What Goes Into a Good Company Name (and Its Domain)

We build and name things for a living — tools, sites, whole product lines — and we keep coming back to the same discovery: a name is not a label. It is infrastructure. You speak it on calls, type it into contracts, print it on invoices, and hand it to machines that will spell it, say it, and look it up on your behalf. A good name quietly does all of that work for decades. A bad one charges rent forever.

So we went and read the evidence properly — the psychology papers, the trademark law, the domain price data, the famous case files, and the new research on how AI assistants hear and cite names. This is what a good name actually consists of. Two updates followed in October 2026 — a second research round on why some names sound cool while others quietly sound wrong, and on domain strategy: the name and the domain as a single decision, made together. Both are folded into what follows, and into the study.

Update, October 26, 2026: the sections "Cool is a property, not a vibe" and "The name and the domain are one decision" are new, as is the pair screen at the end; the audio above was re-recorded for the new sections (version 2). Everything else stands as published on October 25.

The short version. Every candidate name has to pass four kill gates before it earns any points at all:

1. Trademark-clear. No live mark that sounds like it in the classes you'll operate in. Sound counts as much as spelling — respelling a name does not escape a conflict.
2. No bad meaning. Checked with native speakers, in the spoken form, in every language you launch in.
3. The exact-match domain is obtainable at a sane renewal price. Not a cheap first year — a sane renewal.
4. The radio test. Ten strangers hear it once, in a sentence, and at least eight write down exactly the spelling you intended.

Survivors are then scored on fluency, voice-assistant survival, ownability, memorability, and scalability — a name for a company has to fit businesses that don't exist yet. One anchor figure before we start: the same company name costs $11.08 a year to keep on a .com domain — and $2,060.25 a year on a .inc. Same name, same company, 186 times the rent, decided entirely by three letters after the dot. (Prices checked at Porkbun, October 25, 2026.) That is the kind of difference this decision quietly bakes in.

Heard once, spelled once

The best-documented effect in naming is also the simplest: names that are easy to process — easy on the tongue, easy on the ear — are trusted, remembered, and valued more. In a 2006 study (Alter and Oppenheimer, Princeton), stock tickers that were easy to pronounce outperformed hard ones right after their IPOs; a hypothetical $1,000 basket of the easy names beat the hard names by $112 to $333 depending on the window. The effect fades over months — fluency buys you the first impression, not the business — but first impressions are exactly what a name is for.

The most useful translation of that idea is the radio test: could a stranger who hears your name once — on a podcast, in a noisy call, from a friend — type it correctly into a browser? Everything that fails it shares a shape: homophones (site/sight/cite), silent letters, creative punctuation, and spellings that could go three ways. Accenture's rename in 2000–01 is the professional version of the same test: the brief required the new name to be trademarkable in 47 countries, inoffensive in over 200 languages, and available as a URL — before anyone was allowed to fall in love with a candidate.

On length, the evidence points to a sweet spot rather than a race to zero. Longer names are actually recognised more accurately (each extra letter is another cue), while very short ones win on pure recall — and three-letter domains sell for scarcity prices ($1,425 median in a study of 480,000 sales) that have nothing to do with quality. Past about three characters, the market pays the same for five letters as for twelve; what it pays for is a good word. Practical target: two or three syllables, five to ten letters, no initials. The public will shorten a long name anyway — FedEx and Dunkin' were both customer nicknames first — but nobody can lengthen a name that's too compressed to say.

The sound of it

Sound carries meaning before a single definition is learned. Give people two nonsense shapes — one round, one spiky — and ask which is "bouba" and which is "kiki", and 95–98% agree across cultures and languages. Words work the same way. Front vowels (the ee in "teeny") make things feel small, fast, light and precise; back vowels (the oo in "huge") make them feel large, heavy and substantial. Hard opening consonants — p, b, t, d, k, g — are measurably easier to recall, which is one reason so many durable names start with one.

George Eastman understood all of this in 1888, before any of it was measured. He invented Kodak under three rules that still hold up: the name must be short, impossible to mispronounce, and resemble nothing else in existence. He liked K — "a strong, incisive sort of letter" — and built the word around it. One caution from the research: these are population-level tendencies, not a horoscope. Judge a name by how whole strangers process the whole word, never by letter-by-letter symbolism.

