Naming an AI company without dating it to this year

Every technology wave leaves a sediment layer of company names. You can date a startup to within about eighteen months from its wordmark alone. The -ly suffixes are 2012. The dropped vowels are 2009. The two-syllable Latin abstractions with a lowercase logotype are roughly 2021. None of those founders thought they were following a fashion. They thought they were picking something modern.

AI naming is in its own sediment layer right now, and it is thick. If you are naming a company this quarter, the useful question is not what sounds current. It's what will sound like nothing in particular in five years, which is the highest compliment a company name can earn.

The tells that will date you fastest

Some patterns are so dense in the current cohort that using them is effectively announcing your founding year.

• The bolted-on AI. Anything AI, AI anything. It reads as a category badge, and category badges expire. Nobody named a company Something Internet after about 2001, and the ones who did quietly stopped.
• Greek and Roman minor deities. The pantheon has been strip-mined. If your name is a lesser Titan, there are four other companies with the same name and two of them do vector databases.
• Cognition words. Mind, brain, cortex, synapse, neuro, cognition, reason, think. Fine words. Currently doing the work of a uniform.
• Agent, copilot, assistant, autonomous. These describe a product architecture that is being redrawn every six months. Naming the company after the architecture ties you to a moment in it.
• The single Latin verb stem. Cogni-, senti-, veri-, intelli-. Available in bulk, distinguishing nobody.

None of these are wrong in the sense of being bad words. They are wrong in the sense that a black turtleneck was not wrong in 2004, but you would notice one now.

Why the .com question is really a durability question

The naming shortcuts above tend to come with a second problem attached. The good .com went years ago, so the founder takes the .ai, tells themselves the extension is on-theme, and moves on.

The extension is on-theme. That's the difficulty. A .ai says AI company the way a .io said developer tool and a .co said we couldn't get the .com. It's a date stamp with a renewal fee. Plenty of serious companies run on .ai and will keep doing so; the point is that the extension is carrying meaning you may not want it carrying in 2031, when your product has drifted into something adjacent and the letters after the dot still insist on 2025.

There's a practical cost too. Every .ai company spends a small permanent budget on correcting people who typed the .com and landed somewhere else. Sometimes that somewhere else is a competitor. Sometimes it's a parked page with ads. The tax is real, it's just distributed across thousands of tiny moments you never see.

If you can get the .com at founding, get it. If you cannot afford the .com for the name you want, that is useful information about the name, not about your budget. Pick a name where you can own the front door.

What actually holds its value: names that describe nothing

The most durable technology names tend to be either arbitrary or metaphorical, and almost never descriptive.

Consider what survived earlier waves. Stripe. Figma. Notion. Slack. Vercel. Not one of them tells you what the company does, and that is precisely why none of them had to be replaced when the company changed what it does. A name that means nothing can come to mean whatever you make it mean. A name that means something specific will eventually mean something you've stopped doing.

The mechanics of this are worth being concrete about:

• Real words, unexpected context. Short English nouns that have nothing to do with software. Concrete, physical, pronounceable. Most are taken as .coms, but a good many are held by people who would sell.
• Invented but phonetically ordinary. A word that doesn't exist yet obeys English stress patterns, so nobody has to be taught how to say it. Two or three syllables. No unusual consonant clusters.
• Compound of two plain words. Harder to do well than it looks, and it dates faster, because compounds absorb the vocabulary of their era.

If you are working the invented-word route, category domain search is a faster way in than a spreadsheet, because it filters against what is genuinely available rather than what sounds nice in a meeting.

The five-year test, and how to actually run it

Ask three questions about any candidate. Not one, because a name can pass the obvious test and fail the hard one.

Does it survive the pivot?

Write down the most plausible thing your company does in four years that isn't what it does now. Model layer companies become application companies. Application companies become infrastructure. Now say the name out loud in that sentence. If it stops making sense, you've named the product, not the company.

Does it survive the technology becoming boring?

At some point, the machine learning inside your product will be as remarkable as the database inside it. Nobody markets a database. When AI is table stakes, a name that leans on AI as a differentiator is leaning on nothing. Names that referenced the internet in their branding all had to quietly stop doing so, and most of them lost something in the process.

Does it survive being said badly?

Over a bad phone line, by someone with a different accent, to a person who then has to type it. This is the test that kills clever spellings. If you have to say "with a K" or "no E at the end", you have bought yourself a permanent small tax on every introduction.

When paying up for a name is the correct call

Founders routinely spend more on three months of a mid-level engineer than they'll ever consider spending on the name the whole company lives inside for a decade. That asymmetry is hard to defend once you say it plainly.

The reasonable version of the argument: a name you own outright, in the extension people default to, with no spelling caveats and no competitor one letter away, removes a category of friction from every future hire, every press mention, every enterprise procurement conversation. It compounds quietly. It's the sort of thing you cannot retrofit at Series B without spending vastly more, and losing whatever recognition you'd built.

Buying an aftermarket name is less dramatic than founders expect. Escrow handles the sequencing, so neither party has to go first; the buyer pays no commission, and the transfer itself is a mechanical process of unlocking the domain at the losing registrar and moving it. If you'd rather deal with whoever actually holds the name, owner-direct listings cut out a layer.

One deliberate exception

If your entire strategy is to be read as an AI company immediately, by investors and by buyers who are currently allocating budget to a line item called AI, then a dated name may be the right trade. You are optimising for the next eighteen months, not the next ten years. Some companies genuinely should do that.

Just make the trade knowingly. There's a difference between choosing a name that says 2025 because the next two years are what matter, and choosing it because it was the fourth option in a document and everyone was tired. The second one is how most companies end up with names they'd change if changing were free.