The names you can think of are already taken.

Most great brand names were registered ten years ago. The good ones still left aren't the names you'll think of unprompted — they're the ones you have to invent. Here's how we find them.

The names you brainstorm are mostly already taken

Every founder we've watched go through the naming process does the same thing: lists ten names they think are clever, opens a registrar tab, and watches eight of them come back registered. The two left over are usually compromises — a tortured spelling, a weird TLD, a name that doesn't quite mean what they wanted.

That's the unprompted set. The names a literate adult would think of without help. Almost all of them were registered in the dot-com era and have stayed registered since. The useful naming territory in 2026 is the names humans don't reach for first.

What an LLM is actually good at here

We probe a model with your idea and ask it to invent words for the category. LLMs turn out to be unusually good at two things humans struggle with: generating novel compounds that still sound like real brands, and matching a name to the TLD that reinforces it.

We ran the v4 prompt against a SaaS pitch — turning customer-support tickets into a searchable knowledge base for new hires. Three of the model's top picks:

Notice the model routed kbforge to .dev, where the developer audience already lives, and put tribely on .so where indie-startup brands cluster. It also knows which roots are picked over on .com — every .com prediction here is honest about the slim odds.

On predicted availability

Every TLD pairing the model proposes carries a confidence score from a calibrated availability prior. We bucket those into three labels — predicted free, uncertain, predicted taken — and never show a green check. Verification is one click away on the registrar of your choice.

Why the buckets and not the raw 0-100? Because the gap between "85% likely free" and "92% likely free" is not a difference a buyer can act on. The gap between "predicted free" and "predicted taken" is. We collapse the noise.

The hard part: deciding which one to ship

Coining gets you 30 candidates. Picking which one to commit to is a different problem — and the one our research product was built for. Click "research →" on any candidate and we run the full attractor / inference / coherence / pricing stack on that exact (name, TLD) pair, then synthesize a verdict.

Coining is the front of the funnel. The depth is in the per-domain read.

Sample pitch used in this piece: "We're building a SaaS app that turns customer support tickets into a searchable knowledge base — so new hires stop asking the same questions on day one."