Cool is a property, not a vibe

Ask a room why one name sounds cool and another sounds like a law firm, and you'll get shrugs — taste, apparently. Taste turns out to be measurable. The psychology of cool has been studied properly for about a decade, and its findings explain most of what founders feel in their gut and then can't defend in a meeting.

Start with the opposite of cool: the name people quietly recoil from. The best experiment here (psychologist Paul Thibodeau, 2016) found that word aversion lives in meaning, not sound: people who disliked a word were really reacting to what it made them picture — the bodily, domestic or faintly disgusting company it keeps — and sound-alike words with innocent meanings bothered nobody. What this means for naming is unfashionable and useful: a name rarely fails because of how it sounds. It fails because of what it resembles. English even has clusters that drag their meanings with them — the sl- family (sly, slime, sloppy) leans pejorative, the sn- family is nose-and-contempt country (snore, snot, snide), while gl- is light and shine (gleam, glisten, glow). Nobody taught you those associations; you absorbed them from the vocabulary itself. A name built on a dirty cluster starts the race carrying weight it didn't choose.

The positive side has a profile too. Asked across decades which words they find beautiful, English speakers keep choosing the same shape: three syllables, stress on the first, built from soft liquids and nasals — the l, m, n and r sounds. That profile reads organic and premium. It does not read high-tech — which is the first hint that "sounds good" and "sounds cool" are different judgements wearing similar clothes.

Cool itself, measured rather than admired, is bounded autonomy: a deviation from a norm the audience considers illegitimate, kept inside the limits the audience can absorb. Too little deviation and a name reads corporate-follower; too much and it reads deviant — errors and oddities register as mistakes, not rebellion. The admired technology names fit the formula almost suspiciously well: one or two syllables, stress on the first, utterly fluent to say — and exactly one marked feature. A single rare letter (the K, X, Z and V family reads edgy across every study we found). One odd-but-sayable cluster. One unexpected meaning on an ordinary sound. What they never carry is two novelties at once — novelty stacked on novelty is the sound of trying too hard, and there is no reading cool survives less.

The era evidence should humble everyone. The -ify wave, the dropped vowels, the doubled letters: each began life as a practical workaround (the domain was taken; the trademark was crowded) and each curdled, within a decade, into a date stamp — the naming equivalent of last decade's slang. Deviations get adopted precisely because they signal independence; once the crowd copies them, the same research shows early adopters abandon the signal as follower behaviour. A name chosen to sound like this year's cool will sound like this year's year, forever. The names that stay cool are the ones whose single deviation was chosen for a reason that doesn't expire.

Misspelled on purpose

For a decade, startup naming meant dropping a vowel or swapping a letter — Flickr, Tumblr, Lyft. The research now says that trick mostly charges you instead of paying you. A 2023 study in the Journal of Marketing (Costello, Walker and Reczek) tested slightly misspelled brand names against correct ones across 29 experiments: shoppers chose the correctly spelled version 62% of the time versus 48% for the misspelled one. A one-letter deviation cost fourteen points of preference — because people read an unexplained error as carelessness, and unreliability is the one message a new company cannot afford.

Misspellings also break the radio test in a way no radio spot can fix: "Klear — with a K" is a tax you pay out loud, forever. And they don't even buy you the legal escape people assume: trademark confusion is judged on sound and overall impression, so "Klear" does not escape "Clear" if the products are related. The honest exceptions exist — a misspelling can work when it's one small step from the real word and there's an obvious reason for it (Lyft's y carries the "lift" idea) — but the default answer is: spell it the way it sounds, or pick a different word.

Ownable beats explainable

Trademark law runs on a spectrum, and where a name sits on it decides what you actually own:

• Generic (escalator, aspirin — both once trademarks): names the thing itself. Never protectable.
• Descriptive (a name that states what you sell): not registrable until years of use prove the public sees it as you.
• Suggestive (Netflix, Coppertone): hints, doesn't state. Registrable, ownable — the sweet spot for products.
• Arbitrary (Apple, for computers): a real word used for something unrelated. Strong.
• Coined (Kodak, Exxon, Pepsi): invented. Strongest of all — protectable from day one, in every market.

