Watson starts off by searching its massive database of sources for stuff that might be pertinent to the question. Next, it searches through all of the search results and turns them into candidate answers. For example, if one of the search results is an article, Watson might pick the title of the article as a possible answer. After finding all of these candidate answers, it proceeds to iteratively score them to determine which one is best.
The scoring process is very complicated, and involves finding supporting evidence for each answer, and then combining many different scoring algorithms to determine which candidate answer is the best. You can read a more detailed (but still very conceptual) overview here, by the creators of Watson.