Every founder knows the ritual. A whiteboard fills with candidates, the room argues from taste, somebody's favourite wins, and the company spends the next decade paying for a decision made on adjectives. The most permanent choice in a business — the one on the contracts, the invoices, the front door — is usually the least examined.
It doesn't have to be a matter of taste, because the research underneath naming is measured psychology, published and replicated, and most of it reduces to numbers. So we built the numbers into a model: the naming engine inside Company Start, the company-formation tool set we're designing for the Entrepreneur Suite. It takes a candidate name — always as a pair, the name plus its exact domain — together with a profile of your business, and it computes a verdict: gates passed or failed, a score whose weights are derived from how your customers will actually meet the name, a reading of how six simulated audiences will receive it, and a verification record for every constraint you set. This piece is the plain-words tour; the formal specification is published with the tool set, so anyone can check the arithmetic.
The short version. A great name is what survives four absolute gates, scores well on twelve criteria weighted for your business, lands right for the audiences who decide your revenue, and checks out — at the seller's own page, not a price table — against every limit you declared. Everything below is that sentence, unpacked.
The science was never the problem
Start with what is already known, because the model invents none of it.
Names that are easy to process are valued more. In Alter and Oppenheimer's 2006 experiments, participants expected fluently named stocks to gain +3.90% over a year and disfluent ones to lose −3.86% — a gap produced by pronounceability alone. Green and Jame (2013) found the durable version in real markets: companies with shorter, easier names had broader ownership and higher valuations, and Harvard Business Review's summary of their work puts a price on it — one word shorter was worth about $3.75 million to the median firm.
Sound steers judgement before meaning arrives. Klink's phonetic-symbolism experiments (2000) showed front vowels read as smaller, lighter and faster, back vowels as larger and more substantial. Yorkston and Menon (2004) served the same ice cream under two names: as Frosh (back vowel) it was judged richer and better liked than as Frish — the sound promised the product.
Aversion is about association, not acoustics. Thibodeau (2016) found that people who detested a word were reacting to what it made them picture, while sound-alikes with innocent meanings bothered nobody. And cool, measured by Warren and Campbell (2014), is bounded autonomy: break exactly one norm, legibly and on purpose, and you read as confident; break three at once and the same move reads as a mistake.
Finally, the penalty for cleverness: Costello, Walker and Reczek (Journal of Marketing, 2023) tested deliberately misspelled brand names across thousands of shoppers. The correctly spelled version was chosen 62% of the time, the misspelled one 48% — one swapped letter cost fourteen points, because readers took the unexplained error as a sign of insincerity.
Every one of those findings is a dial or a tripwire in the model. None of them is a slogan.
One ruler for every business is the bug
Here is where ordinary checklists fail, including good ones: they weigh every quality the same for every company. But what matters about a name depends on how the business's customers meet it.
A software company's name is mostly seen — typed into search, read in a directory, scanned in a contract. Spelling, distinctiveness and visual shape decide; sound is a minor term. A cleaning company's name is mostly heard — word of mouth, a radio spot, a neighbour's recommendation. Pronunciation, spell-from-hearing and spoken memorability decide; its domain is a contact detail, not a storefront.
So the model takes a business profile — your estimate of where customers first meet the name, as shares across nine channels, from typed search to vehicle signage — and computes the weights from that mix. Under the study's general-purpose weights, the sound block (pronounceability, dictation survival, phonetic fit) carries 7.0 of the model's 24 weight units. Run a software profile and it drops to 3.8. Run a cleaning profile and it holds at 6.9, while the domain block halves. Same twelve criteria, same 120-point scale, same pass mark — different ruler, because the encounter is different. A name is no longer "good" in the abstract. It is good for this business, and the output shows the weights it used, so the verdict can be argued with, reproduced, or re-run.
The gates don't negotiate
Some things are not preferences, and the model refuses to average them. Before any points are scored, a pair must pass four kill gates: a trademark first pass clear in the classes and countries it will operate in (sound-alikes included — respelling cures nothing); no bad meaning, spoken aloud, in the launch languages; the exact-match domain obtainable at a sane renewal price, not a cheap first year; and the radio test — at least eight of ten strangers spelling it correctly after hearing it once.
A gate is an existence condition. Weighting one — "slightly infringing, but the logo is lovely" — would let a strong total launder a hard failure, so gates sit outside the arithmetic entirely, identical for a cleaning company and a software company, because the failures they guard are the least fixable ones a business can buy.
Audiences, simulated and weighted
Scoring tells you how a name performs. It doesn't tell you how it lands. So the model also plots every candidate on eight axes — formality, cool, warmth, playfulness, boldness, gravitas, tech register, approachability — and runs it past six simulated audiences: younger early adopters, professional buyers, enterprise procurement, trades and local consumers, parents, and investors. Each panel returns an approval estimate with an uncertainty band — a range, honestly wide at first, never a fake-precise point.
