Exploring text collections through named entities remains a very common need for scholars. Despite the recent advances of Named Entity Recognition (NER) systems, more efficient and easier to use, the task remains problematic when the data is not born digital and thus more prone to Optical Character Recognition errors. In this paper, we investigate the real influence of noise on the extraction of locations in a collection of ten books in French exhibiting different levels of difficulties for NER systems (digitization quality, complexity of layout, variation in language). We compare the results of various systems on the "clean" version of the documents and on different OCRed versions. We show that NER systems do not yield many more errors in noisy documents, most of the errors being already there on a reference version, and in some cases NER performs better on noisy versions. According to our results, the main problem is rare entities (especially hapax) which are more likely to disappear from of the output.