The trap is the second rung. A descriptive name feels like free marketing — it explains you! — but you can never stop competitors from using the same words descriptively, and it boxes you into the product it describes. Comprehension can be bought later with marketing. Ownability cannot be bought at any price once someone else holds it.

Two stories make it concrete. A small-town Burger King in Mattoon, Illinois, had used the name first, locally — and when the national chain arrived, the courts carved out a roughly twenty-mile zone the giant still cannot enter. Rights follow use, not registration, in the US system. And Google's parent, Alphabet, has coexisted for a decade with BMW's fleet company Alphabet and its alphabet.com — because trademark is about confusion in a field, and fleet management and internet services are different worlds. A good name isn't necessarily an empty word; it's a word that's empty in your field.

The domain is rent, not a purchase

A domain feels like something you buy once. It isn't. It's a lease that renews every year, forever — so the only price that matters is the renewal, and the gap between the first year and the renewal is where the industry hides its margins. Live prices from Porkbun (checked October 25, 2026, non-premium names):

Two different "premium" traps hide in this table. An aftermarket premium — buying a taken name from its owner — is a one-time pain: it renews at the standard rate afterwards. A registry premium is a name the registry itself withheld, and it often renews at the premium price every year. A $2,000-a-year name is a $20,000 decade, and once your email and brand live on it, you can't realistically walk away. Always read the renewal line at checkout, never the banner.

Why insist on .com for anything the general public touches? Not snobbery — leakage. People who remember a name but not its ending guess .com (consumers are measurably several times more likely to assume it), and every guess that lands on a stranger's site is a lost customer. Worse is email: someone will eventually send a contract, an invoice, or something confidential to [email protected] by habit. If a stranger owns that mailbox, that's not a lost lead, it's a breach. The tech endings have their own footnotes, too: .io is a British territory's country code whose long-term future follows a sovereignty treaty, and .ai is Anguilla's — a small island nation now earns a quarter of its government revenue from .ai renewals, which is exactly the kind of landlord you don't want setting your rent in year ten. Products aimed squarely at developers can live happily on .io or .ai. Anything that has to face the general public belongs on .com or a home country domain.

Hold the whole portfolio in view while you decide: the main name, the .com twin, a couple of obvious misspellings held defensively. At ordinary prices that's $100–$250 a year all-in — noise next to payroll. The expense you're actually insuring against is the rename. But notice what a portfolio presupposes: that the name and its domains were chosen as one thing, on the same day, on the same evidence. The next section is about that decision.

The name and the domain are one decision

Everything above treats the name and its domain as separate purchases. In practice they are one asset, and judging them separately is how good names end up on bad domains. There are only a handful of shapes the pair can take. You can hold the bare .com — the asset every other option is measured against, because .com is the ending the public defaults to: vendor tests (labelled as vendor data in the study) put the default at several times the traffic of the alternatives, and every misdirected email to yourname.com is a cost no dashboard shows. You can take a descriptive ending under your own .com's protection, or a country ending as a deliberate brand play (a short ccTLD reads as a designed object, not a compromise). You can wear a modifier — get-, try-, -hq, -app — which reads exactly like what it is: scaffolding, honourable on a product, temporary on a company. Or you can decouple entirely: Meta's legal name, brand and domain are three different objects, as are Alphabet's — a parent name can be chosen for the register and the trademark file while the domain is bought later, separately, at whatever the asset costs.

The sharpest illustration is an email purchase. In 2011 Facebook paid $8.5 million for fb.com — not for the public, who already typed facebook.com, but for its own staff: the short domain was bought primarily so that employee email could move onto it cleanly. Email is where domain strategy stops being marketing. Your domain is read, aloud and in inboxes, by bankers, lawyers and customers deciding whether the message in front of them is real. firstname@yourname, read among the Stripes and Googles in an inbox list, is a credibility test your name takes a hundred times a day. A pair that needs its ending spelled out on a call — "that's name, dot, systems" — is paying a tax on every introduction, forever.

