Many ubiquitous learning (u-learning) applications heavily rely on the accurate retrieval of points of interest (POIs) in a geographical area, as it is at these locations that learning activities are proposed to students. In previous work, semantic technologies have been successfully employed to retrieve such POIs from Linked Open Data (LOD) datasets. However, the recent advancements in Large Language Models (LLMs), and their improved performance in processing factual data, arises as to whether u-learning applications could rely on LLMs to obtain exhaustive lists of POIs in a given geographical area. This poster paper provides empirical evidence about the current limitations of LLMs when carrying out this task in comparison with the capabilities of LOD datasets. More specifically, we compare the capabilities of a LOD dataset (Wikidata) and two LLMs (ChatGPT-o1 and DeekSeek-R1) for providing exhaustive lists of cultural heritage sites of three European cities and regions. Our results suggest that currently available LOD semantic datasets can complement state-of-the-art LLMs in terms of accuracy, completeness, consistency, and validity when gathering POIs for the design of u-learning situations.