The rapid growth of tourist-generated content demands scalable and reliable quality assessment methods. This study introduces an LLM-driven framework that combines parameter-efficient fine-tuning and prompt engineering to evaluate content quality accurately and interpretably. Applied to 484,930 reviews from MaFengWo, TripAdvisor, and Ctrip, the approach achieves superior performance (RMSE=0.3040, NDCG@100=0.500, BERTScore=77.95%) with 9 × higher efficiency. Spatial-temporal-semantic analyses reveal platform-specific quality patterns: MaFengWo exhibits prominent spatial centrality and stable temporal cointegration; TripAdvisor demonstrates simplified core-periphery structures with high volatility; Ctrip presents dynamic multicentricity particularly in Shanghai. Two domestic platforms, MaFengWo and Ctrip, expose systematic deficiency on the theme ‘ Decision-making Plan ’ (92.9∼96.4% lacking operational suggestions), while international TripAdvisor emphasizes ‘ Practical Information ’ and ‘ Consumption Activity ’ but 40.76% neglects original viewpoints. Heterogeneous network analysis identifies the behavioral signatures of high-reliability user—preference attachment, quality stability, and profile homogeneity. This work bridges theoretical rigor with operational scalability, demonstrating the potential of LLMs in content governance for digital tourism.