Why Your AI Keeps Making Things Up About Your Business

Your AI makes things up about your business because it has never been told the truth about your business in a form it can read. It isn't lying and it isn't broken. It has a gap where your company should be, and it fills that gap with the most likely-sounding answer — which is an average of every business that looks vaguely like yours.

Once you see it that way, the fix stops being "write a better prompt" and starts being a much more ordinary question: where is the real answer written down?

What this actually looks like

You've probably seen at least one of these:

It confidently describes a service you don't offer.
It quotes a price you have never charged.
It tells a customer you serve a town two hours away.
It writes a warranty or cancellation policy that sounds reasonable and isn't yours.
It gets one thing wrong in an otherwise perfect email, and you only catch it because you happened to read closely.

That last one is the dangerous version. Obvious errors get caught. It's the plausible ones that go out to customers.

Three reasons, in plain language

It never knew in the first place. These tools learned from an enormous pile of public text. Your prices, your service radius, the way you handle a callback in February — none of that was in the pile. When you ask about it, the model doesn't look anything up. It produces what a business like yours would probably say.

It can't tell you that it doesn't know. This is the part people find hardest to believe. The same machinery that produces a correct answer produces a wrong one, and it produces both in the same confident sentence structure. Confidence is not a signal of accuracy. It's just the house style.

Pasting into the chat isn't memory. Most people discover that if they paste their price list into the conversation, the answers get dramatically better. They do — for that conversation. Next week, in a new window, it's gone, and you're back to the average. Doing that paste every single time is a job, and it's the job most people quietly give up on around week three.

The uncomfortable reason

Here's the one that actually matters, and it isn't about AI at all.

In most small businesses, the answer the AI got wrong isn't written down anywhere consistent. The website says one thing. The proposal template says something slightly different, because it was written two years later. The way you actually do it now is a third version that lives in your head and has never been written down at all.

Hand a machine three versions and it picks one. So does a new employee. So does the person covering your phone while you're on a job site. The AI didn't invent your inconsistency — it's just the first thing that ever read all of your material at once, at speed, and answered from it without asking a follow-up question.

That's genuinely useful information, if you can stand to hear it. A confidently wrong AI answer is a very fast, very cheap audit of what your business has actually written down.

The fork

When it gets your business wrong, you're standing at a fork, and almost everyone takes the same path first.

Path one: manage the AI. Longer prompts. Detailed instructions at the top of every chat. Re-pasting the same background. Correcting it in the moment, then correcting it again next week. This works, a little, and it never stops costing you. You've hired an assistant who forgets everything each morning and gets re-trained daily, by you, forever.

Path two: fix what it reads. Write the real answer down once, in one place, in a form a machine can retrieve every time — and make the system point at that place instead of guessing.

Path one treats the symptom. It's not stupid; it's just permanent. Path two is more work up front and then it's done, and the same work makes your business easier to hand to a human too.

The tell is repetition. If you find yourself typing the same context into a chat window for the third time, that context should be a document, not a habit.

What I got wrong on my own site

I should be straight about this, because I spent a while being annoyed at a problem I had personally caused.

I sell exactly this — structured business knowledge that AI systems can actually use. And for the first three weeks my own website was live, Google had indexed exactly one page of it. My business had no Google Business Profile at all. Every page except the homepage carried no structured data whatsoever — nothing telling a machine what this company is, where it works, who runs it, or what it sells.

So if you had asked an assistant in July what Barnicle Productions does, there was almost nothing authoritative for it to read. Whatever it told you would have been a guess assembled from fragments. And I would have been irritated by a wrong answer that I had done nothing whatsoever to prevent.

It gets slightly worse. I also run a photography business under my own name, so even the accurate information about me is split across two identities with the same name on them. If I'm hard to resolve, that's my doing, not the machine's.

I fixed it the same way I'd fix it for a client: write the facts down in a form machines read, put them in one place, and make sure the place is reachable. Which is a much less exciting sentence than "prompt engineering," and works considerably better.

What actually fixes it

The unglamorous version, in order:

1. Pick the questions that matter. Not everything — the ten or twenty things you answer over and over. What you do, what you don't, where you work, how you price, what happens when something goes wrong, who does what.

2. Write one true answer for each. One. Not a version on the website and a different one in the proposal. If two sources disagree, that's not an AI problem to solve later; that's a decision to make now.

3. Put them somewhere a machine can read. Plain, structured, boring files. Not a slide deck, not a PDF of a scan, not in your head. The format matters less than the fact that it's one place and it's current.

4. Make the AI answer from that, not from memory. This is the real technical step, and it's what separates a system from a chat window: the tool retrieves your actual documents when it answers, instead of recalling an average.

5. Keep a person in the loop where it counts. Anything going to a customer gets a human's eyes first. Not because the system is untrustworthy, but because that's how you find the gaps — and every gap you find gets written down, which makes the next answer better.

That's the whole method. It isn't clever. It's the same reason a good employee handbook makes a new hire useful in a week instead of a quarter.

This is what an AI Business Brain is, if you want the version with a name on it — your business's real answers, structured so that every tool you point at them gives the same answer you would. If you want the deeper technical treatment of how that's organized, there's a field guide, though fair warning, it's written for people who enjoy this sort of thing.

If you take one thing

Next time an AI says something wrong about your business, don't start by rewriting the prompt. Try to answer the question yourself — and then go find the document you'd point a new employee to for that answer.

If you can find it in under a minute and it's current, you have a retrieval problem, and that's a fixable technical thing.

If you can't find it, or you find three versions that disagree, then the AI wasn't making things up. It was reading exactly what you gave it.