We give every client the same test twice.
The first one is open book. The model is allowed to go search. It reads the site, comes back, tells you about the business. That is what is happening when you type a company name into ChatGPT and something correct comes out the other end. It went and looked while you watched, and it felt like being known.
The second one is closed book. Same model, same question, same afternoon. We switch the searching off. Now it has to answer out of whatever stuck the last time one of these companies sent a crawler through your website.
Because that is the other way in, and it is the one nobody explains. These companies send robots out to read websites and haul the text home. GPTBot. ClaudeBot. PerplexityBot. They show up in your server logs with their names on them, and there is more than one kind doing more than one job, which is a whole conversation I am saving for next week. What matters this morning is that whatever they took on their last trip through is all the machine has to work with when you take its internet away.
Open book, seven of our eight clients come back described correctly.
Closed book, seven of eight are wrong.
Not fuzzy. Wrong.
I did not believe it. Not because the number looked too big, but because I have learned the hard way that AI lies to you, and my first move now is never to take the answer. It is to go after it. So I grilled it. Ran it again, ran it a third time, made a different engine check the work. That is not me being careful. That is a reflex I got the expensive way, and I have it because eight businesses pay me to know how this stuff actually behaves instead of how it advertises itself.
It held. Seven of eight.
And then I wanted to know why, because why is the only part that changes how we work on Monday.
You cannot hear the difference
Here is what bothers me about it.
Both answers sound exactly the same.
A model reading a page it found four seconds ago and a model repeating something it learned in March produce identical prose. Same rhythm, same certainty, same eager little helpful tone. Nothing in the answer tells you which one you got. There is no seam to find.
And those two things have completely different weather. One holds up when your site is slow, when a page moves, when a crawler gets turned away at the door, when somebody redesigns and forgets the redirect. The other one does not, because it was never a fact about your business. It was a fact about whether your door happened to be open at the moment somebody knocked on it.
Worth finding out whether AI can reach you at all before you spend a nickel worrying about what it thinks of you.
The wrong answers have an address
When one of these gets a business wrong, it is not making it up from scratch. It goes to the nearest bucket to your name and reads out of that one instead.
We have a client called AE Training. Closed book, ask about them, and you will get Account Executive training. Or Adobe After Effects. Or Aerospace Engineering. Or Adverse Events, which is a pharmaceutical term. It is not reaching randomly into the whole world. It is standing on a street where every mailbox says AE, knocking on one, and reporting back whoever answered.
Then there is Royal Property Services, in Lee’s Summit. General contracting. Remodels, additions, the kind of job that ends with somebody’s new kitchen.
Ask an AI about them and you get cleaning. Janitorial services. Carpet care. Property turnovers.
Which is a strange thing to invent, so it did not invent it. There is a real Royal Property Services in Las Vegas that does exactly that work, and what comes back about my guy is their marketing copy, almost word for word. Enhance the appeal and functionality of any property. Detailed cleaning and carpet care. Professional janitorial services.
That is a Nevada company’s sentence, arriving in an answer about a man who remodels kitchens in Missouri.
It did not dream you up. It grabbed the wrong box.
Put “in Kansas City” in the question and it finds the right company every single time. Take the city back out and it drives to Nevada. One city name was the only thing standing between two businesses that have never heard of each other, and most people never type a city.
What I take from this
Calling it hallucination had me picturing a machine inventing things, which made it feel unfixable and a little spooky. It is not that, and it is dumber than that.
Somebody asked a question, more than one business answers to that name, and the model picked one and committed. Then it stayed perfectly consistent about the wrong pick, which is exactly why it reads as confidence instead of confusion.
Tell three people in this town to meet you at Joe’s.
One of them goes to the gas station on 47th. One goes to Olathe. One goes to Leawood. Nobody invented a Joe’s. All three did exactly what you asked. There is just more than one and you never said which.
And if one of them is not from here, he is halfway to Tulsa, because Oklahoma Joe’s is a different restaurant owned by a different guy. That is not a hypothetical either. Getting mixed up with the other one is why Joe’s changed its name in the first place.
The difference is that those three text each other and sort it out in four minutes. The machine never texts anybody. It sits in the wrong parking lot in that same steady voice, telling everybody who asks that it is at Joe’s.
The engines that handle it best are the ones willing to say so. We asked Claude about a client with a common name and it came back with there are likely multiple businesses with this name operating in different regions, could you give me more context. That is the correct answer, and it is the one that looks least impressive, and I think that is the whole trap. The engine that made something up sounded smarter and was worse.
Next Tuesday I want to get into the robots themselves. I was talking with a guy this weekend who is considerably deeper in tech than I am, and he had no idea there is more than one kind of crawler on his website, or that they do different jobs, or that one of them decides whether an AI will ever know him at all. If he did not know it, I cannot work out why anybody else would.
Until then, before you go asking an AI what it thinks of your business, ask it something easier. Ask it what you do for a living.
Then find out whether it is describing you, or a guy in Las Vegas with a carpet machine.
See you next Tuesday.
Nathan. Lefty Media Co, Kansas City.