Ask ChatGPT "is this worth buying at $12 if it sells for $29.99?" and you get a confident answer with a clean ROI percentage. That percentage rests on fees the model made up.
I sell over $100K a month on Amazon, and every buy I make clears a fee calculation first. The arithmetic was never the hard part. The inputs are: which referral rate applies, which size tier the product lands in, what Amazon added to the fee schedule this year, and what it costs you to prep and ship one unit in. A chatbot working from memory gets those inputs wrong, and on an arbitrage deal a two-dollar miss eats the profit.
Below: where the guesses happen, and how to route the math through a real calculator so the AI keeps the conversation and the calculator keeps the numbers.
Where a chatbot's fee math breaks
A language model learned FBA fees from whatever it read during training. That text included old fee tables, blog posts with outdated examples and forum threads arguing about tier cutoffs. When you ask about a deal, the model reconstructs a plausible fee from that soup.
Four inputs go wrong over and over.
Referral rates depend on category
Amazon charges a referral fee as a percentage of the sale price, and the percentage depends on the product's category. Some categories also have tiered rates based on price, plus a per-item minimum. Amazon publishes the current schedule on its selling pricing page.
The model has no idea which category your ASIN sits in unless you tell it, and even then it is recalling a rate, not looking one up. Pick the wrong category and the referral fee is off by several points of the sale price.
Size tiers depend on exact dimensions
The FBA fulfillment fee is set by size tier and shipping weight. A product a fraction of an inch over a cutoff jumps a tier and costs more to fulfill. The chatbot does not know the packaged dimensions of the item you are looking at. It guesses "small standard" because that sounds right for a toy or a bottle of lotion, and moves on.
Surcharges change every year
Amazon adds fees on its own schedule. On April 17, 2026 it started a 3.5% fuel and logistics surcharge on FBA fulfillment fees in the US and Canada, about 17 cents per unit on average according to Retail Dive, with no end date. Amazon ran a similar 5% surcharge in 2022. A model trained before April 2026 does not know the current one exists.
Then there is Q4. Amazon charges a holiday peak fulfillment fee for part of the season, which matters most when you are buying for it. A model has no way to know whether the sale you are planning falls inside that window this year.
Your own costs are invisible
This is the one that hurts most. Amazon stopped US FBA prep and item labeling on January 1, 2026, and since March 31, 2026 every arbitrage seller puts an FNSKU barcode on every unit. That means every unit you send in carries a prep cost you pay, either your own time and labels or a prep center invoice, plus inbound shipping to the warehouse.
The chatbot does not know your prep center's rate or what your last shipment cost per unit. It leaves those costs out, or it makes up a round number. Either way the profit it reports is a number you will never see in your settlement.
What a one-dollar miss does to ROI
The math structure below uses made-up round numbers to show why small errors matter. None of these fee figures are real rates. They show the shape.
Say you buy a product at $12 and it sells for $29.99.
- Sale price: $29.99
- Minus referral fee (category rate times sale price)
- Minus FBA fulfillment fee (size tier and weight, plus any surcharge)
- Minus prep cost per unit (your labels, your prep center)
- Minus inbound shipping per unit
- Minus buy cost: $12.00
- Equals net profit. ROI is net profit divided by buy cost.
Now suppose the real net profit is $4.00. That is 33% ROI on a $12 buy, a deal I would take.
If the chatbot drops a $0.60 prep fee and lowballs the fulfillment fee by $0.40, it reports $5.00 instead. That reads as about 42% ROI. You buy 40 units on a number that was never real, and the actual deal sits closer to the floor where I walk away.
Flip it the other way. If the model overstates fees by a dollar, it tells you to pass on a deal that clears your bar. You never see that loss in a report, but you feel it in the deals you skipped.
On a $12 buy, every dollar of error moves ROI by more than eight points. That is why I never trust a fee estimate I cannot trace back to the product's real category, dimensions and my real costs. If you want the full breakdown of every FBA fee and where it shows up, I wrote that up in Amazon FBA fees explained.
The fix: let the AI call a calculator
You do not need to stop using ChatGPT or Claude for sourcing. You need to stop letting them do the fee math. The way to do that is an MCP server, a connector that lets the AI app call real tools with real data instead of answering from memory.
