Why LLMs produce the same brand name

Ask four different large language models to name your startup and you will get four names that are functionally the same name. Not similar — interchangeable. The model is not failing to be creative; it is succeeding at exactly what it was trained to do: converge on the statistical centre of “what a good name looks like.” The desk calls this linguistic averaging, and it is the single biggest reason machine-generated names land as consensus mush.

The tell is recognisable in one glance. The saturated 2026 token list — Nova, Flux, Pulse, Nexus/Nexum, Atlas, Orbit, Vectra/Vanta-class, Synth, Neural, Axera — is a diagnostic, not a curse-word list. So is the [Adjective][Noun] compound family: NovaTech, BrightPath, SwiftFlow. Every output is short, punchy, vaguely technical, and ends in a vowel or an x. That shape is not the model’s “imagination”; it is the centroid of a billion brand names it has read. The averaging test is simple and brutal: would an off-the-shelf generator have produced this from a one-line brief? If yes, the name is slop — kill it.

What matters is how the failure is weighted. It is tempting to blame the model, and nearly everyone does. The desk’s breakdown puts the burden elsewhere: method ≈ 70% · brief ≈ 20% · model ≈ 10%. The model is the smallest contributor. The method — prompting a generator directly as if it were a namer — is the largest. When the brief is one line, the model fills the vacuum with the mode of the training distribution. When the method is “generate names,” the model does exactly that, and averaging is the only destination on the map.

The fix in pipeline v2 is a role reversal that is also a reading of the literature. AI is a profiler, never the generator. Generation comes from morphemes + phonesthemes — the smallest meaning- and sound-carrying units — with a required one-line morpheme story per candidate. A candidate is not a string of characters; it is a decision about which two pieces of meaning were joined and why. The engine generates from those pieces; the human panel discriminates. The model, demoted to profiler, is then genuinely useful: it cold-reads the crowd of candidates with zero context, and its disagreement across families is the signal.

The evidence base for this reversal is specific, not rhetorical. Tanj (2026-03) names the averaging phenomenon directly. Focus Lab × Placek (2026-01) formalise the Identify–Invent–Implement pipeline with 107 linguists and seven in-house LMs, placing AI explicitly in the profiling seat. NameStormers (2026-01) ships a name-slop filter. River+Wolf (2024/26) contribute prompt tonality, lateral paths, and VC-portfolio zeitgeist mining. Klink (JCR & JCPY, 2024) supplies the phonestheme evidence that justifies building from sound-meaning atoms in the first place. Vitevitch & Donoso (2012, PMC3289729) close the loop with the finding that higher phonotactic probability raises purchase intent — the empirically grounded reason a constructed name that feels like a real word can outperform an arbitrary one.

One operational consequence deserves its own line. The documented fix for both human groupthink and AI averaging is identical, because both collapse to the same attractor: disguise or diverge the brief. Run parallel lanes — one with the real brief, one “naming a competitor,” one naming an unrelated category. Windsurf came from the unrelated team. The words stop averaging the moment the generator is no longer allowed to know the destination.