Why ChatGPT keeps suggesting domains that are already taken

ChatGPT keeps suggesting taken domains because producing a name and knowing whether that name is registered are two different operations — and nothing in an ordinary chat forces the second one to happen. A language model generates plausible text; registration state lives in a registry that has to be queried at the moment you ask. When those two operations don’t meet, you get a beautifully crafted name delivered with complete fluency and no connection at all to whether anyone can register it.

That’s not a bug awaiting a patch in the next model release, and it isn’t evidence that AI is bad at naming — models are genuinely excellent at the creative half of the job. It’s a category error about what kind of question “is this domain available?” actually is. Understanding the mechanism is worth ten minutes, because it tells you exactly how to use these tools well.

Why doesn’t the model just know which domains are taken?

Because knowing was never part of the operation. A language model’s core job is generation: given your brief, produce text that plausibly satisfies it. Ask for startup names and it does exactly that — often impressively, drawing on everything it has absorbed about what good names look and sound like. But “is this domain registered?” is not a generation question. It’s a lookup question, answerable only by consulting a live registry, and nothing about generating a name obliges that lookup to happen. The name and the availability are produced by entirely different kinds of work, and a chat only guarantees the first kind.

Training data can’t close the gap either, for two reasons stacked on top of each other. First, it’s a frozen snapshot: whatever the model absorbed describes the web as it once was, while domains are registered, dropped, and re-registered continuously. Second — the part people miss — the web was never a registry to begin with. Roughly half of all registered domains host no active website, so a name’s absence from any crawl proves nothing about its availability. Even a model with perfect recall of its entire training corpus would be flawlessly recalling the wrong database.

Doesn’t web browsing solve this?

Less than you’d hope. Even when an assistant can search the web, searching is not the same thing as querying a registry. A search engine indexes content; registration state lives in WHOIS and RDAP, which have to be asked directly. A tool that can fetch URLs could hit an RDAP endpoint — but only if something in the loop decided to, for that exact name. Meanwhile a web search tells you whether content about a domain exists, not whether the domain is registered, and a domain can be registered yet completely invisible to search. When a page does exist, it’s easy to misread: a parked page or a for-sale lander sits on a domain that is very much taken, and summarizing one as “this looks available” is precisely the kind of plausible-sounding mistake a text generator makes.

An authoritative answer requires a tool explicitly wired to the registry — a WHOIS or RDAP lookup, for that exact domain, at that exact moment. And unless the interface shows you the lookup, you have no way to tell — for each of the twenty names in a batch — whether one ran.

That plumbing does now exist, and it’s worth enabling. Registrars have started shipping connectors: GoDaddy’s domain-search MCP server has been available in ChatGPT’s app directory and Claude’s connector directory since early 2026, free and without an API key, and it does query registry data rather than guess. But it only helps if you’ve turned it on and it actually fired for the name in question. The default chat, with nothing connected, still can’t tell you — and that’s the state most people are in when they fall for a name and lose an afternoon to it.

Why are the best suggestions the most likely to be gone?

This is the cruel part: the failure isn’t random — it’s biased toward the names you’ll love most. Models learn what makes a name feel strong from the same corpus of successful brands that everyone else learned from, so they converge on the same constructions: short, pronounceable, two syllables, clean consonants, familiar endings. That space has been systematically mined by founders, domainers, and squatters for two decades. The more closely a suggestion matches the learned pattern of a good name, the more likely a human had the same instinct years ago and registered it.

This is measurable, not just intuitive. When namemy.app ran live registry lookups on 138 Namelix suggestions in June 2026, 42% of the proposed .com names were already registered — and the taken ones clustered among the most pronounceable names at the top of the results. It’s a competitor’s test of a competitor, so weigh it accordingly; the pattern is the point, and the best AI domain name generators for 2026 covers where the field stands. What makes a name brandable is worth internalizing precisely because the model has internalized it too — as has everyone else prompting a similar model with a similar brief. The better it sounds, the longer the odds.

Why does it sound so confident about names nothing checked?

Because the confidence lives in the prose, not in any check. A model has no calibrated uncertainty about registration state — there’s no internal meter that reads “verified” on some names and “guessed” on others. “This one should be available!” is produced by the same process that produced the name itself: it’s simply a plausible sentence to write after suggesting a name. The cheerful phrasing carries exactly the same evidentiary weight as everything else in the reply, which is to say none. Absent a visible lookup, treat any availability remark in a chat the way you’d treat a stranger’s “I’m sure it’s fine” — pleasant, possibly sincere, not evidence.

How do I know if it actually checked?

Ask one deliberately model-agnostic question: did something query an authoritative registry for that exact domain, at the moment you asked? That test is useful precisely because it survives product churn — assistants gain and lose capabilities month to month, so advice built on what any product can do today rots fast, while “did anything actually query a registry just now?” never does. It applies to every tool claiming to check availability, including ours. If the answer isn’t a clear yes, what you’re holding is a candidate, not a name.

So how should you actually use AI for naming?

Exactly as designed: as a divergence engine. The generation half is where language models genuinely shine — they’re tireless, they handle constraints well, they’ll explore ten stylistic directions before a human namer has finished their coffee, and steering them is a skill worth learning (how to prompt AI for name ideas covers the technique). Take that output seriously. Then add the step the chat never guaranteed: paste your whole shortlist into a bulk checker in a single pass — not one name at a time, which is exactly where attachment forms — and verify before you fall for anything. Attachment is the expensive failure in naming — once a name feels like yours, discovering it’s parked behind a four-figure price tag pushes people into bad compromises, and what to do when the domain you want is taken is a page nobody enjoys needing. Checking domain availability in bulk walks through the verification pass, and the broader decision process lives in how to name a startup.

We should be direct about our own position here: Domain Yoga is an AI product too, and it’s subject to the same convergence — we can’t invent scarcity away, only stop you falling for names that don’t exist. The difference isn’t a smarter model, and it isn’t live-checking alone, since plenty of free generators do that. It’s the combination: a large batch, checked and ranked in one pass, so you’re weighing sorted available options instead of vetting a feed by hand. A search returns around 250 availability-checked ideas, ranked for brandability with a scoring method we publish, at $2–$5 per search with no subscription, and credits never expire. If you want the feature-by-feature version of how that compares to naming in a chat, Domain Yoga vs ChatGPT lays it out. Either way, the principle costs nothing: let the model diverge, let a registry decide.