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

Question: What goes into a good site name? For a technology company that will sit as a parent over multiple brands: what does the evidence — from linguistics and psychology, trademark law, domain-market economics, the historical case record, and the new AI-era constraints — actually say about choosing a company name and its domain name, and how should a founder evaluate candidates before committing?

Researched: 25 October 2026; updated 26 October 2026 with a second research round — the social layer (§6: why names sound cool, and why some sound wrong) and domain strategy (§7: the name+domain pair as the asset) — and the pair screen they add to the framework (§8.4). Sources are cited inline throughout; the full source list is at the end. Companion piece: our blog article, What Goes Into a Good Company Name (and Its Domain), which carries the same findings for everyday readers, with an audio reading.

How to read the evidence labels. Where the evidence is peer-reviewed or otherwise independently established, it is labelled established evidence. Where it rests on practitioner, survey or vendor material, it is labelled expert opinion / vendor data, with the source named. The two are not interchangeable, and several famous naming "facts" turn out, on inspection, to be legends — those are flagged explicitly where they appear. Prices marked "verified live" were read on the registrar's own pricing page (Porkbun, https://porkbun.com/products/domains) on 25 October 2026; other pages were retrieved on 25 October 2026. Trademark discussion states US federal law (the Lanham Act) as the worked example; the principles generalise in form to the UK/EU and most Madrid Protocol jurisdictions, though tests and class practice differ. This study is research, not legal advice — §8 builds a professional clearance step into the procedure for exactly that reason.

Summary

A good site name is not a creative prize; it is a piece of infrastructure that has to pass six independent filters at once, and failure at any one of them is expensive in a different currency:

1. Cognitive — it must be heard once, pronounced one way, spelled one way, and remembered a day later. Processing-fluency research (Alter & Oppenheimer 2006; Green & Jame 2013) shows even investors reward fluent names — but the effect is small, short-lived, and easy to over-claim. Sound symbolism (bouba/kiki; front vs back vowels; word-initial plosives) is real but modest: it tunes perception, it does not create a brand.
2. Legal — it must sit high on the trademark distinctiveness spectrum (fanciful / arbitrary / suggestive) and survive a likelihood-of-confusion analysis against existing marks in related goods and services. A domain being available proves nothing about trademark rights; a famous-sounding word can be unusable in your class while a nonsense word is bulletproof — and the same word can lawfully belong to three different companies in three different industries (Alphabet, BMW's fleet business, is the canonical modern example).
3. Economic — the domain is a perpetual liability, not a purchase. What matters is the renewal price, every year, forever: on 25 October 2026 a non-premium .com renewed at $11.08/yr at Porkbun, a .co at $31.20/yr, a .io at $51.80/yr, a .ai at ~$82.70/yr, and a .inc at $2,060.25/yr — a 186× spread for functionally identical DNS. First-year prices are marketing: .xyz was $2.04 in, $14.21 to renew; .info $3.60 in, $22.14 to renew. And the asset being bought is never the name alone: it is the pair — name plus exact-match domain — designed together or assembled by accident (§7).
4. Historical — the case record separates cleanly. Winners (Kodak, Google, Accenture, Verizon, Stripe, Alphabet, Block, Dunkin') share a pattern: short, single-pronunciation, product-agnostic, legally cleared before launch, with the exact-match domain and handles secured as part of the same decision. Failures (Mondelez's pronunciation and Russian-slang problem, Mitsubishi Pajero's three-name fragmentation, Tropicana's recognition collapse) share the opposite: a name chosen for internal logic that external humans could not say, spell, find, or repeat.
5. AI-era — the name is now spoken to machines at least as often as it is typed. Speech-recognition systems systematically replace rare or unfamiliar names with frequent words; AI recommenders disagree with each other (27.2% of cases in one 2026 study) and churn weekly (79% of brand lists changed within four weeks in another), so an unambiguous, entity-distinct name is now a discoverability asset. And for a tools company the name must also be clean on GitHub, npm/PyPI and app stores, where look-alike names are an active attack surface (typosquatting, and the newer "slopsquatting" on model-hallucinated package names).
6. Social — cool is a measurable social reading, not a vibe: bounded autonomy — deviation from an illegitimate norm, held inside the audience's tolerance (Warren & Campbell). Aversion to a name is usually aversion to what it means and resembles, not how it sounds (Thibodeau). Sound symbolism buys category congruence, never quality; era templates (-ify, dropped vowels) date a name the way slang dates a speaker (§6).

The one-sentence distillation for a multi-brand parent: choose a short (1–3 syllables, typically 4–10 letters), coined or arbitrary, single-pronunciation name that describes no product, clears a trademark first pass in software/services classes in your launch markets, is available as an exact-match domain at a renewal price you would happily pay for a decade (the $20–$50/yr band is entirely workable; $2,000/yr prestige endings are not a branding decision, they are a tax), and is available identically as a handle, GitHub organisation, and package scope. Then test it by phone, by dictation, and in three languages before you fall in love with it — and every rule in it applies to the name+domain pair, never to the bare name (§7.4).

1. Naming linguistics & psychology

1.1 Fluency: the core finding, its size, and its limits

Established evidence. Processing fluency — the subjective ease with which something is processed — reliably biases judgement: fluent things are judged more familiar, more likeable, truer, and safer. The naming-specific demonstrations:

• Alter & Oppenheimer (2006), "Predicting Short-Term Stock Price Fluctuations by Using Processing Fluency," PNAS 103(24): 9369–9372. Three studies. In the laboratory study, participants expected fluently named fabricated stocks to appreciate +3.90% over a year and disfluently named ones to depreciate −3.86% (p < 0.0001). In Study 2 (89 NYSE IPOs, 1990–2004), name complexity negatively predicted performance after 1 day (β = −0.23, p < 0.05) and 1 week (β = −0.21, p < 0.05) — but not at 6 months (p = 0.64) or 1 year (p = 0.47). The headline basket comparison: $1,000 split across the 10 most fluent vs 10 most disfluent names differed by $112 after 1 day, $118 after 1 week, $277 after 6 months, and $333 after 1 year (the last only p = 0.10). Study 3 repeated the effect using pronounceability of three-letter ticker codes alone, which strips out semantic content. (Full text: http://pages.stern.nyu.edu/~aalter/PNAS.pdf ; Princeton summary: https://www.innovations-report.com/education/studies-and-analyses/report-60425/)
• Green & Jame (2013), "Company name fluency, investor recognition, and firm value," Journal of Financial Economics 109: 813–834. The more durable follow-up: companies with short, easy-to-pronounce names have higher breadth of ownership, greater share turnover, lower transaction price impacts, and higher valuation ratios (Tobin's Q, market-to-book); fluency-improving name changes are associated with increases in ownership breadth, liquidity and value; fluent closed-end funds trade at smaller discounts and fluent mutual funds attract greater flows. The effect is stronger among small firms — exactly where recognition is scarce. (https://ideas.repec.org/a/eee/jfinec/v109y2013i3p813-834.html ; https://papers.ssrn.com/sol3/papers.cfm?abstract_id=1777256) Harvard Business Review's summary quantifies it: shortening a company name by one word would have added $3.75 million in market value to the median-size firm in the sample, across 2,630 corporate name changes from 1980–2008 (https://hbr.org/2014/01/size-matters).
• Song & Schwarz (2009), "If It's Difficult to Pronounce, It Must Be Risky," Psychological Science 20(2). Hard-to-pronounce (ostensible) food additives were rated more harmful, and hard-to-pronounce amusement rides both more exciting and more sickness-inducing — low fluency reads as novelty, and novelty reads as risk in whichever direction risk is salient. Mediation ran through perceived unfamiliarity. (https://www.psychologicalscience.org/article?ru=https%3A%2F%2Fjournals.sagepub.com%2Fdoi%2F10.1111%2Fj.1467-9280.2009.02267.x)
• The effect generalises beyond companies: Newman et al. (2014, PLoS ONE) found people with easier-to-pronounce names are evaluated more positively and — remarkably — claims attributed to them are judged truer ("truthiness"). (https://journals.plos.org/plosone/article?id=10.1371%2Fjournal.pone.0088671)

Critiques and boundary conditions — do not skip these.

• Alter & Oppenheimer's own data bound the claim: for company names the significant window was days-to-a-week; the longer-horizon basket differences rest on a 10-vs-10 extreme-groups comparison, and the 1-year difference was marginal. This is evidence that fluency buys an easier first hearing, not that it compounds.
• The fluency→risk result has a documented stimulus-sampling problem: a direct-replication literature ("If it's difficult to pronounce, it might not be risky: the effect of fluency on judgment of risk does not generalize to new stimuli") argues the original result may be an artefact of one unrepresentative stimulus set — a general warning against treating any single fluency experiment as a law. (https://www.researchgate.net/publication/314198171_If_It_s_Difficult_to_Pronounce_It_Might_Not_Be_Risky_The_Effect_of_Fluency_on_Judgment_of_Risk_Does_Not_Generalize_to_New_Stimuli)
• A 2025/26 analysis of the modern S&P 500 (MDPI Symmetry 18(3):477) frames the open tension explicitly: in markets dominated by high-frequency trading, ticker symbols are processed as numeric identifiers by machines, arbitrageurs exploit published anomalies, and the human-fluency window may have closed for large liquid stocks — while remaining plausible for human first-contact decisions (customers, hires, partners), which is the decision a brand name actually faces. (https://www.mdpi.com/2073-8994/18/3/477)

Practical reading: fluency is a tie-breaker and a risk-reducer, well-supported at the moment of first exposure. It is not a growth engine, and no study licenses paying a large premium for a marginally more fluent name over a marginally less fluent one — but several studies do license rejecting a name that strangers stumble over.

1.2 Length and memorability — shorter is usually better, except when it isn't

Established / mixed evidence.

• Pathak et al. (2019), summarised by Springer Nature's research community: participants respond faster to shorter brand names; 92 of the 100 most frequent words in the Corpus of American English are monosyllabic, and short words have denser phonological neighbourhoods — familiarity is built into the language itself. The same summary warns, from de Lencastre et al., that initialisms are the exception: letter-strings are processed differently from words, are hard to pronounce and memorise, and "can lead to brands being ignored altogether". (http://communities.springernature.com/posts/the-psychology-of-names-can-brands-appeal-to-our-subconscious-senses?badge_id=multimedia-tools-and-applications)
• Counter-evidence that must be in any honest study: Zee, Willems, Brengman & Desmet's experimental work on auditorily presented non-word brand names (98 participants, 36 fictional brands, 3,528 recognition measurements) found longer names were better recognised than shorter ones — more phonemic material gives memory more to grip and more distinctiveness from competitors — and an inverted-U for attitude by phoneme count: names shorter than about three phonemes were liked less. A 2026 policy brief from the same research line (SMIT Policy Brief #94, "Rightsizing Brand Names") states the challenge plainly: the popular "two-syllable rule" has a surprisingly thin empirical basis, and syllable count is probably the wrong unit — phonemes are better. (https://www.researchgate.net/publication/257474699_Does_brand_spelling_influence_memory_The_case_of_auditorily_presented_brand_names ; https://prod-d2bcaa46b289670c.paddlecms.net/sites/default/files/2026-06/SMIT_Policy-Brief94-Rightsizing-Brand-Names.pdf)

Reconciliation (this study's position): the findings split by task. Recall from scratch and speed of processing favour short; recognition among confusable alternatives favours enough length to be distinctive. Both are satisfied in the same place: long enough to be unmistakable, short enough to say in one breath — in practice 2–3 syllables / roughly 5–10 letters for a coined parent name, and not an initialism. Names above three syllables predictably get shortened by the public anyway, on the public's terms, not yours: Federal Express → FedEx, Kentucky Fried Chicken → KFC, International House of Pancakes → IHOP. (https://frozenlemons.com/blog/how-long-should-business-name-be)

1.3 The "radio test" — heard once, spelled correctly

Expert opinion, but unusually convergent. The radio test (also "phone test"): say the name once, in a sentence, to someone who has never seen it written; if they cannot type it into a browser correctly, the name leaks traffic, email, and word-of-mouth forever. It is the operational form of the fluency literature plus the voice-assistant problem (§5.1). Naming practitioners apply it as a pass/fail gate, not a score — a name either survives being heard or it doesn't. (https://inkbotdesign.com/brand-naming/)

Three corollaries that follow directly, and that recur in the case studies:

• One spelling per sound. If a sound can be spelled two common ways (Claire/Clare, site/sight/cite, -er/-or endings), hearers will split. Homophones are the single most common radio-test failure.
• No silent letters, no diacritics, no punctuation in the spoken form. Stripe's founders learned this in reverse (§4): their first name was literally unpronounceable as written — "/dev/payments".
• Say it in a noisy sentence, not in isolation. "We use ___ at work", "Email me at ana@___.com", "Look up ___" — names that pass cold often fail embedded, because adjacent words create false word boundaries (the classic handle-era example is the "Who Represents / Whore Presents" segmentation trap, cited in practitioner handle guidance: https://www.meticulosity.com/blog/handle-with-flair-crafting-creative-and-catchy-social-media-handles-that-stand-out).

1.4 Phonetics and sound symbolism

Established evidence (small-to-moderate, replicable effects on perception, not on sales).

• Bouba/kiki. Köhler (1929) showed people match takete to a jagged shape and maluma/baluba to a rounded one. Ramachandran & Hubbard (2001) revived it as kiki/bouba: 95–98% of both American undergraduates and Tamil speakers in India matched bouba to the rounded shape and kiki to the angular one; later reviews put the cross-cultural range at roughly 80–95%. Sound–shape mapping is therefore not arbitrary and not English-specific. (https://en.wikipedia.org/wiki/Bouba/kiki_effect ; https://www.iflscience.com/the-bouba-kiki-effect-shows-a-very-strange-facet-of-language-and-human-psychology-76894)
• Front vs back vowels carry size/speed/weight. Sapir (1929): over 80% of participants judged mal (back vowel) the larger table and mil (front vowel) the smaller. Lowrey & Shrum (2007), "Phonetic Symbolism and Brand Name Preference," Journal of Consumer Research: fictitious names differing only in one vowel — front-vowel nillen/gimmel vs back-vowel nallen/gommel — were preferred for a two-seater convertible and a knife when they had front vowels, and for an SUV and a hammer when they had back vowels. The mechanism is fit: front vowels connote small, fast, light, sharp; back vowels connote large, slow, heavy, solid. Shrum et al. (2012) extended the effect across languages. (https://phys.org/news/2007-09-vowel-affect-consumer.html ; https://daily.jstor.org/whats-brand-name-sounds-persuasion/ ; https://www.sciencedirect.com/science/article/abs/pii/S0167811613000475)
• Plosives aid memory. Plosives/stops — p, b, t, d, k, g — especially in word-initial position, increase brand-name memory, recognition and recall (Lowrey, Shrum & Dubitsky 2003; Cortese 1998), and word-initial plosives are disproportionately common among real top brand names (Vanden Bergh et al. 1987; Pogacar et al. 2015, "Sounds good: Phonetic sound patterns in top brand names," Marketing Letters). The same literature is honest that plosive-initial words also appear more in words rated "unpleasant" — memorability and warmth are different dials. Fricatives (s, z, f, v) are genuinely contested: associated with smaller/faster/lighter in Klink's work, negatively associated with electability in Smith (1998), and under-represented in brand names relative to dictionary frequency in Schloss (1981). (https://www.researchgate.net/publication/285416618_Sounds_good_Phonetic_sound_patterns_in_top_brand_names ; https://static1.squarespace.com/static/586d3b8c579fb3b92bfadb89/t/5b0ee09ef950b77426273ea8/1527701664023/Pogacar+et+al.pdf)
• Why Kodak works, in the founder's own criteria. George Eastman: "The letter 'K' had been a favorite with me — it seems a strong, incisive sort of letter. It became a question of trying out a great number of combinations of letters that made words starting and ending with 'K.'" Working with his mother on an anagram set, his three stated requirements were that the name be (1) short, (2) easy to pronounce, and (3) incapable of resembling or being associated with anything else. He registered Kodak on 4 September 1888. That is, in 1888 language: a fanciful mark, plosive-bookended, two syllables, zero prior meaning — the template Exxon (coined for Standard Oil's 1972 rename, chosen in part because it existed in no language and its double-X was distinctive) later followed. (https://www.mondaq.com/unitedstates/trademark/253212/the-best-method-to-the-madness-behind-choosing-a-company-name-or-mark ; https://www.justia.com/intellectual-property/trademarks/strength-of-marks/fanciful-marks/)

Expert-opinion caution. Naming agencies routinely over-extend this literature into per-letter horoscopes ("S means speed, M means trust"). The evidence supports only coarse claims: vowel frontness shifts perceived size/speed; initial plosives help recall; matched sound–product fit shifts preference in forced-choice experiments. Effect sizes are those of laboratory preference tasks. Treat phonetics as a way to choose between two otherwise equal finalists, and to avoid a mismatch (a soft, fricative, back-vowel name for a product selling speed and precision), never as a generation algorithm.

1.5 The distinctiveness spectrum in language (the marketing side)

Practitioners and courts use the same ladder (§2.1 gives the legal version); linguistically it is a trade-off between immediate comprehension and ownability:

The asymmetry that decides parent-company naming: comprehension can be bought with marketing; ownability cannot. A descriptive name explains the first product and then taxes every subsequent one; a coined name explains nothing on day one and nothing needs explaining away on day 1,000. Justia's standard statement of the fanciful trade-off is exact: strongest protection, but "it can take a long time for the public to associate the mark" with anything — the cost is education, paid in advertising. (https://www.justia.com/intellectual-property/trademarks/strength-of-marks/fanciful-marks/)

1.6 Variant spellings — Flickr, Lyft, Tumblr: the evidence is now against the fashion

This is the one area where the last five years of peer-reviewed work directly contradicts 2010s startup folklore.

• Costello, Walker & Reczek, "'Choozing' the Best Spelling: Consumer Response to Unconventionally Spelled Brand Names," Journal of Marketing (2023). Eight studies, nearly 3,000 participants. Consumers consistently dis-preferred unconventionally spelled versions of real words — not fully coined names; Lyft-for-Lift is the paradigm case. Mechanism, demonstrated by mediation: consumers read the misspelling as a deliberate persuasion tactic / marketing gimmick → infer the brand is less sincere → reduce support. In the field study outside a college football stadium, offered a free seltzer, people chose "Clear" over the alternative 62% of the time, but "Klear" only 48% of the time — a 14-point penalty for one letter in a real, unincentivised choice. The penalty disappears only when the spelling is perceived as sincerely motivated (e.g. it is integral to the brand's origin story). (https://www.fastcompany.com/90863710/brand-names-misspelled-words-consumer-study ; https://www.newswise.com/articles/unconventional-spellings-are-a-badd-choyce-for-brand-names ; https://www.eurekalert.org/news-releases/981993)
• Abell & Smith (follow-up, reported 2025) soften but do not reverse this: in six experiments, mild deviations (including simply compounding two correctly spelled words, e.g. SoftSoap) were rated as positively as correct spellings; negativity grew monotonically with distance from the correct spelling (Eazy Clean > Eazy Klean > Eezy Kleen); and relevance rescued a misspelling — Bloo for an oolong tea (a motivated pun on "blue" + "oolong") performed as well as correct spelling, while the unmotivated Blewe performed worse. (https://www.asianspectator.com/index.php/world/64909-misspelled-names-may-give-brands-a-lyft--if-the-spelling-isnt-too-weird)
• Why Flickr/Tumblr/Lyft did it anyway — and why the reason has expired. The dropped-vowel wave (Flickr dropped the e in 2004-era domain scarcity; Tumblr in 2007; Lyft substituted y for i) bought two things: an available .com-able string and trademark distinctiveness. Both are weaker arguments in the 2020s: (a) the consumer penalty above is now measured, not hypothetical; (b) a misspelling fails the radio test by construction — every spoken mention requires a spelling lesson ("Lyft with a Y", "Flickr without the E"), which is a permanent tax on word-of-mouth, voice assistants (§5.1) and email; (c) survivorship bias: for each Flickr there is a long tail of forgotten -rr, -ly and vowel-less names; the survivors succeeded despite the spelling once product and timing carried them, and several (Lyft especially) spent heavily to teach the spelling.

This study's position: for a parent brand intended to last decades and be spoken by non-customers (investors, regulators, recruits, journalists), a variant spelling is a negative-expected-value choice unless the misspelling is (i) one-step, (ii) semantically motivated, and (iii) the correctly spelled form is legally or domain-wise impossible — and even then, a coined name usually dominates it on every criterion in §8.

1.7 How names age and scale beyond one product

A parent name's job is to mean less over time, not more. Three mechanisms:

• Semantic bleaching is an asset at parent level. "Google" (a misspelling of a number), "Stripe" (a stripe), "Block" (a shape/neighbourhood) impose no ceiling. Product-descriptive names impose a hard one: Dunkin' Donuts had to amputate half its name in January 2019 because beverages were already 60% of sales and "Donuts" was "pigeonholing" the business; Kentucky Fried Chicken became KFC in 1991 to escape "fried"; Apple Computer became Apple Inc. in 2007 when phones eclipsed computers. (https://knowledge.wharton.upenn.edu/podcast/knowledge-at-wharton-podcast/five-reasons-to-rename-your-company/ ; https://www.adweek.com/brand-marketing/dunkin-drops-the-donuts-as-the-latest-brand-to-shorten-its-name/)
• Architecture consequence. A neutral parent (Alphabet over Google; Meta over Facebook/Instagram/WhatsApp; Block over Square/Cash App/Tidal) lets sub-brands be descriptive, fashionable, or disposable without renaming the listed/holding entity. The parent name should therefore be judged by a harsher standard than any product name: it must still be sayable, spellable and legally clean when the company does something currently unplanned.
• Genericide is the far-end failure of success — see §2.4. A name can also age by fashion: the 2000-era crop of pseudo-Latin -ia/-ent/-is consultancy coinages (Accenture, Altria, Mondelez) and the 2010-era dropped-vowel crop both now date-stamp their companies to a naming fashion cycle of roughly a decade. Names built from ordinary phonotactics and no fashionable orthographic trick age slowest — Kodak (1888), Sony, Nike are still orthographically unremarkable.

2. Trademark & legal reality

2.1 The Abercrombie spectrum — registrability is a property of the name–product pair

The taxonomy comes from Abercrombie & Fitch Co. v. Hunting World, Inc. (2d Cir. 1976) and is stated identically by the USPTO itself (USPTO, "Strong trademarks", read 25 October 2026: https://www.uspto.gov/trademarks/basics/strong-trademarks):

(Further statements of the spectrum: https://legalclarity.org/what-is-a-descriptive-trademark-and-can-you-register-it/ ; https://revisionlegal.com/trademark/what-is-a-fanciful-trademark-and-why-does-it-matter/ ; https://Trademarkraft.com/blogs/news/understanding-descriptiveness-and-genericness-in-trademarks)

Two consequences founders routinely miss:

1. The same word moves categories with the product. "Apple" is generic for apples, descriptive-adjacent for cider, arbitrary for computers. You do not trademark a word; you trademark a word for specified goods/services, organised in the Nice Classification's 45 classes — for a tech parent, typically Class 9 (software), 42 (SaaS / software services), often 35 (business services), 36 (payments, if relevant) and 41/38 depending on the roadmap. A name can be clear in Class 25 (clothing) and hopeless in Class 42.
2. Descriptive is a trap disguised as clarity. Even after secondary meaning is acquired, a descriptive mark cannot stop competitors fairly using the same words descriptively — you buy a registration and still do not own the language. IPWatchdog's blunt practitioner version: "from a legal standpoint you really should stay away from the descriptive category at all costs" because enforcement scope stays narrow even when registration succeeds. (https://ipwatchdog.com/2017/11/18/selecting-trademark-business/)

2.2 What actually blocks a name: likelihood of confusion, §2(d)

Registration is refused under Trademark Act §2(d), 15 U.S.C. §1052(d), if the mark so resembles a registered (or previously used) mark that, used on or in connection with the applicant's goods/services, it is likely to cause confusion, mistake, or deception as to source. The analysis applies the DuPont factors (In re E.I. du Pont de Nemours & Co., 476 F.2d 1357 (CCPA 1973)) — thirteen factors, of which USPTO office actions repeat that two dominate in most cases: (1) the similarities between the compared marks (sight, sound, meaning, overall commercial impression), and (2) the relatedness of the compared goods/services — with fame of the prior mark, trade channels, consumer sophistication, actual confusion, and number/nature of similar marks on similar goods also weighed, case by case, no factor mechanically dispositive. (USPTO/TTAB office-action text applying §2(d) and DuPont: https://ttabvue.uspto.gov/ttabvue/ttabvue-92076545-CAN-1.pdf)

Operational implications:

• Sound-alike = block, even if spelled differently. Because similarity is judged on sound and impression as well as appearance, a clever respelling does not clear a confusing name — "Klear" does not escape "Clear" in the same field. This kills the common founder theory that a variant spelling is a legal workaround.
• Relatedness expands the danger zone beyond your exact class. Software and software-adjacent services are routinely held related (Class 9 goods vs Class 42 services for the same function are the textbook pair); examiners cite third-party registrations showing one source offering both. A prior mark in an adjacent line can block you; a prior mark in a genuinely remote line usually cannot.
• A weak prior mark blocks less. If the field is crowded with similar marks ("crowded field"), each is entitled only to a narrow scope — which is why fields full of "-ly", "Cloud-", "-hub" names are simultaneously easy to enter and impossible to defend.
• Refusal is not the only cost. Even a registrable name can draw an opposition after publication, a cease-and-desist from a senior user with common-law rights (below), or a UDRP/ACPA fight if the matching domain was registered in bad faith against an existing mark. Budget and timeline in §8 assume a professional search precisely because the expensive failures are the invisible ones.

2.3 Common-law vs registered rights — use creates rights, registration creates national rights

Established (US). Trademark rights arise from use in commerce, not from filing. But their geography differs radically:

(Comparisons: https://www.shopify.com/blog/trademarks-small-business ; https://www.upcounsel.com/what-does-a-trademark-protect ; https://legalclarity.org/registered-vs-unregistered-trademark-key-differences/ ; fee: https://globaltrademarkcompany.com/blog/what-happens-if-i-dont-register-my-trademark)

The canonical illustration is Burger King of Florida, Inc. v. Hoots, 403 F.2d 904 (7th Cir. 1968): the Hoots family's small Mattoon, Illinois "Burger King" restaurant predated the national chain's Illinois presence; the court preserved their common-law right — within roughly a 20-mile radius of Mattoon, where to this day the national Burger King cannot operate under its own name — while the federal registrant took everywhere else. Both sides "owned" Burger King; geography and registration decided where. (https://www.jdsupra.com/legalnews/trademark-rights-explained-common-law-3256401/)

Why this matters to a naming search: a clean USPTO database result does not mean a name is free. Senior common-law users — a regional SaaS firm, an app, an open-source project with real users — will not appear in TESS, but they can still block, oppose, or force a geographic carve-out or a rename. That is why §8 sequences a web/app-store/GitHub common-law sweep before the registrar, not after.

2.4 Why a famous-sounding name can be unusable, and an odd name bulletproof

Four distinct mechanisms, often conflated:

1. Fame cuts against newcomers. A famous mark gets broader protection, including anti-dilution protection (§43(c)) even across unrelated goods for truly famous marks. A name that "sounds like" an admired tech brand inherits its confusion risk, not its equity. You cannot borrow distinctiveness.
2. Descriptiveness of the admired name in your field. "Summit", "Compass", "North" feel big-brand safe because big brands use them — in other fields. In software, each is already used by multiple firms, giving you a crowded-field weak mark at best.
3. Coexistence is normal and lawful across distance. Delta (airline / faucets), Apple (computers / the Beatles' Apple Corps — which litigated for decades precisely because their fields converged into music distribution), and Alphabet: when Google adopted Alphabet in 2015, BMW already owned Alphabet (its fleet-management subsidiary, founded 1997) and alphabet.com; BMW said it would "examine the legal trademark implications" and was "not planning to sell the domain" — and both Alphabets coexist to this day, because fleet management and internet services are remote fields and neither trades in the other's channel. (https://techcrunch.com/2015/08/12/bmw-alphabet-is-more-than-a-name-to-us/ ; https://www.automotive-fleet.com/news/bmw-fleet-unit-not-selling-alphabetcom-to-google) The lesson is precise: identical names coexist when fields are remote; near-identical names die when fields touch. Field distance, not fame, decided it.
4. Coined names are bulletproof because there is nothing to collide with. A fanciful word has, by definition, no prior user, no dictionary meaning to argue descriptiveness over, and no crowded field. Its only vulnerability is accidental phonetic/visual proximity to an existing mark — which is what clearance searches for. Kodak/Exxon/Xerox were not lucky; they were empty, and empty is defensible.

Genericide — the success disease (established). Marks that became the category name have been cancelled: aspirin (Bayer, US), escalator (Otis Elevator), thermos, cellophane, zipper, dry ice. Owners of near-generic famous marks (Kleenex, Band-Aid, Xerox, Google) therefore police usage — mark as adjective, not noun/verb ("Kleenex® brand tissues"). Google survived a verb-based genericide challenge in Elliott v. Google (9th Cir. 2017) because verb use alone does not prove the public treats the mark as the category name — a useful nuance: becoming a verb is not fatal; becoming the only word for the thing is. (https://www.mondaq.com/india/trademark/1618992/once-iconic-now-generic-the-erosion-of-trademark-identity ; https://www.lexology.com/library/detail.aspx?g=ebd53846-e3ce-48cf-be5c-fec4504224a0) For a new parent this is a distant risk, but it argues for pairing a coined name with a generic descriptor from day one ("___ brand tools", not "a ___").

3. Domain economics

3.1 The installed base: .com is not a preference, it is the default

Established data. Verisign's Domain Name Industry Brief (DNIB), Q4 2025: 386.9 million registrations across all TLDs; .com alone 161.0 million, .net 12.5 million; new gTLDs collectively only 12.4% of all registrations; combined .com/.net renewal rate 75.0%. (https://goldsteinreport.com/2025-ends-386-9-million-domain-names-verisign/) Verisign's renewed .com registry agreement reportedly permits up to 7% annual wholesale increases in each of the final four years of the term (from 2026), from a $10.26 wholesale base — even the cheapest safe TLD has a built-in escalator to model. (https://markets.financialcontent.com/chroniclejournal/article/marketminute-2026-1-2-verisign-schedules-2025-year-end-results-as-market-eyes-2026-pricing-power-and-infrastructure-expansion)

3.2 TLD trust and perception: what the surveys do and do not show

Vendor/survey evidence — directionally consistent, methodologically soft. No peer-reviewed TLD-trust literature of comparable quality to §1 exists; treat numbers as indicative, and note most are commissioned by domain-industry firms.

• AtomRadar survey (2025), reported by Nominus: 76% of consumers say familiarity with a TLD increases their likelihood of trusting a site on it; roughly 70% of US consumers implicitly trust .com; trust in .ai reached ~34% among 35–54-year-olds (its strongest cohort) and ~14% among 55–64-year-olds; respondents aged 55–64 were about half as likely to trust .ai/.io as 18–34-year-olds. (https://www.nominus.com/blog/do-ai-domains-build-more-trust-than-com)
• Companion consumer-behaviour figures from the same industry literature: users are reported 3.8× more likely to assume an unfamiliar website ends in .com, and 76% automatically assume an unknown site uses .com; 68% express higher trust for their local ccTLD when seeking regional services. (https://www.nominus.com/blog/cctld-or-gtld-smart-domain ; cf. https://register.domains/en/blog/domain-extensions-com-vs-net-vs-new-tlds)

Type-in / default behaviour is the mechanism, not "trust" in the abstract. People who remember a name but not an extension guess .com. If you are not the .com, three leaks follow (practitioner analysis: https://brandhunt.com/blog/why-com-still-dominates-2026): (1) type-in and word-of-mouth traffic goes to the .com owner (a competitor, a parker, or worse); (2) email misdirection — customers, investors and counterparties send to [email protected] by habit; if that mailbox belongs to someone else you have a confidentiality incident, not just a lost lead; (3) every non-.com use carries an explanation cost ("no, dot-ai") in exactly the spoken contexts §1.3 and §5.1 show are already the weakest channel.

