How the score works
Four dimensions, one formula. Same name, same number, every time — no model opinion, no hidden adjustments.
The formula
overall = 0.17·pr + 0.15·ln + 0.28·or + 0.40·di + 10·(mm/100)
- pr — Pronounceability (0–100)
- ln — Length (0–100)
- or — Orthographic fluency (0–100)
- di — Distinctiveness (0–100)
- mm — Memorability bonus (0–100, adds up to +10 pts)
Weights lean toward the dimensions that actually discriminate — orthographic fluency and distinctiveness. Pronounceability and length saturate near 100 for any short, sayable name, so they carry less weight. One coupling: names scoring below 60 on orthographic fluency have their distinctiveness discounted in the composite, scaling linearly to zero — a random consonant pile is “unique,” but you can’t own a search term nobody can reproduce. Names that spell cleanly keep full distinctiveness, so for real names the overall reconstructs exactly from the four numbers on the card.
Worked example — Figma
The overall reconstructs from the four numbers on the card — no hidden curve.
How to read the number
The scale is absolute, not graded on a curve, so different kinds of names settle into different bands. Simulated across ~1,000 names:
A coinage can outscore a famous brand: the score measures the linguistics of the string, not the brand equity behind it. “Apple” scores in the 60s because the word is un-ownable in search — Apple the company won anyway, which is exactly the kind of judgment this score doesn’t make.
The four dimensions
Pronounceability
How fluently the name maps to a sayable sequence of sounds. Names easy to articulate are processed faster and feel more trustworthy — the processing-fluency effect. Phonemes come from the CMU pronouncing dictionary; coinages are estimated by grapheme-to-sound rules. Scoring is by consonant-cluster legality (Sonority Sequencing Principle) and phonotactic probability — how common adjacent sound pairs are in English, not raw cluster count.
Alter & Oppenheimer (2006, PNAS); Vitevitch & Luce (1999, phonotactic probability tables); Hofmann & Baumann (2020).
Length
Working-memory load. Shorter names are held and recalled more reliably — the word-length effect. A syllable curve peaks at two syllables and declines for longer names, combined with a smaller character-count term. Successful SaaS names average ~2.5 syllables.
Baddeley, Thomson & Buchanan (1975); Zeroual (2022).
Orthographic fluency
Reading and spelling ease — how transparently the spelling maps to the sound. Irregular grapheme-to-phoneme mappings slow reading and invite misspelling. Scored against a letter-bigram frequency table (CMU dictionary corpus): one illegible consonant pair drags the whole word down. Deliberately independent of word frequency to avoid double-counting distinctiveness.
Grapheme-phoneme consistency effects in reading (Coltheart 1978; Seidenberg & McClelland 1989).
Distinctiveness
Memory and searchability. Distinctive items are recalled better — the von Restorff effect — and rarer strings are easier to own in search and trademark. Measured by word frequency on the Zipf scale via wordfreq. Multi-word names average their parts: a phrase of common words reads common, not coined.
von Restorff (1933); Zipf frequency scale (wordfreq corpus, Speer & Schuler 2022).
Reported, not scored
Memorability— alliteration, reduplication, rhyme — is a small additive bonus rather than an averaged dimension. Phonetic repetition aids recall, but it’s zero for roughly 70% of real names, so it can’t carry the scale. Sticky names still get a lift.
Sound symbolism— the bouba/kiki and magnitude effects — is surfaced as a neutral reading, never graded. Front vowels read “small, bright, fast;” back vowels “large, dark, heavy.” Whether “sounds heavy” is good depends entirely on the brand. Grading it would smuggle a subjective judgment into an objective score.
Köhler (1929); Sapir (1929); Ramachandran & Hubbard (2001); Yorkston & Menon (2004, JCR); Klink (2000); Vanden Bergh et al. (1984).