Arbitrage Stalker ships one. The Arbitrage Stalker MCP server exposes a tool called calculate_profit, and it runs on the same calculator behind the app's dashboard, alerts and Chrome extension. When you ask "is this worth buying at $12?", the AI passes the buy cost, the sell price and the ASIN to that tool and reads the answer back to you.
What calculate_profit does
- Resolves fees from the ASIN's real category and dimensions. Without an ASIN, it falls back to a generic rate card and says the result is an estimate.
- Applies the referral rate for that category, the FBA size-tier fee and closing fees.
- Pulls your saved prep and inbound costs from Settings, unless you pass different ones for this deal.
- Handles FBA or FBM. For FBM it estimates outbound shipping from weight unless you give it a number.
- Returns net profit, ROI and margin, with ROI as net profit over buy cost.
It is free to call. Ask about fifty deals in a row and it costs you nothing.
Why this beats pasting numbers into a prompt
You could paste your fee estimates into ChatGPT and ask it to do the subtraction. You would still be the one guessing the fees. The calculator removes the guessing from the loop, and the AI goes back to the job it does well: reading the result, comparing deals and telling you which one to look at first.
The numbers also match. A profit figure from the chat agrees with the figure on your dashboard and in your alerts, because the same engine produced both. Check a deal in Claude on your phone, then again in the Chrome extension at your desk, and you get the same answer.
How to set it up
Connecting takes a couple of minutes, and no code if you use claude.ai or ChatGPT.
Claude
On claude.ai, open Settings, then Connectors. Click Add, then Add custom connector. Name it Arbitrage Stalker, paste https://app.arbitragestalker.com/mcp and click Continue, then Add. Click Connect, log in to Arbitrage Stalker and click Allow. The connector shows up in the Claude desktop and phone apps too.
ChatGPT
In ChatGPT, click Plugins, then Add, then Add custom MCP server. Paste the same URL, leave Authentication on OAuth, tick "I understand", and click Create as a plugin. Connect, sign in and click Allow. In a new chat, click + and choose Arbitrage Stalker before you ask. ChatGPT moves these menus between releases, so if a label changed, look for custom MCP servers or connectors. I covered the ChatGPT side in more detail in ChatGPT for Amazon sellers.
Set your prep and inbound costs first
Before you ask anything, open Settings in Arbitrage Stalker and save your real per-unit prep and inbound costs. The calculator uses those defaults on every answer. If your prep center raised its rate or your last shipment cost more per unit, update the number. Garbage in still means garbage out.
Questions to ask once it is connected
These work in plain English. The AI picks the tool.
- "Is B0XXXXXXXX worth buying at $12? Use my saved prep and inbound costs."
- "Run the profit at $14.50 buy cost and $34.99 sell price, FBM, 2.1 pounds."
- "Compare these three ASINs at the buy costs I paste below and rank them by ROI."
- "What happens to ROI on this one if the price drops to $26.99?"
- "Show me the net profit after prep at a $1.10 prep cost instead of my default."
For the bigger picture of using AI for sourcing, from finding leads to checking what you can sell, read AI online arbitrage.
What the calculator does not do for you
It does not know your buy cost until you tell it. A feed row from a tracked seller has a buy box price but no buy cost, so there is no ROI on that row until you supply what you would pay.
It does not predict where the price goes next. It gives you exact profit at a price. Velocity and competition are separate reads, which the AI can pull from the product lookup in the same chat.
It also does not replace your judgment on the inputs. Keep your saved costs honest, check the sell price you are planning around, and use the result to decide.
Get the math out of the chatbot's head
Fee math from memory is the weakest link in AI sourcing. Hand it to a calculator and the rest of the workflow gets better: the AI reads real profit, real ROI and real costs, and the advice it gives you rests on numbers you can trust.
Every paid plan includes the MCP server, and the pricing page lays out Scout, Hunter and Predator. Create your Arbitrage Stalker account, save your prep and inbound costs, and follow the MCP setup guide for Claude and ChatGPT to connect your AI app.
Let the deals come to you
Arbitrage Stalker watches proven storefronts, scores every lead against your numbers, and tracks each deal to sold and rebuy. Capped feeds keep your finds exclusive.
Start your 7-day free trial