The segmented truth: inside technical audiences, .io (read as I/O) and .ai (read as AI) signal category fluently and carry no penalty — .io can read as more credible for developer tools/SaaS/APIs, and .ai now functions as a category claim (https://register.domains/en/blog/io-vs-com). Outside technical audiences — a parent company's investors, enterprise procurement, press, regulators, older users — the penalty returns. For a parent over multiple brands, the audience is by definition the general one, which is why the parent belongs on .com (or its home ccTLD) even if individual products live happily on .io/.ai/.dev.

ccTLD-as-gTLD risks are real and asymmetric:

• .io is the British Indian Ocean Territory's ccTLD. The UK and Mauritius signed a Chagos sovereignty treaty on 22 May 2025; as of the research date it was signed but not in force (implementing legislation shelved), BIOT still exists in ISO 3166-1, and .io operates normally (registrations grew ~660,000 in 2021 → over 1.6 million by 2025; registry revenue £31.6m in 2024). Retirement could only follow removal of "IO" from ISO 3166-1, and ICANN's ccTLD retirement policy then allows an orderly 5-year wind-down, extendable to 10. Precedents cut both ways: .su has outlived the USSR by 30+ years; .yu, .tp and .zr were retired. Verdict: no near-term risk, a genuine decade-scale tail risk — acceptable for a product, a disclosure item for a parent. (https://assets.openprovider.com/what-is-io-domain/ ; https://en.wikipedia.org/wiki/.io ; https://www.webstacks.com/blog/is-the-end-of-io-domains-near)
• .ai is Anguilla's ccTLD, and its price is sovereign policy. .ai generated roughly $32m in 2023 (over 20% of Anguilla's government revenue), ~$39m in 2024 (~a quarter), with projections near $49m for 2025 and ~90% renewal rates reported. (https://www.jdsupra.com/post/contentViewerEmbed.aspx?fid=2fc14bcb-5fd9-40f8-a0db-83e6b1d6a623) A registry whose national budget depends on your renewal has every incentive to raise it; .ai already renews at ~7× .com (§3.5).

3.3 Length, characters, hyphens, numbers: what sales data actually say

This is the best-quantified part of domain economics, and it corrects two pieces of folklore.

Established / large-sample data:

• Bishopi analysis of 480,667 .com sales (2026): median price by length is a cliff, then a plateau. 3 characters: $1,425 median (≈8× any longer bucket); 4: $234; 5: $181; 6: $194; 7: $194; 8: $186; 9–12: $183; 13+: $193. Beyond the 3-character scarcity tier, the market does not pay for shortness itself — it pays for word quality. (https://bishopi.io/blog/what-480667-com-sales-reveal-about-the-2026-domain-economy)
• NameBio 5-year analysis (reported by DN.com): 2-letter .com averaged ~$1,000,000 (median $900,000); 3-letter .com averaged ~$66,200; excluding 2–3-character names (and outliers such as voice.com, $30m in 2019), the length–price correlation is present but weaker than expected, and for long names the .com/.org price gap narrows. (https://dn.com/en-us/news/detail-792.html)
• Lindenthal (2014), "Valuable words," Journal of the Association for Information Science and Technology — 243,291 Sedo sales, 2006–2013: domains that sold averaged 9.87 characters, vs 12.59 for domains merely listed and 14.36 for all registered .coms; sold names contained digits in 5.68% of cases (vs 7.00% of all .coms) — the market selects for shorter, cleaner strings, but the sold median price was just $500: most domains, even decent ones, are worth very little. (https://www.lindenthal.eu/assets/papers/Lindenthal-Valuable-Words.pdf)
• Word count beats character count. Zetetic's 2004 study of 405 secondary-market sales found number of words was nearly 8× more significant than number of characters in determining value — one-word names dominate premium sales (single-word English names were 63 of the NameBio Top 100 sales in 2024). (https://www.namepros.com/threads/domain-names-words-more-valuable-than-characters.43467/ ; https://www.greengeeks.com/blog/domain-name-worth-how-to-find-your-domain-value/)
• Scale check on headline prices. NameBio recorded roughly 190,300 sales totalling ~$244m in 2025 — an average near $1,300, heavily weighted by wholesale/investor trades. Headline sales (e.g. bot.ai, $1.2m, reported as the largest public .ai sale) describe the top of a lottery, not a market a founder can budget against. (https://sullysblog.com/the-problem-with-comparing-your-portfolio-to-public-domain-sales)

Hyphens and numbers (expert consensus, partially data-supported): valuation guides converge on hyphens/numbers reducing value by roughly 50–80% and .com commanding multiples of alternatives (one 2026 guide: ".com commands 10–100× premiums"; another: .net/.org realise 10–30% of the equivalent .com, .io/.ai 20–50%) — these multipliers are practitioner estimates, not regression coefficients, and are quoted as such. (https://names.center/blog/domain-valuation-guide.html ; https://domaindetails.com/kb/domain-investing/where-to-sell-domain-names) The mechanism is not aesthetic: a hyphen fails the radio test outright ("is that hyphenated?"), and a number forces a second ambiguity (digit or spelled-out? 4 or "for"?). Lindenthal's data show digits are under-represented among sold names, consistent with a discount. Rule: no hyphens, no numbers in a parent domain, without exception; if the clean form is unaffordable, change the name, not the spelling (see §1.6 — the same conclusion from the consumer side).

3.4 Premium and registry-reserved pricing: two different things sold under one word

Confusing these is the most common domain-budgeting error.

(https://manage.resellerclub.com/kb/node/1949 ; https://www.namepros.com/threads/godaddy-premium-domains.1076977/)

A registry-premium name at $2,000/yr is not a $2,000 decision; it is a $20,000 decade and a permanent hostage — once your email and brand sit on it, switching costs make the renewal effectively non-negotiable for the registry to raise. §3.5's .inc figures are this structure at TLD scale.

3.5 Renewal-price traps and the real cost picture, with live prices

Verified live — Porkbun pricing page, https://porkbun.com/products/domains, read 25 October 2026. The page states prices are for non-premium, single-year registrations, renewals and transfers, including ICANN and other fees.

*10-year = first year + 9 renewals at today's prices, ignoring future increases; .com wholesale increases (§3.1) make even row 1 conservative. Cross-checks: DomScan's Porkbun table, same week — .com $11.08/$11.08, .io $28.12/$51.80, .co $15.76/$31.20, .ai $82.70/$82.70, .xyz $2.04/$14.21, .info $3.60/$22.14 (https://domscan.net/domain-pricing/registrars/porkbun); Dynadot's 2026 explainer — .com $10.88/$10.88, .io $28.89/$53.50, .xyz $1.99/$13.17, .inc $257.71 / $2,060.21 (https://www.dynadot.com/blog/why-domains-cost-more); Network Solutions lists .inc at $2,999.99 first year / $3,152.99 renewal, showing registrar mark-up variance on the same registry price (https://www.networksolutions.com/domains/domain-name-pricing/i).

The trap, quantified. A comparison of advertised first-year vs renewal prices across registrars (vendor compilation, directional) found GoDaddy $0.01 → $21.99, IONOS $1.00 → $20.00, Hostinger $2.99 → $15.99, Namecheap $6.49 → $16.98 — loss-leader first years recovered at renewal, when the domain carries your website, email and brand and moving feels risky. (https://sidebysidedomains.com/blog/domain-renewal-trap-exposed.html) Porkbun's own list makes the structural point without any registrar villainy: 384 of its 637 extensions renew for more than they cost to register (DomScan count).

So what does the $20–$50/year band actually buy, and what does it cost in full?

• It buys a credible, non-.com or discounted-.com-adjacent identity with honest renewals: .co ($31.20), most mid-tier new gTLDs at renewal (.online/.site ~$28.84, .agency $27.29, .company $19.05 — Porkbun live list), or .io at its edge ($51.80). It does not buy .ai ($82.70) or any prestige ending.
• The full picture is never one domain. A parent typically holds: the primary, the .com (or the primary is the .com), 2–4 defensive variants (the obvious misspelling, the homophone, the hyphenated form — held defensively, never used publicly), and product domains later. At $20–50/yr, a five-name defensive portfolio is $100–$250/yr, $1,000–$2,500 per decade — trivial against payroll, and that is the correct comparison: domain cost is noise; rename cost (§4.3) is the catastrophe being insured against.
• The two ways $20–50/yr goes wrong: (1) the price was a first-year promo and renewal is $80–$300 (new-gTLD standard pricing does this routinely — always price year 2, not year 1); (2) registry-premium renewal (§3.4) hidden in checkout fine print. The defence is procedural: record the renewal price at purchase, register for multiple years where the renewal is locked, enable auto-renew + registrar lock, and calendar a transfer review 60 days before each renewal (transferring to an at-cost registrar is the standard escape).

3.6 Email deliverability and TLD reputation

Established-ish, with an important correction to folklore. A TLD does not, by itself, determine inbox placement: mailbox providers weigh authentication (SPF/DKIM/DMARC), sending history, content, infrastructure and recipient behaviour, and Interisle's phishing measurements measure reported phishing domains, not inbox placement. (Careful explainer, updated 25 Oct 2026: https://www.suped.com/learn/email-deliverability/what-tlds-should-be-avoided-for-email-domains-due-to-spam-or-reputation-issues) That source flags .top, .cfd, .icu, .cyou (and to a lesser degree .vip, .click) for elevated phishing-domain scores in May 2025–April 2026 Interisle data. Two real risks remain even with perfect authentication: (1) blanket TLD blocking — some corporate gateways block entire cheap/abuse-heavy TLDs; Spamhaus explicitly discourages the practice while acknowledging networks do it (https://www.spamhaus.org/resource-hub/service-providers/getting-the-low-down-from-xyz-registry-on-combating-domain-abuse/); (2) human perception — a security-review or finance recipient who does not recognise the ending treats the mail as suspect before any filter does. For a parent whose domain will carry investor, legal and enterprise-sales email, "technically deliverable" is not the bar; .com / established ccTLD / .org carry no such tax.

4. Case studies — successes and failures

4.1 How the major tech parents/brands got their names

4.2 Names that failed, backfired, or trapped their owners

• Mondelez (Kraft spinoff, 2012) — the pronunciation + foreign-slang double failure. From 1,700 employee entries, vetted — Kraft said — with consumers in 28 languages, Mondelez ("monde" = world + "delez" ≈ delicious, pronounced mon-dah-LEEZ) still launched to "MONDEWHAAAAT?" headlines: most English speakers mispronounce it on sight, and Russian speakers heard an echo of a vulgar term for oral sex (the manda- element), a connection Northwestern's Prof. Irwin Weil confirmed on the record. Shareholders approved it anyway (>90% of votes cast) and MDLZ trades under it today — proof that a bad name is survivable by an incumbent with distribution, and simultaneously that survival is not vindication: the name has carried a mockery tax for 14 years. (https://www.manufacturing.net/operations/news/13174404/kraft-shareholders-give-mondelez-thumbs-up ; https://www.acclaro.com/blog/kraft-foods-a-lesson-in-global-marketing/ ; https://www.marketingweek.com/the-brands-that-got-lost-in-translation/)
• Mitsubishi Pajero — one product, three names, equity that never compounds. Pajero (named for a Pampas cat, Leopardus pajeros) is Spanish slang for "wanker". Mitsubishi rebadged the identical SUV as Montero in Spanish-speaking markets and North America and Shogun in the UK, keeping Pajero elsewhere (including Brazil, where Portuguese doesn't carry the slur) — three badge identities, three advertising pools, three decades of word-of-mouth that cannot merge. The pattern recurs: Hyundai sells the Kona as the Kauai in Portugal (proximity to a vulgar Portuguese word); Mazda's Laputa collided with Spanish la puta. (https://depix.ai/blog/mitsubishi-sold-the-same-suv-under-three-names-because-the-original-meant-wanker-in-spanish ; https://www.bubblestranslation.com/the-worst-marketing-translation-errors-ever/)
• The Nova legend — a failure that did not happen, and why it still matters. The Chevrolet Nova story ("no va" = "doesn't go", therefore it failed in Latin America) is false as history: the Nova sold well in Mexico and Venezuela, and Spanish speakers parse a brand name as a name, not a sentence — the same reason English speakers don't refuse to buy a car called "Notable" ("not a table"). It survives in every business school because the underlying risk is real (Pajero proves it). Cite it as a legend that teaches a true lesson, never as a case.
• Product-trap names that had to be amputated. Dunkin' Donuts → Dunkin' (2019, §1.7); Weight Watchers → WW (2018, escaping both "weight" stigma and diet-category narrowing — and buying an unpronounceable-as-a-word initialism in exchange, §1.2's warning made flesh); IHOP's temporary IHOb stunt (2018) drew ridicule precisely because it advertised how binding the "P" (pancakes) had become. Wharton's synthesis of the wave is the useful generalisation: renames pay when they remove a factual constraint (menu, category, geography) and keep every other recognition cue (Dunkin' kept font, colours, and the name customers already used); they fail when they are cosmetic or thesis-driven. (https://knowledge.wharton.upenn.edu/podcast/knowledge-at-wharton-podcast/five-reasons-to-rename-your-company/)
• Recognition destruction without a rename — Tropicana (2009). Not a name change, but the best-measured demonstration of what a rename risks. PepsiCo spent a reported $35m on the Arnell redesign (orange-with-straw replaced by a glass of juice, wordmark rotated vertical); within two months sales fell 20%, a ~$30m loss (≈$50m including reversion), and the redesign was withdrawn — customers literally could not find the product on the shelf. Gap's 2010 logo change lasted six days. Lesson for naming: existing recognition is a capital asset with a measurable balance; a new company has none to lose, which is exactly why its name decision must be right the first time — it will build the asset once. (https://elvtr.com/blog/five-examples-of-failed-rebranding-that-cost-companies-millions ; https://widartoimpact.com/journal/the-hidden-data-behind-tropicana-s-30-m-redesign-failure-nobody-has-ever-seen)
• Unpronounceable / unsayable coinages as a class. The 2000s pharma/consulting coinage style (Altria — mocked because its claimed Latin derivation doesn't exist; Embarq — Sprint's local-phone spinoff spelled with a Q, greeted by naming consultant Steve Manning with "Em what?") shows the failure mode: a name engineered to be visually distinctive that nobody can say, and therefore nobody does. Embarq's own defence — executives wanted it "visually and verbally distinctive" and to avoid the limiting word "telephone" — got the constraint right (no product words) and the phonetics wrong. (https://igorinternational.com/press/fortune-naming-branding-firm.php ; https://igorinternational.com/press/kcstar-branding-process-business-naming.php)

4.3 What separated the good outcomes from the bad — synthesis

Across §4.1–4.2, five variables explain nearly all the variance, and none of them is "creativity":

1. Constraint order. Winners specified legal/linguistic/domain constraints before falling in love (Accenture's 47-country/200-language/URL brief; Kodak's three rules). Losers generated first and vetted later — Mondelez vetted in 28 languages and still shipped, which is worse: vetting without veto power is theatre.
2. One pronunciation, one spelling, in the wild — tested on strangers, not colleagues (Google, Stripe post-rename, Block vs Mondelez, Embarq, /dev/payments).
3. Product-agnosticism at parent level (Alphabet, Block, Verizon, Accenture vs Dunkin' Donuts, Weight Watchers, and Meta's thesis-commitment as the intermediate case).
4. Exact-match digital assets secured as part of the name decision (Meta's pre-acquired meta.com/@meta; Stripe's stripe.com purchase; Alphabet as the deliberate, affordable exception).
5. A story that survives telling. "Accent on the future", "veritas + horizon", "a collection of letters / alpha-bet" are all post-hoc — that is normal and harmless. What matters is that the story is tellable in one sentence without embarrassment in any launch language. Yahoo's backronym and Google's typo both pass; Mondelez's "delicious world" required a pronunciation guide and a denial, and failed.

