Psycholinguistic norm estimation can use verbal instructions, labeled prototypes, or continuously rated training examples, but these procedures need not measure a construct identically. We compared zero-shot ratings from six large language models (LLMs), a prototype-defined BGE embedding axis, and supervised GloVe and BGE regression across 18 dimensions and a common 1,636-word holdout. Correspondence with human ratings was strongly construct-dependent. BGE ridge had the highest mean correlation (*r* =.734) and led the individual estimators on 13 dimensions, followed by the six-LLM mean (*r* =.685), GloVe (*r* =.652), and the highest-performing individual LLM, Gemini 3.6 Flash (*r* =.640). The six LLMs nevertheless showed highly concordant dimension profiles (Kendall's *W* =.945). The constrained BGE axis averaged *r* =.624 and had reliably higher correlations than Gemini on six dimensions and lower on ten, showing that the prototype-defined readout was competitive on selected constructs. Raw normalized mean absolute error strongly favored BGE, but out-of-fold linear recalibration narrowed the Gemini–BGE gap from.086 to.009 and eliminated the broad lexical-moderation interactions. Finally, the six-LLM mean and BGE retained complementary variation: their cross-fitted combination averaged *r* =.789, while an outcome-free fixed combination averaged *r* =.774. Automated norm estimates are therefore construct- and procedure-dependent measurements, not interchangeable readouts or entries in a single model ranking.