There's a thread running on the rdf-dev list at the moment about an Elevator Pitch for the Semantic Web. Two good contributions so far from Dan Brickley:

it's all about making Web-accessible data easier to automate by agreeing explicit names for meaningful relationships between things we describe in the Web. Some real world examples: "is a critical review of...", "works for the organisation whose home page is", "wrote the document available at", "is pictured in the image available at", "recorded by the dodgy rock band whose corporate home page is..." (much handwaving appropriate at this stage, indicative of invisible nodes and arcs being represented in confined elevator space...). The Web is about links; the Semantic Web is about the relationships implicit in those links

and from Andy Powell:

The Web is (primarily) about human-readable Web pages and the links between them - but where the linkage is very simple and generic ("this page is related to that page"). The Semantic Web is (primarily) about making machine-readable statements about all kinds of things (Web pages, organisations, people, concepts, products, etc.) and the links between them - and where explicit names are agreed for the relationships between things. For example, "this person works for that organisation", "this organisation has a home page at", "this person is author of the page at", "this person's phone number is", "this person is pictured in the image at", "this organisation sells that product", "this products costs that amount for that person", etc.