support, creates its own policy risk.
Which AI tool should I use?
Any of the main chatbots handles this job, including ChatGPT, Claude and Gemini. What matters is the sample you give it and the checks you run on its counts. Check your plan's data settings before pasting anything that is not already public.
Further reads
- How Small Amazon Sellers Use AI for Listings and Reviews — A wider view of AI for a small seller's listings and reviews.
- Amazon A+ Content With AI: The Built-In Generator and Checks — Use findings in A+ content with the checks it needs.
- What Are Diners Really Saying? Using AI to Analyse Your Reviews — See the same theme method on a different kind of review.
- Amazon Seller Assistant Explained: Check Its Advice First — Check what Amazon's own assistant suggests from your data.
- Is AI Amazon PPC Software Worth It? Break-Even for Small Sellers — Weigh paid ad software once the listing is fixed.
- Amazon A-to-z Guarantee Claims: Reply With AI-Drafted Evidence — A seller-side routine for A-to-z claims: the evidence pack by claim type, an AI-drafted reply, and the order defect rate effect.
- Amazon Search Term Report With AI: Find Wasted PPC Spend — Export the search term report, let AI flag terms with clicks but no orders and group them into a negative keyword list, then decide yourself.
- AI Tools and AI Development: The Complete 2026 Guide — the AI hub, including every tutorial in the AI-for-business series.
Sources: Amazon Conditions of Use and Amazon's sell.amazon.com article on Customer Review Insights (both read 5 November 2026); trade-press reports of the Voice of the Customer switch. Confirm menu locations in your own account.