Cost asymmetry (the economic summary of this section): a rename at scale costs $35m–$300m+ in direct marketing alone in the cases above, plus unrecoverable recognition (Tropicana's 20% in 60 days). A rigorous naming-and-clearance process at founding costs, by comparison, low five figures in search/attorney time and a domain budget. Naming is the cheapest insurance a company ever buys, purchased at the only moment it is cheap.

5. The AI-era factors (2020s)

5.1 Voice assistants and speech-to-text: the radio test, enforced by machines

Established mechanism. Automatic speech recognition is a prediction system: acoustic evidence is resolved toward frequent, plausible word sequences. Unfamiliar names, brands and specialist terms — rare in training data, often with unusual spelling–pronunciation mappings — are therefore systematically replaced by familiar alternatives when the audio is ambiguous (explainer with the canonical failure: spoken "Book a table at Noma" → transcribed "Roma" → assistant acts on the wrong entity). (https://www.howaimodelswork.com/2026/06/why-voice-ai-mishears-certain-words.html)

Numbers, with vendor caveats:

• Benchmark word-error rates look reassuring in aggregate — AssemblyAI's comparison put its Universal-2 at 6.68% WER vs Whisper large-v3 at 7.88% — but the same benchmark singles out proper nouns as the hardest category, with double-digit relative error differences between models (vendor benchmark; treat as directional). (https://webflow.assemblyai.com/blog/comparing-universal-2-and-openai-whisper)
• On names specifically, the gap explodes. A 2026 evaluation of 15 state-of-the-art models on San Francisco street names (2,262 utterances, 78 speakers) found an average transcription error rate of 39% on street names; Whisper-Large scored 14% general WER but 27% on street names; speakers whose primary language was not English scored 46% accuracy vs 64% for English-primary speakers. Proper names are context-poor and intolerant of near-misses — a one-phoneme error is a different place, or a different company. (https://www.together.ai/blog/how-speech-models-fail)
• Practitioner domain guidance converges: homophones (to/too/two; sight/site/cite; sole/soul) are the top voice failure for brands; defensive registration of phonetic variants is the standard mitigation, because the assistant, not the user, chooses which spelling to visit. (https://dn.org/?p=26898)

Test that follows (built into §8): the dictation test. Speak the candidate name, in a sentence, to at least three systems (e.g. phone dictation on two platforms, a Whisper-based transcription app) and to a voice assistant as a search/command, five times each, ideally including a non-native speaker. Require ≥90% exact transcription of the intended spelling — a name the machine hears as a different existing word fails outright, however well it scores with humans.

5.2 AI search / model discoverability: being an unambiguous entity

Emerging evidence (2026 studies, commercial research — labelled accordingly).

• Assistants disagree with each other. LLM Scout's analysis of 749 identical commercial queries across ~27,000 AI answers (ChatGPT, Claude, Gemini, Perplexity) found the systems disagreed on whether to mention a tracked brand in 27.2% of cases. (https://www.issuewire.com/pdf/2026/10/new-llm-scout-research-reveals-significant-differences-in-how-ai-assistants-recommend-brands-IssueWire.pdf)
• And they churn. Envoyra's study (2,720 answers over four weeks, reported by Search Engine Land): assistants disagreed on the top brand in four of five software categories, and 79% of question-and-assistant combinations changed their recommended brand list at least once in four weeks; only 144 of 680 combinations were stable. (https://searchengineland.com/ai-brand-recommendations-vary-platform-week-497967)
• Citation is fragmented. Wellows' analysis of 22.7m AI citations (Jan–Jun 2026) reports only ~11% of domains are cited by both ChatGPT and Perplexity, and that brand search demand and entity recognition — not backlink volume — are the strongest predictors of citation frequency (vendor research, directional). (https://wellows.com/blog/llm-citation-trends-for-ai-search/)
• The selectivity cliff. SOCi's 2026 Local Visibility Index (350,000 locations, 2,751 brands) found ChatGPT recommends just 1.2% of brand locations vs a 35.9% appearance rate in Google's local 3-Pack — AI answers have no "page two". (https://www.soci.ai/news/in-ai-driven-discovery-few-brands-are-chosen-most-disappear/)

What this means for the name (inference, labelled as such): there is as yet no peer-reviewed study isolating name form as a variable in model recommendation. But the mechanism above implies three name-level requirements, each independently justified elsewhere in this study: (1) uniqueness — a name shared with other entities (a common dictionary word with many referents, another company, a popular open-source project) forces the model to disambiguate, and disambiguation resolves toward the most-documented entity, which is not you; a coined/arbitrary name gives the model a clean entity to attach facts to. (2) Consistency — one spelling, identical across domain, handles, GitHub, registries and press, because entity resolution aggregates exact strings; variant spellings fragment the signal. (3) A canonical, crawlable home — the exact-match domain hosting a clear "X is a company that…" statement, since retrieval in current systems explicitly searches for official brand pages and documentation. Treat "AI discoverability" claims beyond these three as marketing until better evidence exists.

5.3 Social handles and usernames as part of the name decision

Practical facts (platform rules, checked via checker documentation): handle constraints differ per platform and the intersection is the binding constraint — X: 1–15 characters, letters/numbers/underscore only; Instagram: up to 30, periods/underscore allowed; TikTok: 2–24, letters/numbers/underscore/period, no trailing period; LinkedIn company pages and YouTube handles have their own rules. A name that must be identical everywhere should therefore be ≤15 characters, letters (and at most one underscore) only, no periods/hyphens. (https://theblue.social/tools/username-availability-checker ; https://www.socialcal.app/username-checker)

Rules that follow:

• Exact-match or a principled variant, decided in advance. Acceptable fallback conventions, in descending order of cleanliness: the bare name → name + a category-neutral suffix used everywhere (hq, app, co) → "get/use/go" + name for product accounts. A different improvisation per platform is the failure mode — it looks like impersonation and fragments search.
• Check who holds the taken handle. An active, unrelated holder in your field is a different problem from a dormant squatter or a fan account: the former is a confusion risk (§2.2 logic applies socially), the latter may be claimable via platform trademark processes after registration — another reason trademark filing precedes launch publicity.
• Claim defensively at decision time, including platforms you will not use. The cost is zero; the cost of recovering a squatted handle after announcing the name is not. (https://www.meticulosity.com/blog/handle-with-flair-crafting-creative-and-catchy-social-media-handles-that-stand-out ; https://namecheckly.com/blog/social-media-name-checker)

5.4 App-store, GitHub and package-name collisions — the tools-company surface

For a company that ships developer tools, the name lives in namespaces a consumer brand never touches, and two of them are security surfaces, not just branding:

• Package registries (npm, PyPI, crates.io, RubyGems). Check the exact name and near-variants for: availability, existing packages with the same name in a different scope, and download counts of anything similar. The threat model is documented: in 2017 npm removed 38 typosquatting packages from one account, including crossenv (the real package is cross-env), which exfiltrated environment variables — credentials — from installing machines; PyPI and crates.io mitigate partly by treating - and _ as equivalent so both spellings cannot have different owners. (https://www.webcodingcenter.com/package-managers/supply-chain-security-across-ecosystems--typosquatting-and-dependency-confusion.html)
• Slopsquatting — the 2025 escalation. Spracklen et al., "We Have a Package for You!", USENIX Security 2025: across 576,000 code samples from 16 code models, 19.7% of 2.23 million package references named packages that did not exist (205,474 unique hallucinated names; ≥5.2% for commercial GPT models, 21.7% for open models); on re-running, 43% of hallucinated names recurred in all 10 runs — model-hallucinated names are predictable, so attackers pre-register them and wait for an AI to recommend the install. A brand name that is also a plausible package name must be defensively registered on npm/PyPI/etc. at founding, and a name that collides with an existing popular package invites both confusion and supply-chain suspicion. (https://aiprogrammingmanual.com/slopsquatting/)
• Dependency confusion (Birsan, 2021) is the internal mirror: if your internal packages share names with public ones, builds can be hijacked — use scoped names (@yourbrand/…) internally from day one, which presupposes a brand string clean enough to be a scope.
• GitHub / app stores. A GitHub organisation name that is taken — or that differs from the company name — permanently splits documentation, stars and search. App stores add their own constraints: duplicate/similar app names are a review-rejection and confusion ground, names are length-capped in display, and a taken exact name forces a suffixed listing that reads as a knock-off. None of these namespaces is checked by a domain registrar search; §8 makes them a separate screening row.

6. The social layer: why some names sound cool — and others sound wrong

Added 26 October 2026. The first edition of this study treated a name as a cognitive, legal and economic object. It is also a social object: it is judged, instantly and mostly unconsciously, against the sound of the era and the company it keeps. This section covers that layer — word aversion and word love, sound symbolism as congruence, the psychology of cool itself, and the current high-tech register — from a second research round. The framework in §8 now screens for it.

6.1 Word aversion is semantic, not phonetic

Established evidence. The best-designed research programme on word aversion is Thibodeau (2016, PLoS ONE 11(4): e0153686 — "in semantics not sounds"), five experiments on the most-discussed aversive word in English; across samples, roughly 13–20% of respondents were averse to it. Three findings matter for naming:

• People misattribute their own aversion. Averse participants say the sound disgusts them. But their aversion tracked the word's semantic associations — above all with bodily functions: the same respondents also disliked vomit and phlegm, and were barely affected by the sound-alikes hoist and foist. If sound drove the aversion, the sound-alikes should disgust too. They do not.
• Aversion is socially transmitted. An aversion can be induced experimentally; it shifts with the company the word keeps (culinary vs sexual neighbours change the rating); and many respondents report minding the word only after someone pointed it out. A public "gross word" label does work the phonetics never did.
• Who is averse: younger, more educated, more disgust-sensitive respondents, and women more than men — close to the demographic that reviews, writes about and amplifies new companies.

Consequences for screening. (i) Run the association check before the sound check: list the candidate's nearest semantic and phonetic neighbours and ask specifically about bodily, infantile, medical and excretory neighbours — that is where cringe lives. (ii) A name that has to be defended ("it doesn't actually sound like…") has already lost: the audience's stated reason will be wrong, but the feeling is real, and it spreads socially. §8.4's cross-generational test operationalises this.

6.2 Phonesthemes: clusters that drag meaning with them

Established linguistic pattern (descriptive tradition). Below the whole-word level, English has phonesthemes — sound clusters recurring in words of related meaning at above-chance rates, so a new coinage containing the cluster inherits the flavour (survey: https://en.wikipedia.org/wiki/Phonestheme):

• gl- light and vision: gleam, glisten, glow, glitter
• fl- movement: flap, flee, flicker, flutter, fly
• st-/str- stability, strength, pointedness: stand, static, steel, sting; straight, strand, streak, stripe
• sn- nose and mouth, largely pejorative-adjacent: sneeze, sniff, snore, snort, snot, snarl, sneer
• sl- frictionless motion and a large pejorative set: slide, slick — but also slime, sludge, slobber, sloppy, slouch, sloth, slum, slur, sly
• squ- compression and wet impact: squash, squelch, squish
• -ump heaviness and roundness: lump, clump, thump, slump; -ash violent contact: crash, smash, splash

This study's synthesis: for a premium or high-tech parent, sn-, sl- and squ- onsets and -ump/-udge codas are loaded dice — not because the sounds are ugly, but because the existing word-stock they activate is bodily or pejorative (the §6.1 mechanism, one level down). Conversely, st-/str- and gl- clusters arrive pre-loaded with strength and light.

6.3 Word love: the beautiful-word profile, and its tension with tech

Cultural/historical + expert synthesis. The mirror tradition is phonaesthetics. Cellar door is the canonical exhibit — cited since Cyrus Lauron Hooper's 1903 novel Gee-Boy as beautiful purely for sound, dissociated from sense and spelling, and popularised by Tolkien in his 1955 lecture "English and Welsh" ("More beautiful than, say, sky, and far more beautiful than beautiful"). David Crystal (1995, "Phonaesthetically Speaking") extracted the recurring profile of words English speakers list as beautiful (recurring examples: gossamer, melody, tranquil): *three or more syllables; stress on the first syllable; /l/ the most common consonant, followed by /m, s, n, r, k, t, d/; short vowels favoured (schwa, then the vowels of lid, led, lad) over long vowels and diphthongs; three or more manners of articulation, with approximants (liquids and glides) most common. His near-perfect specimen is tremulous*.

The tension to manage: the beautiful-word recipe (longer, liquid- and nasal-rich, flowing) is almost the opposite of the cool-tech recipe in §6.7. Both are real registers. Beauty reads organic, calm, literary, premium-in-a-soft-way; the tech register borrows only pieces of it — first-syllable stress, sonorant flow between harder edges — which is why Linear and Notion (near-pure phonaesthetic shape) still read as tech when the category context supplies the frame. A parent that chooses a purely beautiful name is choosing the organic register, not a neutral one.

6.4 Sound symbolism is congruence, not quality

Established evidence. §1.4 established that sounds carry trait meanings; this round's experiments sharpen the point: the effects are fit effects, not quality effects.

• Klink (2000, Marketing Letters 11(1)) judged 124 nonsense-word pairs differing on one phonetic dimension at a time. Front vowels read smaller, lighter in colour, milder, thinner, softer, faster, colder, friendlier, weaker; voiceless stops read smaller, faster, lighter, sharper than voiced stops; voiceless fricatives read smaller, softer, more feminine than voiced. Flagged disagreement: secondary summaries conflict on Klink's headline stop-vs-fricative comparison — which pole reads smaller/faster/lighter is reported both ways; the vowel and voicing effects are consistent across summaries, the stop/fricative direction is not, and it is carried in "Could not verify."
• Yorkston & Menon (2004, JCR 31: 43–51) — phonetic embodiment: an ice cream named Frosh (back vowel) was judged smoother, richer, creamier — and evaluated more favourably — than the identical product named Frish (front vowel), automatically and outside awareness. The back vowel promised richness; for ice cream, richness is the product.
• Lowrey & Shrum (2007, JCR 34: 406–414): front vowels preferred for a convertible and a knife (small, fast, sharp categories), back vowels for an SUV and a hammer (large, slow, dull). Names containing negative-valence vowel sounds were least preferred regardless of category. And under articulatory suppression (participants counting aloud while reading), the phonetic-symbolism effect was eliminated — people quite literally mouth a name to judge it. (§8.4's badge test exists because of this result.)
• The frequency code (Ohala's framing) organises the pattern: front/high vowels and voiceless obstruents signal smaller, lighter, faster, non-threatening; back/low vowels and voiced obstruents signal larger, heavier, more intense — because small vocal tracts make high sounds and large ones make low sounds, across species. Bouba/kiki (§1.4) is the shape-level version: ~95% agreement in Ramachandran & Hubbard (2001), replicated across 25 languages, 9 language families and 10 writing systems (Ćwiek et al.).

Premium is two signals, and the evidence points both ways — reported as such. Pathak, Calvert & Lim (2017) found back vowels over-represented in luxury vs basic real brand names; Pathak, Velasco, Petit & Calvert (2019, Psychology & Marketing 36(10)) found, via Zipf's law, that names longer than three syllables are perceived as more luxurious, monosyllables suiting basic brands. Motoki, Park, Pathak & Spence (2022, Journal of Business Research 150) found higher-frequency sounds (front vowels, fricatives, voiceless consonants) raise evaluation but lower potency, and — in hospitality luxury tests — increased luxury perception, with the letter v specifically increasing behavioural intention, via conspicuousness. Resolution (synthesis): potency/weight (back vowels, voiced stops, length) and conspicuous distinctiveness (rare letters, higher-frequency sparkle) are different premium signals that studies keep separating. Enterprise gravitas buys the first; consumer-luxury sheen buys the second. A tech parent usually wants potency without heaviness: one back vowel or one voiced frame inside an otherwise light name.

6.5 Letter-level effects: why K, X, Z, V read edgy

Established evidence + synthesis. Schloss (1981) analysed real top brand names and found they began disproportionately with a, b, c, k, m, p, s relative to dictionary words; Vanden Bergh (1990) replicated the skew. (A stronger version — 54% of 1979's top-200 brands beginning with c, p or k vs 19% of English words, k alone at 6% vs 1% — circulates in secondary sources; only the weaker over-representation claim was verified in this pass. The percentages are flagged in "Could not verify.") Lowrey, Shrum & Dubitsky (2003; copytesting data, 480 ads) found semantic appositeness, paranomasia (puns) and initial plosives positively related to brand-name memory — for less familiar brands especially.

