Automatic subject indexing is increasingly implemented using Large Language Models, whose performance, however fluctuates sharply across disciplines. A two-dimensional framework—breadth (surface term coverage) and depth—is introduced to to quantify the structural quality of disciplinary controlled vocabularies. Based on the German GND authority file across ten disciplines, several indicators are computed, Spearman correlation analysis reveals that higher breadth is significantly improves indexing accuracy, while excessive depth correlates negatively with performance. These findings confirm that performance hinges not on maximizing either dimension, but on achieving a breadth–depth balance. This work provides the empirical evidence that breadth–depth balance, rather than maximizing either alone, governs cross-disciplinary indexing performance, offering actionable guidance for building equitable, discipline-sensitive, LLM-compatible knowledge bases.