Why pasting your Amazon reports into ChatGPT isn't profit tracking

A bit of background on why MarginLens exists, and why I don't think the "just put your reports in ChatGPT" answer holds up once you've done it for a fortnight.

I sell cushions on Amazon UK. Donut cushions, the ones for piles and tailbone pain. Not glamorous, but they sell, and for a long time I genuinely did not know what I was making on them.

I'm going to talk about two things I tried before building my own tool, because both of them are what people will tell you to do, and both of them left me with the same problem.

The year on Sellerboard

I used Sellerboard for about a year. I want to be fair to it: it does what it says. It pulls your orders and fees in, takes your cost of goods off, and gives you a profit figure for the day, the week, the month. The numbers were mostly right. When I check MarginLens against it now we land within a few pounds of each other, and the gap is usually a setting rather than a mistake.

So why did I stop? Two things it never told me.

The first was cash. Amazon pays out every two weeks, and what lands is not your profit, it's a lump of sales minus fees minus refunds minus whatever ads you ran, from a window that doesn't line up with the calendar. Sellerboard would tell me I'd made £2,400 profit in a month. It would not tell me that £900 of that was still sitting with Amazon, or that the payout landing on Thursday was going to be short because of a refund wave from a bad batch three weeks ago. I was looking at a profit number and a bank balance and they had nothing to do with each other.

The second was VAT, and this is the one that actually hurt. I'm VAT registered. Every £19.99 cushion has £3.33 of VAT in it that isn't mine. So the real question every quarter was "how much of what's in the account do I owe HMRC", and the honest answer was "I'll find out when the accountant does the return". Twice I spent money that turned out to be VAT. Not because I was careless, because nothing I was looking at separated the two.

It cuts the other way as well. Amazon charges VAT on its fees, the referral fee, the FBA fee, storage, all of it, and that VAT comes back off your bill. Sellerboard never told me how much that was. It showed fees as one number. Same with ads: the advertising invoices have VAT on them too, and I'd already paid it, so it's mine to claim back, but the ad spend figure just said £45 and left me to work out what was in it. So I had VAT I owed mixed in with my sales, VAT I'd already paid buried in my fees and my ads, and one profit number sitting on top of all of it. Pulling those apart took a spreadsheet and an afternoon, every quarter.

Profit tracking on its own doesn't solve that. You can know your margin to the penny and still get the VAT bill wrong.

The month on ChatGPT

So I did what everyone's doing now. Downloaded the settlement report, the orders report, the fee preview, dropped them into an AI chat and asked what I'd made last week.

The first answer was good. Honestly, better than I expected. It found the right columns, totted up the sales, took the fees off and explained itself.

The second answer used a different method. I'd pasted a slightly different set of reports and it decided fees should come off before VAT this time. Both answers were defensible. They weren't comparable. By the end of the fortnight I had four numbers for the same week and a chat history I couldn't make sense of.

And it had the same blind spot Sellerboard had, only worse. It didn't know my cost of goods unless I typed it in every time. It didn't know I was VAT registered unless I said. It didn't know that when Amazon grades a returned cushion "customer damaged" I write it off, but when it's "sellable" it goes back on sale. It didn't know the refund on Tuesday was for an order from August. None of that is in the reports. It's in your head, or it's nowhere.

The thing that finally put me off was how confident it was. When it couldn't see the fee events for an order (they arrive about a week after the sale) it estimated, silently, and the number looked exactly as solid as one it had the real fees for. A week later the real fees came in 25p higher per unit and nothing moved, because nothing was watching.

What I actually wanted

Three numbers, every morning, on my phone, worked out the same way as yesterday:

What did I make. Which bits of that are settled and which are still a guess. And of the money that's mine, how much is actually mine once VAT is put aside.

That's it. That's the gap. Sellerboard gave me the first one. ChatGPT gave me a different version of the first one each day. Nobody gave me the other two, and the other two are the ones that decide whether you can pay yourself this month.

So I built it. MarginLens keeps settled and estimated apart on the number itself, so you can see when a day's figure is still moving. It keeps the VAT on sales, the VAT on fees and the VAT on ads as separate lines, so the "put aside" figure is just there, and what you've already paid on fees and ads is counted back, once. And it tells you what's still with Amazon versus what's landed.

Where I do use AI

I'm not anti-AI about this. I use it every day. The point is what it's for.

Adding up fees, VAT, refunds and ads the same way every day is a job for software with rules you set once. It has to be boring. Reading the result and saying "your returns on the grey cushion are mostly 'didn't fit', go and look at the size guide in your images" is a job for something that can read. That bit AI is good at, as long as the numbers it's reading are right.

So in MarginLens the returns analysis has a button that writes out everything the app knows about a product, the return reasons, the listing's own words, your notes, as one prompt. You paste it into ChatGPT or Claude and let it think. Correct numbers in, judgement out. That's the right way round.

If you'd rather keep pasting reports

Fine, but do three things. Use the same reports and the same question every time so the method stays put. Put your cost of goods, your VAT status and your returns rule at the top of every prompt. And don't treat any figure as settled unless the fee events for it are in the file.

If that sounds like a lot to do before breakfast, well. That was my morning too.

Spoken out loud by me, written up and put in order by AI. The opinions, the cushions and the VAT mistakes are all mine.