What a machine can check, and what needs taste

A computer can count the syllables in reddit. It cannot count the syllables in a name that does not exist yet.

Syllable counts come from a pronunciation dictionary, and a pronunciation dictionary contains words. Your invented name has no entry in it, so everything downstream is a guess wearing a progress bar. Our own function is called estimateNameSyllables, and the warning it produces ends with the words verify against pronunciation evidence when available.

That gives you the shape of the whole problem in one line: a machine knows the most about the names you are least likely to use. The more original the candidate, the thinner the evidence underneath any number attached to it.

So where does machine-checkable end and taste begin? Sort every claim a naming tool makes into four kinds, ranked by what happens when you push on the evidence underneath. The first three narrow the field. The fourth picks the name, and no machine has ever reached it.

Tier 1 — string facts

Genuinely computable, cheap, and unglamorous. How long it is. Whether it contains forge, nexus, flux, or pulse. Whether eleven of your forty candidates are the same Adjective+Noun compound wearing different hats. Whether the domain string, run together and lowercased, spells something you will have to explain.

None of this tells you a name is good. All of it tells you when a name is careless, which is a different and more useful service. Sold as anything more than that, it is being oversold.

Tier 2 — phonetics, once the word exists

Phoneme count, stress placement, sonority, spelling transparency. The evidence here is the best in the field.

Laham, Koval and Alter's name-pronunciation effect (Journal of Experimental Social Psychology 48(3), 2012) ran five studies showing that easy-to-pronounce names, and the people carrying them, get judged more favorably. Study 5 is the one that should make you sit up: in a field sample of 500 US lawyers, attorneys with more pronounceable surnames sat higher in their firms' hierarchies.

This is a finding about people's surnames. Extending it to brand names is an inference. It is a reasonable inference, most of the field makes it, and it remains an inference.

Tier 2 also inherits Tier 1's trap. Phonetic analysis is solid for words that already exist and degrades to estimation exactly when you coin something, which is exactly when you wanted the help.

Tier 3 — a language model with good examples

Imagery. How many distinct associations a name sparks. Whether there is an origin story a human would retell. Whether the name sits too close to something biological.

Every naming tool lives here. Most of them present it as Tier 1. The clearest evidence for why this tier cannot be promoted comes from the moist research. Paul Thibodeau's A Moist Crevice for Word Aversion (PLOS ONE, April 2016) ran five experiments across more than 2,500 participants. Around 18% of people are averse to the word. The aversion is semantic: it tracks associations with bodily function, not sound. And here is the part that matters for anyone building a checker: the averse participants consistently blame the phonology. Thibodeau calls it "aversion dumbfounding."

The people who have the reaction cannot correctly report what causes it. A rule that predicts the reaction from the sounds is therefore predicting from the wrong variable, confidently, on behalf of people who are themselves confidently wrong about it.

Anything in Tier 3 is a judgment call with examples attached. Rendering it as a number does not promote it to Tier 2. It just hides which tier you are in.

Tier 4 — the part that decides

Whether the name is worth defending for ten years. Whether you can say it in a funding meeting without flinching. Whether it splits the room, and whether that is the point.

No machine reaches this tier. Every number you collected in the first three was narrowing a field it could not choose from, and the last step was always going to be yours.

Feeling runs through all four, which is the tier model's blind spot

A ladder from computable up to taste implies that emotion sits at the top, arriving only once the machinery has finished. It does not. The feeling a name produces is partly in the sounds, mostly in the associations, and eventually in whatever you spend ten years attaching to it. All three of those are different tiers, and only the first two exist on the day you choose.

The sounds carry real signal, and it replicates. Ćwiek and colleagues ran the largest test of the bouba/kiki effect to date (Philosophical Transactions of the Royal Society B 377:20200390, 2022): 917 participants, 25 languages, nine language families, ten writing systems. Seventeen of the 25 languages validated the effect, including pairs as unrelated as Japanese, Swedish, French and Zulu.

The effect is lopsided. Bouba matched the round shape reliably in 22 of 25 languages; kiki matched the spiky shape in only 11. Softness travels. Sharpness mostly does not. So a tool telling you your name "sounds sharp and technical" is standing on roughly half the evidence of one telling you it sounds round and friendly. The test also failed outright for Chinese, Romanian and Turkish, which is worth knowing before you pick a name for more than one market.

The sounds are still the small half. The strongest measured emotional reaction to an ordinary English word is the moist aversion above, and it is semantic. The feeling is inherited from what the word is already attached to.

A name's emotional content is mostly borrowed, and building a brand is the slow work of changing what it is borrowed from. Apple in 1976 was a fruit and a slightly odd joke. Amazon was a river. Orange was a color. None of them changed phonetically.

So you are choosing a container you will spend a decade filling. The honest question on day one is whether it can hold what you might become, which no analysis of the string can answer, because the thing that fills it does not exist yet.

That has one practical consequence worth carrying out of this page. Does it feel like us? is a poor question on day one and an excellent one on day three thousand. Early on, the questions that pay are the plain ones: can people say it, spell it, and find it, and does it already mean something you would rather it did not.

Ask what grade the evidence gets

There is one question that sorts honest tools from confident ones, and you can ask it of any flag any product raises: what is this resting on?

There are only four possible answers. Replicated research. A single decent lab study. Professional consensus among people who have named a lot of things. Or a house rule somebody chose on a Tuesday.

All four are legitimate reasons to warn you about a name. A house rule that says no fifth candidate ending in -ify is useful and worth keeping, and it has nothing to do with science. The failure is presenting the fourth kind in the typography of the first, which is what a percentage does.

If a tool cannot tell you which of the four a given flag is, the flag is decoration. If it can, you can overrule it intelligently, which is the only reason to show you a flag at all.

The scorecard

What any naming tool, including this one, is entitled to say out loud:

The claim	What would have to be true	Verdict
"This domain is available"	A registry query, with the time it ran	Fair, if the timestamp is shown
"This name is a cliché"	A maintained list, and the matched fragment shown	Fair, if it shows its working
"Your shortlist has collapsed into one formula"	Counting across candidates	Fair, and underused
"This name is easy to pronounce"	Phonetic analysis of a word in the dictionary	Fair for real words, an estimate for coinages
"This name is memorable"	A recall study nobody has run on your name	Unsupportable
"This name scores 87 / 100"	A validated model mapping name features to outcomes	Unsupportable
"This name is available across 40 platforms"	40 live queries at the moment you asked	Fair or fabricated, depending on caching

Print it. Use it on us.

The one rule that must never be encoded

A single blended score.

Every dimension in naming trades against another. Euphony fights memorability. Fluency fights distinctiveness. Descriptive clarity fights trademark strength. Collapse those into one number and the arithmetic hands you the candidate that offends no axis, which is a precise description of a committee name.

A naming tool with one number has quietly optimized for the name nobody objects to.

The names people actually defend tend to split the room on first hearing. An averaging model treats a split room as a defect and marks it down, every time, by construction.

So what is taste, concretely

Elimination is the machine's job, and it is real work: catch the carelessness, catch the collision, catch the fifth name this week ending in -ify, and show the evidence for each so you can overrule it.

Choosing is yours. Taste is knowing which trade you are making and being willing to be disliked for it, which no scorecard on this page will ever do for you.