Why AI makes up domain availability

What is the model actually doing?

Predicting the next word. It has read a great many pages where a domain name is followed by “is available” or “should be free”, and it reproduces that pattern. It has no record of registrations, no connection to a registry, and no way to know that someone registered the name last Tuesday.

This is not a flaw specific to one model. Every language model without a tool behaves this way, because none of them has a source to be right from.

Why does it sound so sure?

Because the training text sounds sure. Domain suggestion pages rarely say “I have no idea whether this is taken”, so the model doesn’t either.

Models are also tuned to be helpful, which nudges them toward the answer you’re hoping for. Ask “is this available?” and “yes, it appears to be” is the helpful-sounding reply. The shape of your question does part of the work.

How do I test it in 30 seconds?

Ask the model for ten “available” domain names for any business. Then look each one up at the registry over RDAP, or on any tool that queries the registry rather than DNS. Count how many were already registered.

Do it with your own brief rather than taking anyone’s word for the number, including ours. The point of the test is that it takes half a minute and the result is yours.

What’s the fix?

Connect the model to a checker. Claude, ChatGPT and most coding assistants can call external tools through connectors or the Model Context Protocol. Then the model writes the names, the tool checks each one at the registry, and the model reports the tool’s answer instead of guessing.

The tool has to be strict, or you’ve just automated the guess. It should return three states — taken, registrable, unverified — and it should never turn a timeout into “available”. That distinction is the whole job.

Does this mean AI is bad at naming?

No. It’s good at naming and bad at checking, and those are different jobs. Use it for the first and something with a registry connection for the second.

If you’re going to use a model for the writing half, it’s worth briefing it properly — five prompts that produce names worth checking, including the one instruction that stops it labelling everything “available”. Then see how the checking half works.

Plainly

Language models predict text and have no connection to domain registries, so their availability claims are guesses. Connect a checker that queries the registry, and the guessing stops.

What Webault does about this

Webault uses a model to write names and then checks every one at the registry before showing it. The model never decides whether a name is free; the registry does.

Questions

Related reading

• Prompts that get better name ideas from any LLM — five prompts that produce names worth checking.
• How we check 532 registries without guessing — why the check is RDAP and not DNS, and what a timeout is allowed to mean.
• Why did a domain show as available, but I couldn’t buy it? — the eight reasons, and how to spot each one.