In the Resource Description Framework, literals are composed of a UNICODE string (the lexical form), a datatype IRI, and optionally, when the datatype IRI is rdf:langString, a language tag. Any IRI can take the place of a datatype IRI, but the specification only defines the precise meaning of a literal when the datatype IRI is among a predefined subset. Custom datatypes have reported use on the Web of Data, and show some advantages in representing some classical structures. Yet, their support by RDF processors is rare and implementation specific. In this paper, we first present the minimal set of functions that should be defined in order to make a custom datatype usable in query answering and reasoning. Based on this, we discuss solutions that would enable: (i) data publishers to publish the definition of arbitrary custom datatypes on theWeb, and (ii) generic RDF processor or SPARQL query engine to discover custom datatypes on-the-fly, and to perform operations on them accordingly. Finally, we detail a concrete solution that targets arbitrarily complex custom datatypes, we overview its implementation in Jena and ARQ, and we report the results of an experiment on a real world DBpedia use case.