Literals are terminal nodes for the RDF data model, where is encoded actual data such as decimals ("12.5"ˆˆxsd:decimal), dates ("2017-01-26"ˆˆxsd:date), or any other information ("apple green"ˆˆex:color). RDF and SPARQL engines can test equality or compare literals whose datatype they know (which is the case for xsd:decimal and xsd:date). But when a datatype is unknown (like ex:color), then there is no direct means to “discover” its semantics. This paper tackles this prob- lem and show how : (i) data publishers can publish the definition of arbitrary custom datatypes on the Web, and (ii) generic RDF or SPARQL engines can discover them on-the-fly, and per- form operations uniformly. We discuss different possible solutions that leverage the Web of Data principles, and describe a simple one based on dereferencing and JavaScript, that can be used for arbitrarily complex custom datatypes.