Large language models (LLMs) have demonstrated remarkable capabilities in text generation and contextual reasoning across various domains. However, their performance in urban data applications remains unsatisfactory. Existing studies primarily use LLMs as tools to process urban data, focusing on feature extraction or prediction rather than improving the inherent ability of LLMs to comprehend cities for supporting diverse tasks across different urban contexts. To address this gap, we propose LLM-Urban+, investigating the capability of LLMs in understanding urban areas by injecting urban knowledge. LLM-Urban+ leverages Points of Interest (POI) and human mobility data to transform structured urban data into compressed textual descriptions and spatial embeddings, enabling LLMs to establish urban area cognition by capturing semantics and spatial correlations, ultimately extending their contextual reasoning capabilities to urban scenarios in a zero-shot setting. The approach is evaluated through the generation of functional textual descriptions and generalizable representations for urban areas, supporting diverse tasks such as functional assessment, similarity analysis, and open-ended Q&A. We validate LLM-Urban+ in New York City and Chicago, demonstrating its effectiveness in enhancing LLMs’ understanding of urban semantics and enabling various downstream applications. The results highlight the potential of LLM-Urban+ to advance urban data analysis and broaden the scope of LLM capabilities.