Why K/X/Z/V specifically read as edgy or high-tech — four mechanisms (this study's synthesis): (i) rarity — k, x, z, v are among the least frequent letters in ordinary English, so they are conspicuous (Motoki's v result is the experimental version); (ii) kiki geometry — their letterforms are angular, matching the jagged/fast/sharp pole; (iii) phonetics — /k/ is a voiceless plosive (small/fast/sharp per Klink), /z/ and /v/ are voiced fricatives (energy with friction), /ks/ is a two-consonant cluster no soft word wants; (iv) semantic borrowing — x and z are the letters of algebra, genetics, radiation and science fiction, so they import a technical register wholesale. The dose matters: one marked letter inside a fluent frame reads designed; three read gamer tag — §6.6's bounded-autonomy rule applied to spelling.

6.6 Cool is bounded autonomy

Established evidence. Warren & Campbell (2014, "What Makes Things Cool? How Autonomy Influences Perceived Coolness," Journal of Consumer Research 41(2): 543–563; six studies) found that perceived coolness rises when something expresses autonomy — deviation from a norm — but only when the deviation is appropriate, under three conditions:

1. The norm deviated from is seen as unnecessary or illegitimate. Rebelling against a rule nobody respects is cool; rebelling against a respected one is just rude or weird.
2. The deviation is bounded. Coolness is curvilinear in autonomy (Studies 3, 4a, 4b): it rises as deviation increases from zero, then falls when deviation becomes large or perpetual. A small, occasional departure reads as confidence; a total departure reads as instability.
3. The audience leans countercultural. Consumers higher in counterculturalism — who see norms as overly repressive — tolerate and reward more autonomy.

Study 1 (an autonomously styled water bottle, cooler than an equally liked normative design) and the final study established that coolness is distinct from liking: consumers prefer cool brands specifically when their goal is to express autonomy and stand out, not when the goal is to fit in.

Naming translation (this study's synthesis): a candidate name is a deviation claim. Ask: which norm does it break, does our audience consider that norm legitimate, and how big is the break? Stripe breaks the norm "company names describe the company" — a norm tech audiences consider illegitimate — by a small amount (it is still an ordinary English word). An unpronounceable scholarly coinage breaks the norms of pronounceability, spellability and word-likeness simultaneously: deviation past the peak of the curve, i.e. weird, not cool. This is the precise mechanism behind "obscure scholarly words feel weird rather than cool": obscurity is unbounded deviation with no legible intent, and intent is what makes deviation read as autonomy rather than error.

Brand coolness has ten faces and a life cycle (established evidence). Warren, Batra, Loureiro & Bagozzi (2019, "Brand Coolness," Journal of Marketing 83(5): 36–56; focus groups, depth interviews and nine quantitative studies) find cool brands perceived as extraordinary, aesthetically appealing, energetic, high status, rebellious, original, authentic, subcultural, iconic, and popular — a measurable scale, though no brand needs all ten and some conflict (popular vs subcultural). The life cycle matters more than the list: brands become cool first to a small niche (where subcultural, rebellious, authentic, original dominate), and only some graduate to mass cool (popular, iconic); coolness in turn predicts brand attitude, satisfaction, word-of-mouth intention and willingness to pay. Translation: a parent-company name does not need to be popular-cool at launch; it needs to be niche-cool to its first audience (here, developers and technical buyers) without foreclosing mass cool later. Names that open with mass-cool signals (a fashionable suffix everyone already uses) skip the niche phase — and the niche phase is where the authenticity is earned.

Two factors of cool (established evidence). Dar-Nimrod et al. (2012, "Coolness: An Empirical Investigation," Journal of Individual Differences 33(3); North Americans, N=918) factor coolness into two orthogonal factors: cachet coolness (the primary factor: friendly, attractive, competent, confident, trendy, driven — status-promoting and socially desirable) and contrarian coolness (rebellious, rough, emotionally controlled/muted, ironic); replicated in Dar-Nimrod, Ganesan & MacCann (2018). Translation: every candidate sits somewhere on these two axes, and a parent must choose its primary axis deliberately. Stripe/Notion/Linear are cachet names (competent, confident, clean) with a whisper of contrarian; a contrarian-led parent name will thrill a niche and tax every enterprise sales cycle for a decade. The cringe failure mode is usually accidental contrarianism — a name the founders meant as cachet that the market reads as a joke it isn't in on.

Subcultural capital and the "trying too hard" line (cultural theory). Thornton (Club Cultures, 1995, adapting Bourdieu) defines subcultural capital as the status currency inside a scene, "embodied in the form of being 'in the know', using (but not over-using) current slang and looking as if you were born to perform the latest dance styles." The parenthesis is the whole theory of cringe: over-use is the tell. Insiders under-signal; aspirants over-signal. A name stacked with era markers (dropped vowel and an X and an -ify) is over-used slang in corporate form — it announces that the company learned the register from outside, recently, and is anxious about it. The signal being read is not the features but the restraint.

Why fashions die: divergence and abandonment (established evidence). Berger & Heath (2008, "Who drives divergence?", JPSP 95(3): 593–607; seven studies) show people select tastes that distinguish them from out-groups and abandon tastes when out-groups adopt them, to avoid signalling an undesired identity (in their field study, undergraduates stopped wearing a wristband once the "geeky" dorm next door adopted it). Translation: naming templates are tastes. When a template (-ly, -ify, dropped vowels, .io, an -ai suffix) is adopted by enough later or lower-status companies, early adopters' own names are unaffected — but new names in the template inherit the template's new meaning: "follower." Era fatigue (§6.7) is this mechanism running on a 7–10-year clock.

6.7 The high-tech register, and era fatigue

Cultural/historical. The canonical fictional tech-company name, Cyberdyne Systems (Terminator), is cyber- + dyne (the unit of force). It reads as authentic because real Cold-War engineering firms built the template first: Rocketdyne (1955) and Teradyne (1960 — the name chosen to signify 10¹² dynes, i.e. 10 meganewtons). The Greek/Latin prefix + hard consonant frame + technical suffix (-dyne, -ex, -ix, -on, -um, -tek) template is therefore not "fiction-sounding"; it is engineering-sounding, and fiction borrowed the credibility. Prefixes meta-, cyber-, neo-, omni-, proto- work the same way: they are claims about category position (beyond, machine-mediated, new, all, first) compressed into a learned morpheme. Meta is the limiting case — a prefix that ate the whole name (§7.1d).

Era fatigue, documented (expert/practitioner + historical).

• The -ify wave: Wall Street Journal coverage of the post-Spotify wave reported naming consultant Christopher Johnson counting 102 startups ending in "ify," up from a handful five years earlier (Xtify, Stackify among the followers).
• The dropped-vowel wave began as a domain workaround, not an aesthetic. Flickr's founders wanted Flicker; flicker.com was taken by an unwilling seller, so the e was dropped. Twitter launched as Twttr because twitter.com was taken, then "bought the vowels" once it could. The aesthetic of an era was, at origin, two companies' domain constraints — the strongest possible evidence for §7's claim that name and domain were always one decision.
• Crunchbase News (2019), "Startup Names May Have Passed Peak Weirdness" (with naming consultant Foden): creative misspellings — dropped vowels, y-for-i, extra consonants — are getting less popular because founders are tired of spelling the name out to everyone who hasn't seen it written. They persist at seed stage (Houwser, Swytch, Wurk) — i.e. the pattern now signals early and unfunded rather than edgy. Crunchbase's 2017 analysis of 1,000+ seed-funded startups found the same fashions from the other side of the peak.
• Vowel-dropping has a measured cost: Pathak and colleagues (2022, as reported in trade coverage) found fake brand names without vowels were read as more "rugged" but also less competent, less sophisticated and less sincere — with the accompanying observation that surviving dropped-vowel names are exactly those where "when you drop vowels, the readability still exists."

What the admired current names share (this study's synthesis, applying §§6.1–6.6). Stripe, Notion, Figma, Linear, Vercel, Arc, Meta — sonically: 1–2 syllables (Arc: one; Figma the outlier shape at two open syllables), first-syllable stress, pronounceable on first sight by any English speaker, spellable after one hearing. Exactly one marked feature each. Stripe: the /str-/ strength cluster (§6.2) + voiceless plosive close (kiki-crisp, §6.4). Vercel: the rare initial V (§6.5's conspicuousness letter) inside an otherwise soft frame that echoes excel/accelerate. Meta: a Greek prefix promoted to a full name — one deviation, huge but singular. Arc: maximal brevity as the single deviation. Figma: an invented word built from the most ordinary CV syllables (fig + ma) — deviation in meaning-invention, zero deviation in sound. Linear and Notion: no phonetic deviation at all — real dictionary words; their deviation is purely semantic (a maths/quality word, an idea word, claimed as a company). No template suffix. None ends in -ly, -ify, -hub, -base, -hq, -ai. Each is sui generis in shape while ordinary in sound — bounded autonomy rendered phonetically, and under-use rendered orthographically. What dated names share: the inverse profile — ordinary invention (a template word + fashionable suffix) combined with marked surface (misspelling, dropped vowel, doubled letter). The deviation is spent entirely on the surface, where it dates fastest, and none is left for meaning, where it would compound.

7. Domain strategy: the name + domain pair is the asset

Scope note: §3 covered domain economics (prices, renewal traps, premium structures). This section covers only what interacts with perception, coolness and architecture: how the pair is read, by whom, and in what order to decide. Survey figures are vendor data (domain-industry firms) unless labelled otherwise.

7.1 The seven approaches

(a) Exact-match bare name on .com — the default asset. Every data point in this section is, directly or indirectly, a measurement of the distance from this configuration. The revealed-preference evidence is the upgrade pattern in §7.2: successful companies buy their way to (a) as soon as they can afford to, which is the market's own valuation of all the alternatives below. (This study's synthesis from the sources below.)

(b) Bare name on an alternative ending while someone else holds the .com. The economics of this are a leak, and it is measurable:

• GrowthBadger's 2019 study (vendor research, method as reported by the firm) found trust ratings close — .com 3.5/5 vs .co 3.4, .org 3.3 — but recall accuracy .com 44% vs .co 33%, .org 32%, and respondents 3.8× as likely to guess .com when trying to remember an address, whatever the real extension. (Vendor data.)
• Atom's AtomRadar survey (US consumers 18–64, plus investors): 77% of consumers say trust is damaged when other brands use the same name at a different extension, and 77% consider exact-match domains important; add-on words impact trust for 78% of consumers, a hyphen for 35%. (Vendor data — Atom sells domains and naming services; the direction of bias is obvious, so use for direction, not precision.)
• The leak's cost depends entirely on who holds the .com: a parked page (tolerable), an unrelated company (a standing confusion and trademark problem, §3.2), a competitor (untenable). An older SEER Interactive recall study is a useful counterweight: exotic extension was not a significant predictor of recall failure (p=0.11) once length was accounted for — short domains were ~9.88× more likely to be recalled correctly than long ones. The leak is real but it is a length-and-default effect, not extension magic. (Expert/practitioner study, as reported.)
• When (b) works: when the audience is technical (§7.3), the .com holder is inert, and — critically — when the company treats (b) as a stage, budgeting the upgrade, as Notion did (§7.2).

(c) Name-plus-word domains (get-, try-, use-, hey-, join-, go- prefixes; -hq, -labs, -systems, -app, -works suffixes). No controlled research exists on these; the following is consistent practitioner consensus (expert opinion):

• get- is the most neutral call-to-action and the classic scaffolding pattern — Dropbox (getdropbox.com) and Facebook (thefacebook.com, a different scaffolding) both treated theirs as temporary and paid to remove it.
• try- signals "evaluate us" — it frames the company as an uncommitted experiment and caps enterprise credibility; use- signals tool/utility ("use X" = X is an instrument), acceptable for a dev-tool product, wrong for a parent; hey-/join-/go- are consumer-community registers.
• -app names one product and dates fastest; it structurally caps a parent company's ambition. -hq ("our home base") is company-ish and faintly Web-2.0. -labs signals experimental/R&D — credible for a parent of multiple bets, but reads as "not the main company," which is either honest architecture or an own goal. -systems is enterprise/engineering gravitas with a sci-fi corporate register (§6.7: cf. Cyberdyne Systems) — cold, competent, deliberately unlovable. -works/-studio import a services/agency register.
• Cost context (practitioner figures, 2026): .com ≈ $12–16/yr, .io $45–60, .ai $60–100, .co/.so $25–45 — trivial next to the perception effects, except that premium-renewal alternative endings compound (§3.4).

(d) Deliberate decoupling: legal name ≠ brand name ≠ domain. The three layers do different jobs: the legal name serves regulators, courts, investors and contracts (Meta Platforms, Inc.; Alphabet Inc.); the brand name serves the public (Meta; Google); the domain serves typing and email (meta.com; google.com). Alphabet is the worked example of parent-level decoupling: Larry Page stated Alphabet was not intended to be a big consumer brand, and that its companies "should have independence and develop their own brands" — so the parent lives at abc.xyz, a domain that is itself a wordplay proof of the alphabet concept, while the consumer brands keep their own names and domains. Rule (this study's synthesis): decouple upward freely (the parent legal name can be as dull as compliance wants), decouple downward never (the brand the public says, types and emails should be one string). The failure mode is accidental decoupling — a brand name, a different trading name, and a third thing in the domain — which reads as assembled, not architected (§7.4).

(e) Country-code endings used as brands: .io, .ai, .co, .tv. These work when the letters mean something in the product language rather than naming a country — "IO" (input/output) to engineers, "AI" as a category claim, "TV" as a medium. Nobody experiences .tv as Tuvalu: Verisign paid ~$5m/year (≈1/12 of Tuvalu's GNI) for the rights, GoDaddy took over on 14 Dec 2021 at ~$10m/year — the registry itself is a media business. .ai is a category claim, with a category's risks: Anguilla's registrations grew from ~144,000 (2022) to 354,000 (2023) and past ~850,000–1 million by early 2026 (sources vary), at ~$140/two years standard registration and ~90% renewal (§3.2 gives the revenue picture). A .ai parent domain says "we are an AI company" every time it is typed — an asset only while the company wants that exact category, and a dating mechanism exactly like the -ify wave if it doesn't. .co is the worst leak profile of the major alternatives despite Google treating it as generic (§7.3): it sits one letter from .com, and the canonical cautionary tale is Overstock's O.co, where a majority of curious users reportedly ended up at O.com. Sovereign risk (recap of §3.2): .io is the British Indian Ocean Territory's code; the UK–Mauritius sovereignty treaty was signed in May 2025 and was not in force as of mid-2026 in the source consulted; under ICANN's retirement policy, if the ISO code were removed, a 5-year wind-down (extendable to 10) would follow — precedents run both ways (.yu, .tp, .zr retired; .su persists 30+ years). No near-term risk; a real long-tail risk for decade-scale brand equity.

(f) Domain hacks and wordplay endings — dated, with one exception. del.icio.us, blo.gs and bit.ly defined the form in 2004–2010. The hack's flaw is structural: the brand name and the pronounceable word diverge — nobody's company is called "del-icio-us" — so the pair fails every spoken test in §8.4 (radio, email dictation, keynote) while passing only the screenshot test. bit.ly survives because it is infrastructure, and even Bitly operates publicly as "Bitly" on bitly.com. (Note also that .ly is Libya's ccTLD, under Libyan law.) Verdict: viable as a secondary short-link domain, not as a parent identity. Hacks are the dropped-vowel era in domain form: clever once, now a date stamp.

(g) Parent architecture: subdomains/subpaths vs each brand on its own domain. Search reality first: Google's John Mueller — search is "fine with using either subdomains or subdirectories… keep things together as much as possible… use subdomains where things are really kind of slightly different"; Search Console nonetheless verifies subdomains as separate properties, and practitioner sources disagree on how much authority subdomains inherit from the root (flagged, not resolved). The architecture rule (this study's synthesis): put a brand on a subpath or subdomain of the parent (parent.com/product, app.parent.com, docs.parent.com — Apple and Google both run major properties this way, e.g. tv.apple.com, developers.google.com) when the brand should borrow and repay parent trust and share the parent's email namespace and reputation. Give a brand its own domain when it needs its own email namespace, its own (possibly cooler, more subcultural) identity, insulation from the parent's reputation, or clean sale/spin-out optionality — the Alphabet logic in (d). The parent company's own name+domain is the trust root; that is precisely why the parent pair must be the most conservative, most exact-match pair in the portfolio, even if product brands take more deviation.