Whose verdict counts is, again, computed from the business: the profile declares which audiences decide the revenue, and the aggregate is weighted accordingly — per-panel results always shown beside it, so a decisive audience's "no" can't be averaged into invisibility. A technology brand typically weights the young adopters and the professional buyers about equally: cool to the first, safe to the second.
And some judgements are caps, not averages. The profile can declare hard bounds on an axis that hold no matter how well a name scores — the canonical one is playfulness ≤ 5: cool, yes; childish, no. A name that breaches a declared cap is flagged in the open and cannot receive an unqualified pass until the owner waives the flag, explicitly, on the record.
One honesty rule is printed on every output: these are simulated audience priors — not surveyed opinion. They are the model's structured estimate, built from the published evidence — a starting position that earns narrower uncertainty only as real outcomes accumulate against its predictions.
Your constraints are law — and verified means verified
The person using the tool sets the filters: maximum renewal per year, maximum first-year price, a total budget for the domain set, the endings allowed, the jurisdictions the name must clear in, the timeline it has to fit. The agent works inside those limits — and then proves each surviving pair against each one, because the market's own numbers cannot be taken at face value.
We learned that in the model's first live run, this week. Four short, attractive names on a fashionable tech ending passed the screen easily: the ending's standard renewal, about $51 a year, sat comfortably under the engagement's cap of $100 a year. Checked at the registrar's own page, all four were registry-premium names: $273.44 for the first year, renewing at $1,092.18 a year — nearly eleven times the cap, every year, for as long as the company exists. A fifth name the registry data showed as free was simply taken. Had the price table been trusted as proof, the tool would have presented five "available, in-budget" pairs of which exactly zero were.
So the rule is structural: registry data and price tables are for screening — they can kill a candidate, but they cannot pass one. A pair earns the label verified only with per-constraint evidence read at the seller's live page, timestamped, in the output. And when the constraints kill every candidate — it happens, and it is information, not failure — the model escalates instead of improvising: it reports which constraint binds, shows the nearest miss, and states what relaxing each limit would admit. It never quietly loosens one: a tool that edits your facts to lengthen its shortlist has stopped working for you.
A model that keeps its receipts
The model is versioned, and it is not finished — it is perfectible, on the record. Every engagement logs the business profile, the weights it produced, the model's ranking, the panels' predictions, and the name actually chosen. On a set cadence those logs are reviewed: where the chosen name ranked below a rejected one, or a panel's prediction missed a known outcome, the model's tables are adjusted in small, bounded steps — each revision changelogged, with the evidence that motivated it, never a silent self-rewrite. A model that quietly revised its own tables would quietly change what "good" means; this one shows its working, its version, and its doubts.
Where this lives
This engine is pair-screen, the scoring layer of Company Start — the naming, clearance and formation tool set in the Entrepreneur Suite, currently in design. Its companions do the checking legwork (registry and trademark first passes, country by country) and, later, the formation paperwork; the engine is the judgement layer they feed. The evidence underneath it is our naming study, published in full; the method is specified formally in ALGORITHM.md, published with the tool set's design, because a claim like "mathematically great" should come with the maths attached.
Name a company once, and live with it for decades. The guess was never necessary.
Caveats, honestly
The psychology cited here is peer-reviewed, and we have kept its limits attached: the fluency effects are real but buy first impressions, not businesses; the sound-symbolism effects are population tendencies, not horoscopes; the audience panels are simulations whose uncertainty bands are wide until real engagements calibrate them — they are labelled as simulations everywhere they appear, including in the tool's own output. The premium-pricing specimen is from a live run on October 25, 2026; registrar prices move, which is rather the point of checking them live. And the engine itself is a design-stage method, published as a method — not a product you can buy today.
The evidence base in full — every figure, the trademark framework, the domain data, the case files, and the complete scored framework — is published alongside this article: What Goes Into a Good Company Name (and Its Domain) — the study.
Sources
- Alter & Oppenheimer (2006), "Predicting short-term stock price fluctuations by using processing fluency," PNAS 103(24).
- Green & Jame (2013), "Company name fluency, investor recognition, and firm value," Journal of Financial Economics 109; Harvard Business Review summary (2014), "$3.75 million per word."
- Klink (2000), on phonetic symbolism, Marketing Letters 11(1); Yorkston & Menon (2004), "A sound idea: phonetic effects of brand names on consumer judgments," Journal of Consumer Research 31.
- Thibodeau (2016), "In semantics not sounds: word aversion," PLoS ONE 11(4).
- Warren & Campbell (2014), "What makes things cool? How autonomy influences perceived coolness," Journal of Consumer Research 41(2).
- Costello, Walker & Reczek (2023), "'Choozing' the best spelling: consumer response to unconventionally spelled brand names," Journal of Marketing.
- The method, formally: ALGORITHM.md (with DESIGN.md §5), in the Company Start tool set — published with the tool set's design in the agentpreneur repo.
The full source list, with links, is in the companion study.
← All field notes