I gave my AI one rule to stop it from making things up. The first claim it caught was mine.
It’s September 28, 2026, five days after Amazon announced its Selling Partner plugin, the one that puts Amazon’s Seller Assistant inside Claude. I’m writing my response to it, working through the draft with my AI.
There’s a standing rule in my AI’s memory, one I put there myself: never make a claim you can’t back up with data. I added it because AI makes things up, and an AI that makes things up about an Amazon business is worse than no AI at all.
So I hand it a point my team had been passing around in Slack since the launch. Amazon’s MCP doesn’t connect to ads.
It won’t take it. It flags the claim as unverified and asks where it comes from.
I don’t have a source. I have something I know.
I’m a little annoyed. I run a company that builds Amazon data tools. This is my field. But the rule is the rule, and I wrote it. So I tell it to go search the internet and prove me right.
It comes back.
Amazon’s own announcement doesn’t say the plugin can’t touch ads. It doesn’t say it can, either. It lists what the plugin does: listings, inventory, metrics, sales analytics.
The clearest line saying “no advertising” comes from a company that sells a competing Amazon MCP.
And Amazon has a second MCP server, built for Amazon Ads, in open beta since February 2, 2026.
So my fact was true of one product, unconfirmed by Amazon, and wrong about Amazon.
It got messier when I went back over the sources for this piece. A trade publication ran a headline saying the plugin “Supports Ads.” Read the article and it means the assistant can point to an advertising opportunity in a recommendation, not that it runs campaigns. Same question, three different answers depending on whose page you read.
The sentence I published was narrower than the one I started with: “At the time of writing, the plugin’s announced scope covers listings, inventory, metrics and sales analytics. Running ad campaigns, bids and keywords is not on the list.” I can defend every word of that.
I put the rule in to keep the AI honest. I didn’t expect it to hold me to the same standard. It’s accurate to a fault. Sometimes it makes me prove something I already know in my head is true.
And sometimes what I know in my head is only mostly true.
What you “know” about your Amazon business
Every seller runs on a few facts nobody has checked in a while. This SKU is our best seller. That campaign pays for itself. Returns went up because of the new supplier. Most of them are mostly right.
“Mostly right” is how a SKU that loses money after ads, fees and COGS keeps getting restocked.
Agencies have it worse, because their claims go in front of a client. “ACoS went up because a competitor started bidding on your brand terms” is a story until someone pulls the search term report. If the client pulls it first and it says something else, the next report gets read with a red pen.
Here is the part most people miss about AI: it will build on whatever premise you hand it. Tell a chat window your hero SKU is profitable and ask how to scale it, and you get a confident scaling plan. The hallucination didn’t start with the AI. It started with you, and the AI dressed it up.
A no-claim-without-data rule stops both. The AI can’t invent a number, and it can’t quietly accept yours.
Steal the rule
Here’s a version you can paste into your AI’s standing instructions: project instructions in Claude, custom instructions in ChatGPT, or a CLAUDE.md file if you work in Claude Code.
Never state a number, a fact about my account, or a claim about what a tool or platform can do unless you checked it in this conversation. That includes claims I make. If I state something as fact and you haven’t verified it, say so, and tell me what it would take to check. If the data isn’t available, say “I don’t have enough data to answer this accurately” instead of estimating.
Three things make it work:
- “That includes claims I make.” Without that line, the AI holds itself to the standard and lets yours through.
- A way out. “I don’t have enough data” has to be an acceptable answer, or the AI will fill the gap with a guess to keep you happy.
- A second opinion on anything that matters. I have a different model check the first one’s work. Two models rarely make the same mistake the same way.
Expect some friction. The rule will slow you down on things you’re sure about. That’s the cost, and on the ones you turn out to be wrong about, it pays for itself.
The rule needs something to check against
A rule like this is only as good as what the AI can check. For a question about Amazon’s product, that’s the open web. For a question about your business, it’s your own Amazon data.
That’s what the Seller Labs MCP Server is for. It connects AI tools like Claude to a seller’s own Amazon data, with one profit row per SKU per day covering sales, fees, ad spend and COGS. So when the rule tells your AI to prove that your hero SKU makes money, it can run the query instead of taking your word for it, or making up its own.
I wrote the rule to catch the AI’s mistakes. The first one it caught was mine.