7.2 The upgrade pattern: the strongest evidence in this section

Revealed preference beats survey data. Companies repeatedly launch on compromised domains and pay real money to reach the bare .com as they scale (expert/practitioner compilation; prices as reported in secondary sources):

• Dropbox operated on getdropbox.com until October 2009, when it acquired dropbox.com and renamed the company Evenflow, Inc. → Dropbox, Inc. in the same month — legal name, brand name and domain consolidated in one move; the price is reported at ~$300,000 in secondary sources.
• Notion: notion.so → notion.com (retaining .so). Mercury: mercury.co → mercury.com. Metronome: getmetronome.com → metronome.com. Monograph, Nio and others follow the same path.
• Facebook: thefacebook.com → facebook.com, bought for $200,000 in 2005; then FB.com bought from the American Farm Bureau for $8.5 million (2010/11) — explicitly primarily for internal email addresses (§7.3).
• Tesla: teslamotors.com → tesla.com after ~a decade of negotiation, reported at ~$11 million in secondary sources.

Reading (this study's synthesis): nobody upgrades away from the bare .com. Every compromised configuration in §7.1 is, in the market's own behaviour, a temporary configuration — which means choosing one permanently is choosing to pay the apology tax (§7.4) forever instead of the purchase price once. It also means launching on a compromised domain is legitimate if the pair is screened as a stage-one pair with a credible upgrade path (an inert .com holder, a plausibly purchasable domain), and illegitimate if the .com is a competitor's fortress.

7.3 Software-company specifics

Do developers judge non-.com domains? The honest answer is: there is no peer-reviewed evidence either way, and the best available data is one vendor survey. AtomRadar reports familiarity among tech-employed respondents of 77% for .io and 89% for .ai — far above the same survey's 51% familiarity for .co among consumers generally. (Vendor data.) Developers demonstrably built the .io/.ai conventions, so within the in-group a clean alternative ending signals "technical, current" rather than "couldn't afford the .com" — Thornton's subcultural capital (§6.6) in domain form. The same survey's investor cut adds the counterweight: 73% of investors say a single-English-word .com helps significantly or is a major factor in securing funding. (Vendor data.) So the audiences split: developers forgive the extension; mainstream customers leak (§7.1b); investors price the .com. A parent company must satisfy all three, in that order of increasing strictness.

The email test. The parent domain becomes every employee's address and appears in every inbox, signature, calendar invite and LinkedIn profile; it is spoken ("I'm sarah at X dot io") far more often than the website is typed. Three consequences: (i) mis-remembered extensions misdeliver mail to whoever holds the .com — a confidentiality issue, not just a marketing one (§3.2); (ii) the extension is pronounced thousands of times a year by your own staff — "dot io" and "dot ai" are sayable, a hack or a get- prefix is a small daily humiliation; (iii) the market has priced this test exactly once, visibly: Facebook paid $8.5 million for FB.com primarily so employees could have fb.com email (§7.2). No defensible study isolates TLD effects on email trust — the statistics circulating in marketing blogs (e.g. "75% see custom-domain email as a legitimacy signal") trace to low-quality sources and are not cited here as data.

Handles and namespaces are the same sweep. The GitHub org name, npm/PyPI package scope (@name/…), X/LinkedIn/YouTube handles, and app-store developer name are the same scarcity problem as the domain, judged by the same developer audience, daily (§§5.3–5.4). A name whose GitHub org is a squatted inactive account or whose npm scope is taken will be lived with by exactly the people the company most wants to impress. No study evidence exists; this is uniform practitioner practice.

SEO and type-in reality for a tools company. Google states the TLD is not a ranking factor (beyond genuine country ccTLDs geotargeting their own country), and treats the generic ccTLDs used as brands (.ai, .io, .co, .me, .tv) as generic for geotargeting. For a tools company whose customers arrive via search results, registries, directories, integrations and AI agents clicking links, type-in traffic is a minor channel — so the SEO/type-in argument for .com is weak. What remains is the human channel: word of mouth, conference conversation, podcast mentions, and email — all of which pass through memory, where the 3.8× .com default (§7.1b) operates. The correct conclusion is not ".com doesn't matter" but ".com matters in the spoken channel, not the search channel" — and a parent company's name lives disproportionately in the spoken channel (hiring, fundraising, partnerships).

7.4 The hand-in-hand rule

Designed vs assembled (this study's synthesis). A pair feels designed when the spoken name, the written name, the domain and the email address are the same string with nothing left over: Stripe → stripe.com → @stripe → [email protected]. It feels assembled when any layer needs an apology — a prefix ("we're at get-…"), an ending you must spell ("…dot i-o, not dot com"), a hack, or a legal name that matches neither. Every apology is a recurring tax, paid in keynotes, sales calls, support chats and email dictation, by every employee, forever. Conversely, one moment of visible design in the pair — Alphabet's abc.xyz — is a decade of free signalling that the company sweats details.

Decision order (this study's synthesis, applying §§7.1–7.3):

1. Generate pairs, never bare names. A candidate does not exist until it is name + specific domain + price.
2. Screen each pair in one pass: the cool screen (§8.4) for sound/semantics/social reading; trademark first pass (§8.2 Phase 4); domain price and renewal terms (Phase 5); the full namespace sweep (Phase 6).
3. Kill pairs that fail any gate rather than falling in love with a name and retrofitting a domain — retrofitting is how assembled pairs happen (and, per §6.7, how the dropped-vowel era happened).
4. Secure in one sweep: the domain (plus the defensive .com or the closest leak-plugging variants, if the primary is an alternative TLD), the trademark filing, the GitHub org, the package scopes, the handles.
5. Only then announce. Dropbox's October 2009 move — domain acquired and company renamed in the same month — and Facebook's 2005 domain purchase ahead of its relaunch are the precedents for doing it in this order.

8. A practical framework

8.1 Scored criteria for a candidate name + domain

Score each finalist 0–5 per criterion, multiply by the weight, and apply the kill rules first — a failed kill rule rejects the name regardless of total score. Weights reflect the evidence and cost asymmetries in §§1–7: legal and spoken-form failures are the least fixable, so they weigh most.

Maximum weighted score: 5 × (3+3+3+2+1+3+2+2+2+1+1+1) = 120, plus 4 kill gates. In practice, require ≥85/120 with no criterion scored 0 among the ×3 rows.

8.2 Step-by-step evaluation procedure

Phase 0 — Write the brief before generating anything (Accenture order, §4.1). One page: what the parent must cover (current + plausible businesses), launch markets/languages, audience (for a parent: investors/enterprise/press as well as users), posture (fast/precise vs large/trustworthy — picks your vowel register, §1.4), the kill constraints (classes, languages, renewal ceiling, handle requirements), and who holds veto power at each gate. Names generated without this brief cannot be judged, only liked.

Phase 1 — Generate widely, by method (target 200–500 raw candidates). Use all five generation methods, because each explores a different part of the spectrum (§1.5):

• Real word, arbitrary (Apple/Stripe/Block method): thesaurus and concrete-noun lists from adjacent domains (geology, sailing, tools, fauna) — spellable by definition, needs the hardest clearance work.
• Coined (Kodak method): build from phoneme inventory — CVCVC structures, plosive onsets, chosen vowel register, 2 syllables; generate systematically (syllable tables) rather than by vibes; screen each coinage for resemblance to any existing word in launch languages.
• Compound / portmanteau (Microsoft, Netflix, Verizon method): two real morphemes fused at a clean syllable boundary; check the fusion doesn't create a third word or a descriptive claim.
• Suggestive metaphor (the sweet legal spot, §2.1): name the benefit's image, not the product (Compass, Summit-type images — then clear them hard, this is the most crowded field).
• Myth / history / place (Nike, Amazon method): names with a pre-built story and usually a real-word spelling; check cultural/religious sensitivity and that the referent's baggage is one you want permanently.
• (Variant spellings and initialisms are not generation methods in this framework — §§1.2, 1.6 — they are, at most, a last-resort repair, scored as in criterion 10.)

Generate pairs, not bare names (§7.4). Every candidate that survives Phase 2 is immediately paired with its specific domain and renewal price; a candidate without a pair does not advance. This is the one change the second research round (§§6–7) makes to the procedure's front end — it is what prevents assembled pairs, and it is why Phase 5 below is a confirmation step rather than a first look.

Phase 2 — Meaning screen (cheap, kills most). Dictionary + slang + native-speaker pass in launch languages (K2); accidental-word/segmentation check; "say it as a customer would" read-aloud inside the team. Expect to kill 60–80% here.

Phase 3 — Sound screen. Radio test (K4) and pronunciation-unity check on survivors (10 strangers; §1.3 method); dictation test (§5.1) on the top ~10. This order matters: these tests cost an afternoon and kill names that trademark searches (which cost money per name) would have passed.

Phase 4 — Trademark first pass (K1). Knockout search per §8.1 on the top ~5: USPTO + WIPO Global Brand Database + EUIPO/UKIPO as applicable; web common-law sweep (search engines, app stores, GitHub, company pages) for unregistered users in adjacent fields (§2.3). This is a screening pass, not clearance — survivors go to a trademark attorney for a full search and opinion before any public use or filing; then file intent-to-use in the parent classes (9/42 first) in home market + Madrid designations for launch markets.

Phase 5 — Domain at a registrar (K3, criteria 7–8). For each survivor, record in the spreadsheet: exact-match availability per TLD, first-year AND renewal price from the registrar's own pricing page (the §3.5 method — never the promo banner), premium status (aftermarket vs registry, §3.4), and the .com holder's identity/status (active competitor / parked / for sale at what price). Decide buy-vs-choose-different-name with the 10-year portfolio cost in front of you.

Phase 6 — Handles and namespaces (criterion 9). The §5.3/5.4 matrix: socials, GitHub org, npm scope + PyPI name, app-store name search, YouTube. Claim/register everything available the same day the decision is made — domains, handles, org names, package names (defensive, even if unused), before any external announcement, job posting or filing becomes public and tips off squatters.

Phase 7 — Final validation and lock. 24-hour recall test (criterion 4); model probe (ask 3–4 assistants "What is [name]?" and category questions — you want "no established entity" now and an unambiguous referent later); a final counsel review; then register domains multi-year, enable registrar lock + auto-renew, set the renewal calendar, and document the name's story, pronunciation, and usage rules (adjective use, §2.4) in the brand sheet.

8.3 The classic traps, in priority order

1. Falling in love before clearance. Every expensive story in §4 starts with attachment preceding the trademark/domain facts. Procedure (§8.2 order) is the entire defence.
2. Believing a domain purchase confers trademark rights — or vice versa. They are separate systems; you need both, and each can be blocked by the other's incumbents (§§2.2–2.3, 3.4).
3. Choosing descriptive for the SEO/clarity sugar-rush. Weakest legal position, no power to stop fair descriptive use by competitors, and a product ceiling built in (§§2.1, 1.7).
4. Respelling to escape a conflict or a taken .com. Fails consumers (Costello penalty), fails the radio test, and usually fails legally too — confusion is judged on sound (§§1.6, 2.2).
5. Pricing the domain by its first year. Promo-in, premium-forever-out; .inc's $258 → $2,060 is the type specimen (§3.5).
6. Taking a non-.com parent while a stranger holds the .com. Traffic, email and trust leak to them indefinitely; Alphabet is not your precedent unless you are Alphabet (§§3.2, 4.1).
7. Skipping the foreign-language spoken check. Written checks pass Pajero-type failures because slang lives in pronunciation and colloquial registers; Mondelez was checked in 28 languages and shipped anyway — checks need veto power (§4.2).
8. Naming the thesis, the technology, or the year's fashion. Meta (thesis), any "AI/Cloud/Crypto-" prefix (technology), dropped vowels and -ify suffixes (fashion): all date or confine the parent (§§1.7, 4.1).
9. Forgetting the non-web namespaces. Handles, GitHub, npm/PyPI, app stores — discovered at launch, when the squatter/collision is maximally expensive and, for packages, a security issue (§§5.3–5.4).
10. Initialisms and unpronounceable coinages. WW and Embarq: if people cannot say it, they will not, and the market will rename you on its own terms (§§1.2, 4.2).
11. No defensive registrations. Misspellings, homophones and phonetic variants left unregistered are left for typosquatters, phishers and speech-recognition error to find (§§3.5, 5.1).
12. Treating the name as finished at launch. No usage rules → genericide drift; no monitoring → infringements harden into rights; no renewal calendar → the domain itself lapses, the one unforgivable failure (§2.4).
13. Deciding the name and the domain separately (added 26 October 2026, §7.4). Falling in love with a bare name, then discovering the pair is an assembled compromise — the leak, the misspellings and the spelled-aloud extension were locked in at the moment of attachment. Generate pairs, screen pairs, secure pairs in one sweep.
14. Over-signalling cool (added 26 October 2026, §6.6). Stacking marked features — rare letters, dropped vowels, fashionable suffixes, in-group slang — overshoots bounded autonomy into deviance: it reads as trying too hard, the one reading cool cannot survive. One deviation, deliberately chosen and congruent, is the whole budget.

8.4 The pair screen: cool, judged on the name + domain together

Added 26 October 2026. Apply after — never instead of — the kill gates (§8.1) and the procedure (§8.2). Every test below judges the name+domain pair (§7.4); the evidence base is §6 (cool) and §7 (domain strategy).

1. The keynote sentence. "Hello, we're X." Said on a stage, cold, to people who have never seen it written. Pass = no spelling, no gloss, no laugh you didn't intend. (Warren & Campbell, §6.6: the deviation must read as intentional; a name that needs explaining reads as error.)
2. The headline test. Put the pair in a trade-press headline — and in a critical headline ("X lays off 200"). Pass = it reads as a company, not a product feature, a typo, or a joke — and it survives the hostile version.
3. The job-title test. "VP of Engineering, X" on LinkedIn; "X Fellow"; "X, Inc. v. —" in a court caption. Parent names live in these frames for decades; product-cute names fail here first. (Dar-Nimrod's cachet axis, §6.6, asserting itself.)
4. The radio/phone test — for the pair. Say the name once, aloud, to someone with a keyboard. Pass = they type the domain and the email correctly, first try, including the extension. This single test kills hacks, dropped vowels, creative misspellings and (for mainstream-facing parents) alternative endings — and is the operational form of the 3.8× .com-default data (§7.1b).
5. The email test (§7.3). firstname@pair, ceo@pair, support@pair — read them in an inbox list among Stripe's and Google's mail. Pass = indistinguishable in dignity.
6. The conference-badge test. The name on a badge, read across a table, by a non-native speaker, at the end of a long day. Pass = pronounceable from the badge alone. (Lowrey & Shrum's articulatory-suppression result, §6.4, is why this matters: people mouth names to judge them; a name that can't be mouthed can't be judged kindly.)
7. The cross-generational reading. A 55-year-old CFO and a 22-year-old engineer each say it aloud and say what the company probably does. Pass = neither produces the bodily/infantile/wrong-category reading (Thibodeau's association check, §6.1) and neither produces the era-template reading ("oh, like a 2016 app?").
8. The hoodie test — and its limits. Would staff wear the name on a hoodie, unironically, in public? It is a useful niche-cool probe (Warren et al., §6.6: cool starts subcultural) and practitioner folklore — no study or source for the test itself was found, and it is stated here as folklore, not evidence. Its limits: it selects for the taste of young tech in-groups and against enterprise buyers, regulators and anyone the parent must also impress; a name can pass the hoodie test and fail the job-title test catastrophically. Use it as a veto-detector for embarrassment, never as an approval criterion.
9. The namespace sweep (§7.3). GitHub org, package scopes, handles, app-store name: all clean, all matching, all secured in the same sweep as the domain. A pair that fails here is not a pair.
10. The ten-year test. Read the name as a 2036 incumbent: "X, founded in the AI era…" Pass = the name contains no era template (-ify, -ly, dropped vowel, .ai-as-identity, an -ai suffix) that Berger & Heath's abandonment mechanism (§6.6) will have re-labelled "follower" by then.

Phonetic recipes (this study's synthesis, from §6).

• Recipe A — cool/high-tech parent: 1–2 syllables (3 maximum for a parent), first-syllable stress; an otherwise fluent sonorant/liquid frame; exactly one marked feature — one rare letter (K/X/Z/V), or one voiceless-plosive frame, or one unusual-but-sayable cluster; front-vowel brightness for speed or a single back vowel for weight — never both signals at once; no template suffix; the semantic field checked against bodily/infantile/wrong-category neighbours (§6.1) before phonetics are even discussed.
• Recipe B — organic/calm/premium-soft: Crystal's profile (§6.3) — 3 syllables, first-syllable stress, liquids/nasals (/l, m, n, r/), short vowels, open syllables — optionally with length-as-luxury (§6.4). Choose this knowing it reads organic and literary rather than technical; it suits a parent whose authority is taste, not engineering.

The cringe list (this study's synthesis). Obscure scholarly/Greek/Latin words no one can pronounce or spell on first encounter (unbounded deviation, §6.6). Stacked marked features: X + Z + dropped vowel + fashionable suffix in one name (over-used slang, §6.6). Any era template used a decade late: -ly, -ify, dropped vowels, doubled letters, -hub/-base, an -ai suffix on a non-AI company (§§6.6–6.7). Category-locking shapes for a parent: -app, get-/try- presented as the final identity, a .ai identity for a company that may not always be an AI company (§7.1). sn-/sl-/squ- onsets and -ump/-udge codas for anything premium (§6.2). Names whose nearest neighbours are bodily, infantile, or a service the company doesn't offer (§6.1). Pair-level cringe: hacks, an extension that must be spelled aloud, a legal name / brand / domain that are three different strings by accident (§§7.1, 7.4).

Worked pairs (from §6–§7's analysis).

Stripe / stripe.com. One syllable; /str-/ strength cluster (§6.2) + voiceless plosive close = kiki-crisp (§6.4); a real word whose deviation is purely semantic. Cachet axis, near-zero contrarian. The keynote, headline, job-title, radio and badge tests are all trivially passed — which is the point: the coolness is in the restraint.

Figma / figma.com. Two open CV syllables anyone on earth can say; the invention is entirely in the referent (a figment made collaborative). Demonstrates Recipe A's central claim — deviation can live in meaning while sound stays ordinary — and Lowrey, Shrum & Dubitsky's finding that appositeness aids memory: the name explains itself retroactively once you know the product.

Linear / linear.app. The name is Recipe B's diction (liquid onset, real word, quality claim) wearing Recipe A's brevity — and "linear" is a promise about the product (no curves, no detours). The pair is the caveat: linear.app is a name-plus-word configuration (§7.1c) that names a product, not a parent; under §7.2's logic it should be treated as scaffolding with an upgrade budget, because a parent company will outgrow "-app" framing exactly as Dropbox outgrew "get-".

Vercel / vercel.com. The initial V is the single marked feature (§6.4's conspicuousness letter) inside a soft frame that echoes excel/accelerate; the pair only became designed after the company abandoned Zeit (German for "time" — unpronounceable-on-sight for most of its market, on zeit.co) and rebuilt name and domain together. The before/after is the cleanest illustration in this study that a rebrand is usually a pair repair, not a name repair.

Notion / notion.com (via notion.so). A dictionary word claiming the idea-space itself; phonaesthetically pure Crystal shape (two syllables, nasal + liquid flow). Launched on notion.so — a defensible stage-one pair for a developer-adjacent audience (§7.3) — then upgraded to the bare .com as it crossed to the mainstream, keeping .so. The pair's history is §7.4's decision order executed correctly.

Dropbox / dropbox.com (via getdropbox.com). A rounded, utterly apposite compound (the box you drop things in) that spent its early life in scaffolding; in October 2009 the company bought the bare domain and renamed the legal entity in the same month. If one case proves the hand-in-hand rule, it is this one: the company itself treated get- as a temporary brace and the consolidation as a single transaction.

Meta / meta.com. The deviation is enormous — renaming one of the world's largest companies after a Greek prefix meaning "beyond/self-referential" — but it is one deviation, executed at the brand layer while the legal layer (Meta Platforms, Inc.) absorbs the formality and the domain layer (meta.com) stays a single clean string. It passes §7.1d's rule (decouple upward, never downward) and Warren & Campbell's test: the norm broken ("parent companies have descriptive names") is one its audience had already stopped respecting.

Alphabet / abc.xyz. The deliberate anti-pair: a parent that refused consumer-brand status (Page, §7.1d), given a domain that is a conceptual proof rather than an address. It works because the brands underneath it (Google, YouTube) carry their own exact-match pairs. Copied by a company whose parent is the public brand, it would fail every test above — the exception that defines the rule.

Quibi / quibi.com — the rejected case. Quibi ("QUI-ck BI-tes," coined with Siegel+Gale) raised $1.75 billion, launched 6 April 2020 and shut down 1 December 2020, selling its library to Roku for under $100 million. The name failed the screen on every axis this study measures: pronunciation was genuinely ambiguous in coverage (kwih-bee / kwee-bye / kee-bee) — failing the badge and radio tests before launch; its semantic neighbours were infantile and gustatory, and pre-launch market research reportedly found a substantial number of respondents thought it was a food-delivery service (a characterisation Quibi denied — both sides reported here); launch marketing spent $63 million (MediaRadar), including a $5.6 million Super Bowl spot whose script itself joked "A what?" — the company paid Super Bowl prices to announce that its name required explanation. The pair (quibi.com, exact-match) was flawless; the name was unbounded deviation with no legible intent — Warren & Campbell's curve, past the peak.

"Omakase" — the road not taken, correctly. Katzenberg's preferred original name for Quibi was "Omakase" (chef-selected, high-quality sushi), set aside after CEO Meg Whitman objected. The screen agrees with Whitman, for reasons orthogonal to taste: an untranslated Japanese culinary term is an obscure-scholarly-word deviation (§6.6) for a mass audience — unspellable from hearing, unpronounceable from reading for most of the market, and semantically locked to food, the precise wrong-category neighbour Quibi ended up fighting anyway.

Could not verify / open questions

• No peer-reviewed literature on TLD trust was found comparable to the §1 psychology literature; all TLD-trust numbers in §3.2 come from industry-commissioned surveys (AtomRadar via a registrar blog, and similar). Direction (familiarity → trust, .com default) is consistent across them; magnitudes should not be quoted as settled.
• Model-discoverability name effects (§5.2) are inference from mechanism + commercial citation studies; no study yet isolates name form as an experimental variable in model recommendation.
• Meta/Reality Labs cumulative loss figures (>$70bn) come from 2025–26 secondary reporting (Bloomberg-derived), not re-verified against Meta's 10-K segment disclosures in this pass.
• Registrar prices are point-in-time (25 Oct 2026, Porkbun) and vary by registrar (cf. Network Solutions' .inc pricing, ~50% higher); .ai's 2-year-term checkout presentation differs by registrar. Any purchase decision should re-read the live pricing page and the checkout renewal line on the day.
• The trademark section is US-federal in detail. EUIPO absolute-grounds practice, UK IPO practice, and first-to-file jurisdictions (notably China) differ in ways that matter for filing order; the framework's attorney step (Phase 4/7) is where those are resolved.
• The frequently repeated claim that Nokia "Lumia" is Spanish slang for prostitute is widely reported but contested by linguists; it is therefore treated only as a reported example of the §4.2 pattern, not asserted as fact. The Nova story is asserted only as a debunked legend.
• Klink (2000), stop-vs-fricative direction (§6.4). Secondary summaries of Klink's headline comparison conflict (stops vs fricatives as the smaller/faster/lighter pole). The vowel and voicing effects are consistent across sources; the stop/fricative direction is reported as unresolved. Resolving it requires the original Marketing Letters paper, which was not accessible in full text during the second research round.
• Schloss (1981) percentages (§6.5). The claim that 54% of top-200 brands of 1979 began with c/p/k (vs 19% of English words), and that k alone was 6% of brand names vs 1% of words, circulates in secondary sources; only the weaker claim (over-representation of a, b, c, k, m, p, s initials) was verified against a research summary. The percentages are flagged as secondary wherever used.
• No peer-reviewed evidence was found isolating (i) developer/technical-buyer preference among TLDs, (ii) TLD effects on email trust or perception, or (iii) market readings of get-/try-/-hq/-labs patterns. §7 relies on one vendor survey (AtomRadar) and practitioner consensus, and is labelled accordingly. The commonly circulated email-legitimacy statistics trace to low-quality marketing blogs and were deliberately not used.
• The hoodie test (§8.4) has no identifiable source or study; it is reported as practitioner folklore with stated limits.
• .ai figures (§7.1e) vary by source and year (registrations ~850k–1m+ by early 2026; revenue ~$32m 2023 / ~$39m 2024). Figures are given as reported, not reconciled. The second round's sources were read via web index on 25–26 October 2026; vendor survey methodologies (AtomRadar, GrowthBadger) could not be independently audited, and domain-upgrade prices (§7.2) rest on widely reported company histories and secondary compilations rather than primary filings.

Sources

All sources read 25 October 2026. "Verified live" = read on the seller's own page in a live session on that date; all other sources were retrieved the same day.

Psychology & linguistics (§1) - Alter & Oppenheimer 2006, PNAS full text: http://pages.stern.nyu.edu/~aalter/PNAS.pdf ; Princeton/innovations-report summary: https://www.innovations-report.com/education/studies-and-analyses/report-60425/ - Green & Jame 2013, JFE: https://ideas.repec.org/a/eee/jfinec/v109y2013i3p813-834.html ; SSRN: https://papers.ssrn.com/sol3/papers.cfm?abstract_id=1777256 ; HBR summary: https://hbr.org/2014/01/size-matters - Song & Schwarz 2009: https://www.psychologicalscience.org/article?ru=https%3A%2F%2Fjournals.sagepub.com%2Fdoi%2F10.1111%2Fj.1467-9280.2009.02267.x ; non-generalisation critique: https://www.researchgate.net/publication/314198171_If_It_s_Difficult_to_Pronounce_It_Might_Not_Be_Risky_The_Effect_of_Fluency_on_Judgment_of_Risk_Does_Not_Generalize_to_New_Stimuli - Newman et al. 2014, PLoS ONE: https://journals.plos.org/plosone/article?id=10.1371%2Fjournal.pone.0088671 - MDPI 2026, ticker fluency in modern markets: https://www.mdpi.com/2073-8994/18/3/477 - Lowrey & Shrum 2007 via Phys.org: https://phys.org/news/2007-09-vowel-affect-consumer.html ; JSTOR Daily: https://daily.jstor.org/whats-brand-name-sounds-persuasion/ ; extension: https://www.sciencedirect.com/science/article/abs/pii/S0167811613000475 - Pogacar et al., "Sounds good": https://www.researchgate.net/publication/285416618_Sounds_good_Phonetic_sound_patterns_in_top_brand_names ; PDF: https://static1.squarespace.com/static/586d3b8c579fb3b92bfadb89/t/5b0ee09ef950b77426273ea8/1527701664023/Pogacar+et+al.pdf - Bouba/kiki: https://en.wikipedia.org/wiki/Bouba/kiki_effect ; https://www.iflscience.com/the-bouba-kiki-effect-shows-a-very-strange-facet-of-language-and-human-psychology-76894 - Name-length synthesis (Pathak et al.): http://communities.springernature.com/posts/the-psychology-of-names-can-brands-appeal-to-our-subconscious-senses?badge_id=multimedia-tools-and-applications - Zee et al., length & recognition: https://www.researchgate.net/publication/257474699_Does_brand_spelling_influence_memory_The_case_of_auditorily_presented_brand_names ; SMIT Policy Brief #94: https://prod-d2bcaa46b289670c.paddlecms.net/sites/default/files/2026-06/SMIT_Policy-Brief94-Rightsizing-Brand-Names.pdf - Costello, Walker & Reczek, Journal of Marketing: https://www.fastcompany.com/90863710/brand-names-misspelled-words-consumer-study ; https://www.newswise.com/articles/unconventional-spellings-are-a-badd-choyce-for-brand-names ; https://www.eurekalert.org/news-releases/981993 - Abell & Smith on misspelling severity: https://www.asianspectator.com/index.php/world/64909-misspelled-names-may-give-brands-a-lyft--if-the-spelling-isnt-too-weird - Misspelling-wave histories: https://go2tutors.com/15-brands-that-embraced-misspelling-and-made-it-their-thing/ ; https://www.oyzta.com/blog/misspellings-as-strategy-when-incorrect-spelling-makes-for-great-tech-branding/ - Kodak/Eastman: https://www.mondaq.com/unitedstates/trademark/253212/the-best-method-to-the-madness-behind-choosing-a-company-name-or-mark

Trademark & legal (§2) - USPTO, "Strong trademarks": https://www.uspto.gov/trademarks/basics/strong-trademarks - Spectrum explainers: https://legalclarity.org/what-is-a-descriptive-trademark-and-can-you-register-it/ ; https://revisionlegal.com/trademark/what-is-a-fanciful-trademark-and-why-does-it-matter/ ; https://Trademarkraft.com/blogs/news/understanding-descriptiveness-and-genericness-in-trademarks ; https://www.justia.com/intellectual-property/trademarks/strength-of-marks/fanciful-marks/ ; https://ipwatchdog.com/2017/11/18/selecting-trademark-business/ - §2(d) / DuPont, USPTO-TTAB office action text: https://ttabvue.uspto.gov/ttabvue/ttabvue-92076545-CAN-1.pdf - Common-law vs federal: https://www.shopify.com/blog/trademarks-small-business ; https://www.upcounsel.com/what-does-a-trademark-protect ; https://legalclarity.org/registered-vs-unregistered-trademark-key-differences/ ; https://www.jdsupra.com/legalnews/trademark-rights-explained-common-law-3256401/ ; https://globaltrademarkcompany.com/blog/what-happens-if-i-dont-register-my-trademark - Genericide: https://www.mondaq.com/india/trademark/1618992/once-iconic-now-generic-the-erosion-of-trademark-identity ; https://www.lexology.com/library/detail.aspx?g=ebd53846-e3ce-48cf-be5c-fec4504224a0

Domain economics (§3) - Porkbun domain pricing (verified live, 25 Oct 2026): https://porkbun.com/products/domains - DomScan Porkbun price table: https://domscan.net/domain-pricing/registrars/porkbun - Dynadot, domain pricing explained: https://www.dynadot.com/blog/why-domains-cost-more - Network Solutions TLD pricing: https://www.networksolutions.com/domains/domain-name-pricing/i - Verisign DNIB Q4 2025: https://goldsteinreport.com/2025-ends-386-9-million-domain-names-verisign/ ; .com price-increase provision: https://markets.financialcontent.com/chroniclejournal/article/marketminute-2026-1-2-verisign-schedules-2025-year-end-results-as-market-eyes-2026-pricing-power-and-infrastructure-expansion - TLD trust surveys: https://www.nominus.com/blog/do-ai-domains-build-more-trust-than-com ; https://www.nominus.com/blog/cctld-or-gtld-smart-domain ; https://register.domains/en/blog/domain-extensions-com-vs-net-vs-new-tlds ; https://register.domains/en/blog/io-vs-com ; .com-default analysis: https://brandhunt.com/blog/why-com-still-dominates-2026 - Sales-data analyses: https://bishopi.io/blog/what-480667-com-sales-reveal-about-the-2026-domain-economy ; https://dn.com/en-us/news/detail-792.html ; Lindenthal 2014: https://www.lindenthal.eu/assets/papers/Lindenthal-Valuable-Words.pdf ; https://sullysblog.com/the-problem-with-comparing-your-portfolio-to-public-domain-sales ; https://www.greengeeks.com/blog/domain-name-worth-how-to-find-your-domain-value/ ; https://www.namepros.com/threads/domain-names-words-more-valuable-than-characters.43467/ ; valuation multipliers (practitioner): https://names.center/blog/domain-valuation-guide.html ; https://domaindetails.com/kb/domain-investing/where-to-sell-domain-names - Premium/renewal: https://manage.resellerclub.com/kb/node/1949 ; https://www.namepros.com/threads/godaddy-premium-domains.1076977/ ; renewal-trap compilation: https://sidebysidedomains.com/blog/domain-renewal-trap-exposed.html - .io / Chagos; .ai / Anguilla: https://assets.openprovider.com/what-is-io-domain/ ; https://en.wikipedia.org/wiki/.io ; https://www.webstacks.com/blog/is-the-end-of-io-domains-near ; https://www.jdsupra.com/post/contentViewerEmbed.aspx?fid=2fc14bcb-5fd9-40f8-a0db-83e6b1d6a623 - Email / TLD abuse: https://www.suped.com/learn/email-deliverability/what-tlds-should-be-avoided-for-email-domains-due-to-spam-or-reputation-issues ; https://www.spamhaus.org/resource-hub/service-providers/getting-the-low-down-from-xyz-registry-on-combating-domain-abuse/

Case studies (§4) - Google: https://www.howtogeek.com/789647/why-is-google-called-google - Yahoo: https://en.wikipedia.org/wiki/History_of_Yahoo - Alphabet: https://en.wikipedia.org/wiki/Alphabet_Inc ; https://www.techradar.com/pro/why-google-rebranded-to-alphabet-and-why-it-paid-dividends ; BMW/Alphabet.com: https://techcrunch.com/2015/08/12/bmw-alphabet-is-more-than-a-name-to-us/ ; https://www.automotive-fleet.com/news/bmw-fleet-unit-not-selling-alphabetcom-to-google - Meta: https://searchengineland.com/facebook-is-now-meta-tech-giant-announces-rebrand-375607 ; https://www.mondaq.com/unitedstates/trademark/1127228/facebook-unfriends-its-own-brand-with-the-launch-of-meta--what-does-facebooks-rebrand-teach-us - Accenture: https://www.scribd.com/document/454927758/Accenture-Translate ; https://www.eweek.com/enterprise-apps/new-name-new-challenges/ - Verizon: https://www.britannica.com/money/Verizon-Communications ; http://www.marketingprofs.com/articles/2000/205/playing-the-name-game-again - Stripe: https://summify.io/discover/patrick-collison-ceo-of-stripe-the-tim-ferriss-show-podcast-l73FKk/ - Block: https://cointelegraph.com/news/breaking-square-rebrands-to-block-as-focus-shifts-to-blockchain - Dunkin'/KFC/WW renames: https://knowledge.wharton.upenn.edu/podcast/knowledge-at-wharton-podcast/five-reasons-to-rename-your-company/ ; https://www.adweek.com/brand-marketing/dunkin-drops-the-donuts-as-the-latest-brand-to-shorten-its-name/ - Mondelez: https://www.manufacturing.net/operations/news/13174404/kraft-shareholders-give-mondelez-thumbs-up ; https://www.acclaro.com/blog/kraft-foods-a-lesson-in-global-marketing/ ; https://www.marketingweek.com/the-brands-that-got-lost-in-translation/ - Pajero & translation fails; Nova myth: https://depix.ai/blog/mitsubishi-sold-the-same-suv-under-three-names-because-the-original-meant-wanker-in-spanish ; https://www.bubblestranslation.com/the-worst-marketing-translation-errors-ever/ ; https://www.slideserve.com/dayshaun/false-powerpoint-ppt-presentation - Tropicana / Gap: https://elvtr.com/blog/five-examples-of-failed-rebranding-that-cost-companies-millions ; https://widartoimpact.com/journal/the-hidden-data-behind-tropicana-s-30-m-redesign-failure-nobody-has-ever-seen - Altria / Embarq criticism: https://igorinternational.com/press/fortune-naming-branding-firm.php ; https://igorinternational.com/press/kcstar-branding-process-business-naming.php ; practitioner radio-test framing: https://inkbotdesign.com/brand-naming/

AI era (§5) - ASR mechanism & name errors: https://www.howaimodelswork.com/2026/06/why-voice-ai-mishears-certain-words.html ; https://www.together.ai/blog/how-speech-models-fail ; https://webflow.assemblyai.com/blog/comparing-universal-2-and-openai-whisper ; voice & domains: https://dn.org/?p=26898 - Model brand recommendation: https://www.issuewire.com/pdf/2026/10/new-llm-scout-research-reveals-significant-differences-in-how-ai-assistants-recommend-brands-IssueWire.pdf ; https://searchengineland.com/ai-brand-recommendations-vary-platform-week-497967 ; https://wellows.com/blog/llm-citation-trends-for-ai-search/ ; https://www.soci.ai/news/in-ai-driven-discovery-few-brands-are-chosen-most-disappear/ - Handles: https://theblue.social/tools/username-availability-checker ; https://www.socialcal.app/username-checker ; https://namecheckly.com/blog/social-media-name-checker ; https://www.meticulosity.com/blog/handle-with-flair-crafting-creative-and-catchy-social-media-handles-that-stand-out - Packages / supply chain: https://www.webcodingcenter.com/package-managers/supply-chain-security-across-ecosystems--typosquatting-and-dependency-confusion.html ; slopsquatting / Spracklen et al. USENIX Security 2025: https://aiprogrammingmanual.com/slopsquatting/

<strong>The social layer (§6)</strong> <em>(second round, read 25–26 October 2026)</em> - Thibodeau (2016), word aversion — "in semantics not sounds," <em>PLoS ONE</em> 11(4): e0153686: <a href="https://journals.plos.org/plosone/article?id=10.1371/journal.pone.0153686">https://journals.plos.org/plosone/article?id=10.1371/journal.pone.0153686</a> - Phonesthemes: <a href="https://en.wikipedia.org/wiki/Phonestheme">https://en.wikipedia.org/wiki/Phonestheme</a> ; Phonaesthetics / <em>cellar door</em> / Crystal (1995): <a href="https://en.wikipedia.org/wiki/Euphony">https://en.wikipedia.org/wiki/Euphony</a> ; Tolkien's "cellar door": <a href="http://annasmol.net/2019/04/13/tolkiens-cellar-door/">http://annasmol.net/2019/04/13/tolkiens-cellar-door/</a> - Klink (2000) via Shrum &amp; Lowrey review summary: <a href="https://www.researchgate.net/publication/253096931_Sounds_Convey_Meaning_The_Implications_of_Phonetic_Symbolism_for_Brand_Name_Construction">https://www.researchgate.net/publication/253096931_Sounds_Convey_Meaning_The_Implications_of_Phonetic_Symbolism_for_Brand_Name_Construction</a> ; Yorkston &amp; Menon (2004), <em>JCR</em> 31: 43–51: <a href="https://citeseerx.ist.psu.edu/document?doi=a494a378ca90a287cb3f146fd0a8f8c231fabb3e&amp;repid=rep1&amp;type=pdf">https://citeseerx.ist.psu.edu/document?doi=a494a378ca90a287cb3f146fd0a8f8c231fabb3e&amp;repid=rep1&amp;type=pdf</a> ; Lowrey &amp; Shrum (2007), <em>JCR</em> 34: 406–414: <a href="https://phys.org/news/2007-09-vowel-affect-consumer.html">https://phys.org/news/2007-09-vowel-affect-consumer.html</a> - Warren &amp; Campbell (2014), <em>JCR</em> 41(2): 543–563: <a href="https://EconPapers.repec.org/article/oupjconrs/doi_3a10.1086_2f676680.htm">https://EconPapers.repec.org/article/oupjconrs/doi_3a10.1086_2f676680.htm</a> ; Warren, Batra, Loureiro &amp; Bagozzi (2019), <em>Journal of Marketing</em> 83(5): <a href="https://www.ama.org/2019/08/07/10-characteristics-of-brand-coolness-and-how-to-engineer-them/">https://www.ama.org/2019/08/07/10-characteristics-of-brand-coolness-and-how-to-engineer-them/</a> ; Dar-Nimrod et al. (2012): <a href="https://psycnet.apa.org/record/2012-15112-006">https://psycnet.apa.org/record/2012-15112-006</a> ; Berger &amp; Heath (2008), <em>JPSP</em> 95(3): <a href="https://ssrn.com/abstract=1130823">https://ssrn.com/abstract=1130823</a> - Pathak, Velasco, Petit &amp; Calvert (2019), <em>Psychology &amp; Marketing</em> 36(10): <a href="https://psycnet.apa.org/record/2019-42692-001">https://psycnet.apa.org/record/2019-42692-001</a> ; Motoki, Park, Pathak &amp; Spence (2022), <em>JBR</em> 150: <a href="https://c8be56c2-c6ee-46c8-b736-6aab19293bbd.filesusr.com/ugd/b21206_0a139ee91f954f9db8f845411cede09d.pdf">https://c8be56c2-c6ee-46c8-b736-6aab19293bbd.filesusr.com/ugd/b21206_0a139ee91f954f9db8f845411cede09d.pdf</a> ; Schloss (1981) / Lowrey, Shrum &amp; Dubitsky (2003): <a href="https://www.researchgate.net/publication/285416618_Sounds_good_Phonetic_sound_patterns_in_top_brand_names">https://www.researchgate.net/publication/285416618_Sounds_good_Phonetic_sound_patterns_in_top_brand_names</a> ; <a href="https://www.researchgate.net/publication/250174194_Enhancing_Recall_and_Recognition_for_Brand_Names_and_Body_Copy_A_Mixed-Language_Approach">https://www.researchgate.net/publication/250174194_Enhancing_Recall_and_Recognition_for_Brand_Names_and_Body_Copy_A_Mixed-Language_Approach</a> - Bouba/kiki: <a href="https://en.wikipedia.org/wiki/Bouba/kiki_effect">https://en.wikipedia.org/wiki/Bouba/kiki_effect</a> ; Thornton (1995), <em>Club Cultures</em> (subcultural capital): <a href="https://bookshop.org/p/books/club-cultures-music-media-and-subcultural-capital-sarah-thornton/ff0113eabf4752c1">https://bookshop.org/p/books/club-cultures-music-media-and-subcultural-capital-sarah-thornton/ff0113eabf4752c1</a> - Cyberdyne / Rocketdyne / Teradyne naming history: <a href="https://www.speculativeidentities.com/research/cyberdyne-systems">https://www.speculativeidentities.com/research/cyberdyne-systems</a> ; Crunchbase News (2019), "Startup Names May Have Passed Peak Weirdness": <a href="https://news.crunchbase.com/venture/startup-names-may-have-passed-peak-weirdness/">https://news.crunchbase.com/venture/startup-names-may-have-passed-peak-weirdness/</a> ; (2017): <a href="https://news.crunchbase.com/startups/startup-are-still-making-weird-name-choices/">https://news.crunchbase.com/startups/startup-are-still-making-weird-name-choices/</a> ; Retail Dive (2022), Pathak vowel research: <a href="https://www.retaildive.com/news/retail-brand-names-vowels/633644/">https://www.retaildive.com/news/retail-brand-names-vowels/633644/</a>

<strong>Domain strategy (§7)</strong> - Domain-upgrade compilation: <a href="https://bernard.digital/en/list-upgrade-domains/">https://bernard.digital/en/list-upgrade-domains/</a> ; Dropbox history: <a href="https://en.wikipedia.org/wiki/Dropbox_(service)">https://en.wikipedia.org/wiki/Dropbox_(service)</a> ; domain-upgrade case studies: <a href="https://smartbranding.com/how-businesses-transformed-after-upgrading-their-domains-facts-first/">https://smartbranding.com/how-businesses-transformed-after-upgrading-their-domains-facts-first/</a> ; FB.com purchase: <a href="https://techcrunch.com/2011/01/11/facebook-paid-farm-bureau-8-5-million-to-acquire-fb-com/">https://techcrunch.com/2011/01/11/facebook-paid-farm-bureau-8-5-million-to-acquire-fb-com/</a> - Alphabet/Meta decoupling: <a href="https://www.marketingweek.com/why-google-is-rebranding-itself-as-alphabet/">https://www.marketingweek.com/why-google-is-rebranding-itself-as-alphabet/</a> - Name-plus-word patterns (practitioner): <a href="https://nametastic.com/blog/domain-name-ideas">https://nametastic.com/blog/domain-name-ideas</a> ; <a href="https://www.colinkeeley.com/blog/how-to-pick-a-great-startup-name-domain">https://www.colinkeeley.com/blog/how-to-pick-a-great-startup-name-domain</a> ; <a href="https://wbcomdesigns.com/the-ultimate-guide-to-naming-your-online-software-business/">https://wbcomdesigns.com/the-ultimate-guide-to-naming-your-online-software-business/</a> - .co / O.co: <a href="http://register.domains/en/blog/should-i-buy-a-co-domain-complete-guide-2025">http://register.domains/en/blog/should-i-buy-a-co-domain-complete-guide-2025</a> ; Google &amp; TLDs: <a href="https://www.searchenginejournal.com/google-confirms-ranking-preference/522477/">https://www.searchenginejournal.com/google-confirms-ranking-preference/522477/</a> ; <a href="https://aamax.co/blog/does-tld-affect-seo-google-search-central">https://aamax.co/blog/does-tld-affect-seo-google-search-central</a> - .io sovereignty/ICANN: <a href="https://ccnso.icann.org/sites/default/files/field-attached/ccpdp-ret-policy-11feb19-en.pdf">https://ccnso.icann.org/sites/default/files/field-attached/ccpdp-ret-policy-11feb19-en.pdf</a> ; <a href="https://domainincite.com/tag/io-icann">https://domainincite.com/tag/io-icann</a> ; .ai/Anguilla: <a href="https://quasa.io/media/anguilla-s-ai-domain-boom-millions-in-revenue-but-there-s-a-catch">https://quasa.io/media/anguilla-s-ai-domain-boom-millions-in-revenue-but-there-s-a-catch</a> ; .tv/Tuvalu: <a href="https://en.wikipedia.org/wiki/.tv">https://en.wikipedia.org/wiki/.tv</a> - Domain hacks: <a href="https://dn.org/the-psychology-behind-memorable-domain-name-hacks/">https://dn.org/the-psychology-behind-memorable-domain-name-hacks/</a> ; <a href="https://diffsense.com/ly-the-real-story-behind-the-shortest-domain-on-the-internet-76">https://diffsense.com/ly-the-real-story-behind-the-shortest-domain-on-the-internet-76</a> ; Subdomain vs subpath SEO: <a href="https://www.siegemedia.com/seo/subdomain-seo-guide">https://www.siegemedia.com/seo/subdomain-seo-guide</a> ; <a href="https://visualmodo.com/what-is-a-subdomain-and-how-does-it-impact-seo-performance-ultimate-guide/">https://visualmodo.com/what-is-a-subdomain-and-how-does-it-impact-seo-performance-ultimate-guide/</a> - Quibi case: <a href="https://en.wikipedia.org/wiki/Quibi">https://en.wikipedia.org/wiki/Quibi</a> ; <a href="https://www.slashfilm.com/574744/quibi-alternate-name/">https://www.slashfilm.com/574744/quibi-alternate-name/</a> ; <a href="https://gizmodo.com/quibi-almost-had-an-infuriatingly-worse-name-1844054119">https://gizmodo.com/quibi-almost-had-an-infuriatingly-worse-name-1844054119</a> - Vendor studies (labelled as such wherever used): Atom / AtomRadar domain survey: <a href="https://www.atom.com/radar/domain-name-data-and-insights/">https://www.atom.com/radar/domain-name-data-and-insights/</a> ; <a href="https://www.thedomains.com/2024/11/14/atom-77-of-consumers-say-trust-is-damaged-when-other-brands-use-the-same-name-at-a-different-domain-extension/">https://www.thedomains.com/2024/11/14/atom-77-of-consumers-say-trust-is-damaged-when-other-brands-use-the-same-name-at-a-different-domain-extension/</a> ; GrowthBadger (2019) extension trust/recall study, as reported: <a href="https://smartbranding.com/backed-by-data-why-domain-extensions-matter-to-your-brand/">https://smartbranding.com/backed-by-data-why-domain-extensions-matter-to-your-brand/</a>