Watson gets the keywords from the document and searchs for them, showing the results in a side panel
The results are surprisingly relevant and often very useful
It saves you the hassle of carefully thinking about relevant search queries. The document you are editing becomes one big search query and the results (a few hundred links, usually) come almost instantly. So far I have only tried it on existing documents, but it should be interesting to try to use it while writing a new document, augmenting my knowledge and memory with the power of the Internet. :)
Obviously, in the future automatic context-sensitive search will become much more common. Computers will be presenting us with hints, ideas, reminders, based on what we are doing, reading, talking about. The amount of available data, including our photos, e-mails, posts, conversations, etc. is growing very fast, as well as the amount of relevant information online. In the past it made sense to perform research (originally a literature search, later a google search) before writing a paper or preparing a presentation. But increasingly this becomes redundant and useless, as the amount of useful and relevant information often far exceeds our need in it.
So it makes sense to pay less attention to getting all the information. The computer should just give us something and we will use it. This is how our memory works - we are not trying to produce a perfect recollection of our past actions or of school knowledge that we acquired - we simply use what our subconsciousness gives us. A good illustration to the expected future of search is in Charles Stross's Accelerando, where search agents are running on human exocortices.