Cybersickness remains a persistent challenge in immersive game design, yet most empirical research relies on laboratory-based trials with limited ecological validity. In contrast, this study analyzes over 1.3 million user-generated reviews from the Steam platform to examine how players experience, mitigate, and discuss cybersickness in first-person virtual reality (VR) games. Using large-scale computational text analysis, the study integrates topic modeling, lexicon-based sentiment analysis (VADER), and Random Forest classification to identify sickness-related narratives, self-reported mitigation strategies, and emotional framing in player reviews. The results show that cybersickness is more frequent in negative reviews; however, it also appears in a non-negligible subset of positive reviews. Reviews mentioning cybersickness tend to be longer and are associated with shorter playtimes at the time of review. Furthermore, surface-level textual and behavioral features, mainlyy word count and sentiment polarity reliably distinguish sickness-related reviews from other forms of player dissatisfaction. The analysis also reveals commonly reported mitigation strategies, including session control, seated play, and locomotion adjustments, while showing considerable inter-individual variability in their perceived effectiveness. Based on these results, the study derives design recommendations emphasizing player-centered comfort options and customizable locomotion systems. Overall, the results demonstrate the value of large-scale computational text analysis for capturing real-world VR experiences and informing inclusive, comfort-aware game design beyond laboratory settings.

Studies of user-generated video game reviews make sampling decisions that limit how far their findings can be generalized. How often these decisions are reported has not been measured. This abstract-level systematic mapping study screened 868 records from Scopus and Web of Science and coded the 245 eligible studies on eight dimensions from titles, abstracts and author keywords. Agreement with independent manual coding gave a median per-column Cohen’s kappa of 0.92 and a screening kappa of 0.93, and a second coder agreed with the final coding at a median of 0.89. Language scope was absent from 93.1% of abstracts, corpus size from 49.8% and the review platform from 41.2%; only 4.1% reported all three. Platform disclosure improved over time; language disclosure did not. In the 56 accessible full texts of a random sample of 75, nearly all missing platforms and corpus sizes were stated, but language scope remained unstated in 36. Among studies naming a source, 72.2% used Steam. Transformer-based methods rose (odds ratio 1.56 per year, p < 0.001), while dictionary-based sentiment analysis did not decline in absolute use. Syntheses in this area need full-text extraction, review-based findings largely describe one storefront, and language scope is underreported in abstracts and full texts alike.