Abstract
Rating systems on accommodation platforms suffer from a familiar problem: nearly every listing displays a nearly perfect score, so the number that is supposed to separate good listings from bad ones barely varies. Whether the review text accumulating beneath those scores still carries usable information is an open question. I ask a dynamic version of it: does the text guests have already written predict where a listing's displayed rating moves next? Treating text and ratings as parallel channels that aggregate guest experience at different speeds, I construct a prespecified sentiment index from the complete review history of each listing in a two-wave panel of more than two hundred thousand listings across 34 U.S. markets. Because the broader project had explored these data before, I locked the model and its falsification checks in advance and reserved half of the markets, untouched, for a single confirmatory estimation. On those held-out markets, warmer past text predicts a small but precisely estimated upward movement of the displayed rating over the following year. Listings that received no new reviews show no such movement, the association survives host fixed effects, and no single market drives it. The results indicate that the platform's rating aggregation discards information its own review text retains. I discuss what this leading-indicator property implies for the design of reputation displays. The text index is a defined dictionary-based instrument that has not been validated against human judgment, and I state that boundary plainly.