Answers are grounded in this paper and cite their sources.
Artificial intelligence enables researchers to incorporate customer-generated text into firm-level analysis. Conventional review measures summarize evaluative direction, activity, and disagreement, but do not explicitly identify the subjects and experiences discussed. This study examines whether semantic representations of customer reviews provide incremental information about firm sales. Using Amazon reviews matched with publicly listed firms, we aggregate Sentence-BERT embeddings of review bodies into firm-quarter centroids and construct Customer Semantic Change (CSC) and latent semantic components. These measures augment a benchmark containing abnormal review tone, ratings, volume, dispersion, financial controls, and fixed effects. Adding CSC and 20 components increases adjusted R2 from 0.0492 to 0.0539. In a single chronological holdout, combining CSC with three training-selected components increases test R2 from 0.008076 to 0.014655. These results indicate modest incremental fit and holdout performance under the reported specifications. Exploratory component interpretations suggest variation in product domains, compatibility, and ownership experiences, but do not separate customer experiences from product composition. The study contributes a framework for constructing and evaluating firm-quarter semantic measures alongside conventional review indicators. The findings do not establish causal effects or practical decision benefits.
No takes yet. Share an insight, caveat, or question.
Ding et al. (2026) studied this question.
Synapse has enriched 8 closely related papers on similar clinical questions. Consider them for comparative context: