That is a good idea! Like Google advertising, where the ads are selected based on the textual context -- words in your search string or email or website -- but instead of ads as the output, short video clips. For long text blocks, run the videos alongside a scrolling reader such that the video and text stay synchronized. The word "respect" prompts an Aretha Franklin clip, maybe. Appeal to the [people who think they are] multitaskers! You'll have nerd users searching for which Shakespearean passages lead to the most amusing clip playlist!

If the video bandwidth is all Youtube's, and the UI is Javascript, your ongoing cost is mostly the server time to run the text against your corpus and spit out a playlist.

I love when startup ideas spring from a misinterpretation of how someone else summarizes their product.