Due to technological limitations in early archaeological work, a large number of sites were documented only with vague textual descriptions, lacking precise geographic coordinates. This has posed significant challenges for subsequent archaeological investigations, site protection, and research. Traditional field survey methods are time−consuming and labor−intensive, making them unsuitable for large−scale or high−throughput site localization tasks. To address this issue, this study proposes an intelligent localization framework for archaeological sites based on large language models (LLMs). By designing tailored natural language prompts, the framework guides LLMs to automatically extract geographic descriptions—such as landmarks, directions, and distances—from online archaeological literature and records. It then integrates this information with high−resolution satellite imagery analysis to infer and calculate the precise geographic coordinates of the target site. This pipeline achieves full−process automation, covering text−based information extraction, remote sensing data retrieval, and spatial reasoning, thereby significantly improving the efficiency and intelligence level of site localization. Validation experiments conducted on the Laohushan Site, Sanxingdui Site, and Liao Zhongjing Site show that the deviation between the automatically inferred locations and actual surveyed coordinates is within one kilometer, with the best accuracy reaching approximately 10 meters—meeting the basic precision requirements of archaeological applications. Localization errors primarily stem from ambiguities in textual descriptions and uncertainties in image interpretation. Compared with traditional manual methods, this approach greatly reduces labor costs and offers enhanced scalability and application potential. The proposed method provides a novel technical pathway and tool support for the digitalization of archaeological information, site conservation, and further research.