Linguistic and emotional characteristics of LLM-generated marketing texts: a comparative analysis using Empath

Large language models (LLMs) such as GPT-5.1, Gemini, and Perplexity are increasingly used to generate marketing content. This study analyzes 90 marketing texts generated from 30 standardized prompts submitted to the three aforementioned LLMs under deterministic conditions. Using Empath, an open-source lexical analysis tool, we extracted 194 semantic categories and identified 15 focal variables representing emotion, persuasion, achievement, and business-relevant dimensions. Prior to inferential testing, floor-effect checks identified three categories (negative emotion, joy, and fear) with insufficient variance for analysis. Across the remaining 12 testable categories, MANOVA indicated no significant multivariate differences. Univariate ANOVA with Bonferroni and Benjamini–Hochberg corrections revealed no statistically significant differences after correction for multiple comparisons. A full exploratory scan of all 194 Empath categories confirmed this null result. ANCOVA controlling for significant word count differences across models further weakened observed trends. Cross-prompt stability analysis revealed that, in this dataset, Gemini produced the most consistent linguistic style across prompt types, while Perplexity exhibited the most variability. These findings suggest that, in this sample, contemporary LLMs produce broadly convergent marketing language under deterministic prompting conditions, with implications for model selection, prompt engineering, and automated marketing communication.