You read every one-star review last night.
You still cannot say what to fix first.
That is not a discipline problem. Forty complaints about the clasp, twelve about sizing, nine about the box arriving crushed, and a scattering of things nobody could do anything about, all arrive in one list, in the order they were written, with no counts on them. So you fix whichever one you read last, or the one from the angriest customer, and three months later the clasp is still breaking. What is missing is arithmetic on the pile: which complaint, how many, and can you actually move it.
The pile, sorted by what caused it
Small, tight and runs narrow are one complaint
Word counting splits them into three, and none of the three looks big enough to bother with. Grouping by what the buyer meant puts them back together, and sizing turns out to be 19% of your negatives rather than three separate 6% items you kept deciding to leave until next month. Each group comes back with the raw quotes under it, so when a count surprises you, you go read the actual reviews and check.
A misleading photo is not a customer service problem
Every group gets tagged with who would have to move: the copy, the images, the product, the packaging, or the reply template. That matters because the costs are nothing alike. Rewriting a bullet is twenty minutes. Changing the clasp is a new mould, a sample round and eight weeks. Knowing which bucket a complaint falls into before you start is most of the decision already made.
Then it writes the fix, not a chart
The sizing group goes to the listing skill, which drafts the bullet stating the real measurement and the FAQ entry that heads the question off before it becomes a return. The photo group goes to the creative skill for a replacement image brief. Both land as to-do items with the evidence still attached, so a month later you can see what you actually shipped and whether that count came down.
Where review analysis goes wrong
These four are written into the skill itself, because each one is a way to produce a confident number that means nothing.
Forty reviews is not your reputation
Three weeks on a slow-moving SKU is a sample too thin to carry a percentage, and a file with only the one-stars in it is not a sentiment distribution however you slice it. When the numbers cannot support the question you asked, you get told that instead of a chart.
It cannot tell you which ones are fake
Text alone gives you patterns that look coordinated, and that is all it gives you. Anything sold as fake-review detection from review text is a guess with a confidence score painted on it. If you actually suspect manipulation, the route is a report to the platform, not a dashboard.
Nothing is scraped on your behalf
By default it reads the file you put in the folder and nothing else. Automatic collection stays off until you confirm the platform terms and your own authorisation, and logins, captchas and rate limits are never worked around. In practice that matters less than it sounds, since Voice of the Customer plus the review export covers most of it.
Shipping a fix is not a promise of a rating
You get the complaint, its share, the evidence and what would have to change. Whether the rating moves depends on your volume, on how long the old reviews keep weighting the average, and on whether you fixed the real cause. Anyone promising you a star is selling something.
Questions from sellers who already have the export open
Do I have to connect an API, or can I just drop the CSV in?
Drop the CSV in. That is how most people run this and it needs no setup at all. There are connectors for Seller Central, Shopify, eBay, Etsy, Walmart and WooCommerce if you want the download step gone, but every one of them means registering a developer application on the platform first, so it only pays off if you are pulling data weekly.
I already paste reviews into ChatGPT. What is different?
Two things, and neither is the model. A paste is capped by what fits, so on 800 reviews you are summarising a sample and calling it the whole; here it reads the file off your disk. And a chat ends when you close it, whereas these groups become to-do items and the bullet rewrite happens in the same place, so next month you can look at what you changed instead of starting over with a fresh summary.
Where does buyer data go?
Names, order numbers and addresses are stripped before analysis starts, and the files stay in your folder and are read from disk. What leaves the machine is whatever the agent quotes to the model provider you picked, which is your own OpenAI or Anthropic account if you connected one and ours if you are on a plan. The source is public under MIT if you would rather check that than take our word for it.
Can I run my competitor's reviews through it?
Yes, and it is often more useful than your own, because their unfixed complaints are your bullet points. Put both files in and the groups get compared side by side, which separates what is wrong with your product from what is just how the category behaves. Same rule on getting hold of it: whatever the platform lets you export.
What does it cost?
The app is free and open source under MIT. Model calls cost money for everyone including us, so either connect your own OpenAI, Anthropic or Google account and pay them directly, or take an Orkas plan. Nothing is priced per review or per report, which matters when the file is four thousand rows.
The export is already in your downloads folder
Free, macOS and Windows. Use your own model key or an Orkas plan.