Platform Visibility and Spatial Representation in Industrial Heritage Regeneration

User-generated content (UGC) is increasingly used to inform heritage planning, yet platform visibility may not reflect the spatial elements that require management attention. Using Hangzhou Steel Park as a case study, this study compares 3521 online posts (containing 18,827 images in total) from four platforms with 2689 protocol-based on-site images. A multimodal large language model converted images into structured descriptions, which were analysed with BERTopic. Three linked models compared online and on-site visual topics, tested robustness through downsampling, and examined post-level image–text integration. Online UGC concentrated on blast furnaces, night lighting, check-ins, and visitor experience, whereas the on-site reference provided broader coverage of renovated buildings, vegetation, lawns, corridors, and other spatial elements. The leading blast-furnace topic remained dominant after downsampling, suggesting that corpus size alone is unlikely to explain its prominence. Post-level discourse analysis remained strongly experience-oriented. These findings show that UGC is valuable for monitoring visitor attention but should be complemented by systematic on-site evidence for spatial planning and heritage management.