This paper develops register sub-corpora for the Web-crawled Finnish Internet Parsebank. Currently, all the documents belonging to different registers, such as news and user manuals, have an equal status in this corpus. Detecting the text register would be useful for both NLP and linguistics (Giesbrecht and Evert, 2009) (Webber, 2009) (Sinclair, 1996) (Egbert et al., 2015). We assemble the subcorpora by first naively deducing four register classes from the Parsebank document URLs and then developing a classifier based on these, to detect registers also for the rest of the documents. The results show that the naive method of deducing the register is efficient and that the classification can be done sufficiently reliably. The analysis of the prediction errors however indicates that texts sharing similar communicative purposes but belonging to different registers, such as news and blogs informing the reader, share similar linguistic characteristics. This attests of the well-known difficulty to define the notion of registers for practical uses. Finally, as a significant improvement to its usability, we release two sets of sub-corpus collections for the Parsebank. The A collection consists of two million documents classified to blogs, forum discussions, encyclopedia articles and news with a naive classification precision of >90%, and the B collection four million documents with a precision of >80%.