Startups vote on this with real money. The upgrade pattern is so consistent it might as well be a law: launch on the assembled pair (the get- domain, the modifier, the alternative ending), then buy the bare .com the moment the company can afford it — Dropbox out of getdropbox.com in 2009, Facebook out of thefacebook.com before it, Notion and Mercury after. Nobody upgrades away from the bare .com. Which means the assembled pair was never a saving; it was a loan, repaid later at a price set by a seller who can see your funding announcements.

So the working rule of this whole piece, restated as one line: judge pairs, not names. Generate candidates already attached to their specific domain and its renewal price. Screen each pair once, whole. Secure the winner — domain, defensive twin, handles, package names — in a single sweep, the same day, before the announcement and before the attachment sets in. A name you love whose pair fails is not a near-miss; it is a different, worse asset wearing a good name. And the framework at the end of this piece now screens pairs exactly that way.

What the case files say

The winners share a pattern, and it isn't creativity. Accenture wrote its constraints first. Verizon picked from 8,500 candidates. Stripe began life as "/dev/payments" — a programmer joke nobody could pronounce or spell — and renamed itself, buying stripe.com cheap in 2010, before it mattered. Google is a typo of "googol" that happened to satisfy every criterion: two syllables, one pronunciation, nothing else like it, and a .com that was free precisely because the word was new.

The failures are equally consistent. Kraft's Mondelez was chosen from 1,700 employee entries and vetted in 28 languages — and still launched to headlines reading "MONDEWHAAAAT?", with a sound Russian speakers heard as a vulgar term. Vetting without veto power is theatre. Mitsubishi's Pajero meant something rude in Spanish slang, so the same SUV has been sold under three different names (Montero, Shogun, Pajero) for decades, its reputation never compounding into one brand. And the names that describe the product eventually amputate it: Dunkin' Donuts became Dunkin' when drinks outsold donuts; Weight Watchers shrank to the unsayable "WW".

The price tags: Verizon reportedly spent over $300 million launching its coined name; Accenture's rebrand carried a reported $175 million ad blitz; Tropicana's 2009 redesign — not even a rename, just new packaging — lost 20% of sales in two months, roughly $30 million, before it was withdrawn. Set against that, doing this properly at the start costs a few thousand in searches and legal review, plus the domain. Naming is the cheapest insurance a company ever buys — and it's only cheap once, at the beginning.

The machines are listeners now

Two things changed in the 2020s. First, names are increasingly spoken to software. Voice recognition resolves unclear audio toward common words, so an unusual name gets "corrected" into a famous one — ask for a table at Noma and you may book Roma. Measured on real names, the error rates are startling: one 2026 evaluation of fifteen speech models found an average 39% error rate on street names, and speakers whose first language isn't English fared far worse. So add a fifth test to the radio test: the dictation test. Say the name, in a sentence, to three different dictation systems, five times each. If they don't type back exactly your spelling at least nine times out of ten, the name fails — no matter how it looks on a slide.

Second, AI assistants are becoming the front door. Studies of hundreds of thousands of AI answers in 2026 found the major assistants disagree on which brands to mention in about 27% of identical questions, and brand recommendation lists churn week to week. The name-level defence is boring and effective: be unique (a model forced to disambiguate between you and a famous namesake will pick the famous one), be consistent (one spelling everywhere — domain, social handles, code registries — because systems stitch your identity together from exact strings), and keep a canonical home page on your exact-match domain that says plainly what the company is.

The handles deserve a line of their own, because they bind tighter than domains do. To be identical everywhere, a name has to fit the strictest platforms: 15 characters or fewer, letters only if you can manage it. And for anything that ships software, check the unglamorous namespaces too — GitHub, npm, PyPI. Researchers found that AI code models invent plausible-sounding package names that don't exist about 20% of the time, and attackers pre-register those names to catch the installs. A brand that isn't defensively registered in the package registries is a supply-chain incident waiting for a victim.

The checklist

For anyone naming something that has to last, the whole study compresses to this:

Kill gates — fail any one, discard the name:

1. Trademark first pass clear in your classes (sound-alikes included).
2. No bad meaning, spoken aloud, in launch languages.
3. Exact-match domain obtainable, renewal price sane.
4. Radio test: 8 of 10 strangers spell it right after hearing it once.

