We report the results of an experiment in which we use a Large Language Model (LLM) to generate a socio-economic dataset from unstructured online data. Specifically, we design a query to obtain the geolocation of the production facilities of a firm and loop this query over the set of top 2000 global firms. Such a dataset is not publicly available and is required to perform economic assessment of climate impacts. The LLM provided meaningful results for 75% of the firms, recovering a total of 35809 production sites. The rate of correct answers among those is of the order of 70%.