Then score what's left on: fluency and one agreed pronunciation · dictation survival (9 of 10 exact) · ownability (suggestive or stronger on the trademark ladder; can you own the first page of search results for the bare name within a year?) · memorability after 24 hours · sound that fits the posture (fast and light, or large and solid) · scalability — no product word, technology word, or fashion in a company name · portfolio cost at renewal prices · the right ending for the audience · identical handles and package names, claimed the same day you decide.

Then run the pair screen — ten quick judgements on the name-and-domain together, detailed in the study: the keynote sentence said cold on a stage; a trade-press headline, including a hostile one; the job-title test ("VP of Engineering, X"); the radio test applied to the pair, email included; the inbox-dignity test; the conference badge read across a table; the cross-generational reading (a CFO and a junior engineer each say what the company probably does); the hoodie test, folklore rather than evidence and useful only as an embarrassment alarm; the namespace sweep; and the ten-year test — the name read as a 2036 incumbent, with no era template left in it. Survivors of all of that are rare, which is rather the point.

And the traps, in the order they actually bite: falling in love before clearance · believing a domain gives you trademark rights (or the reverse) · going descriptive for the quick clarity hit · respelling your way out of a conflict · pricing by the first year · building on a non-.com while a stranger holds the .com · foreign-language checks nobody could veto with · naming the current product, technology, or trend · forgetting the handles and registries until launch day. Two more, added with the October 2026 update: deciding the name and the domain in separate steps — the assembled pair is the price of falling in love with a bare name — and stacking signals of cool until the name reads as trying too hard, the one reading cool cannot survive.

Caveats, honestly

Some of this evidence is strong — the fluency studies are peer-reviewed and replicated, the trademark framework is statute and case law, the price data is arithmetic. Some is softer: the TLD-trust percentages come from industry surveys, not journals, and the claims about how AI assistants treat names are inference from how those systems demonstrably work, not controlled experiments on names themselves. Where the study relies on either, it says so. Prices are point-in-time — checked at Porkbun on October 25, 2026 — and renewal terms change; re-read the live pricing page before any purchase. And two claim types in the new sections deserve their labels in the open: the traffic and trust figures in the domain-strategy section are vendor measurements, and the hoodie test is practitioner folklore — we say so where they appear, and the study grades both accordingly.

The full study — every figure above with its source, the trademark spectrum and DuPont factors in detail, the domain sales datasets, the case files, and the complete scored framework with its procedure — is published alongside this article: What Goes Into a Good Company Name (and Its Domain) — the study.

Sources

• Alter & Oppenheimer (2006), "Testing the association between the fluency of a name and its market performance," PNAS — and Green & Jame (2013), "Company name fluency, investor recognition, and firm value," Journal of Financial Economics.
• Costello, Walker & Reczek (2023), on unconventional brand-name spellings, Journal of Marketing.
• USPTO, "Strong trademarks" guidance; the Abercrombie spectrum; In re E. I. du Pont de Nemours (1973).
• Verisign, Domain Name Industry Brief, Q4 2025 (386.9M registrations; .com 161.0M).
• Bishopi (2026), analysis of 480,667 .com sales; Lindenthal (2014), "Valuable words," JASIST.
• Porkbun domain pricing page — read live October 25, 2026 (first-year and renewal, incl. fees).
• Case files: Accenture (2001 rename brief), Verizon (2000), Stripe (2010–11), Alphabet (2015), Meta (2021), Block (2021), Mondelez (2012), Mitsubishi Pajero, Tropicana (2009).
• AI-era: Together AI (2026), speech-model evaluation on proper names; LLM Scout and Envoyra (2026), brand-mention studies; Spracklen et al. (USENIX Security 2025), package hallucinations ("slopsquatting").
• The social layer: Thibodeau (2016) on word aversion as semantics; Crystal (1995) on the beautiful-word profile; Klink (2000), Yorkston & Menon (2004) and Lowrey & Shrum (2007) on sound symbolism; Warren & Campbell (2014) and Warren, Batra, Loureiro & Bagozzi (2019) on cool as bounded autonomy.
• Domain strategy: Facebook's 2011 fb.com purchase (reported at the time by TechCrunch and Reuters); the Dropbox, Notion and Mercury upgrade histories; modifier and ending readings graded as vendor data and practitioner synthesis in the study.

The full source list, with links, is in the companion study.