The issue here is that only a tiny percentage of the information transmitted is of interest to someone doing surveilance.

No one cares about the data of the YouTube videos you watch, the content and images and JavaScript and CSS and what-have-you of the web pages you visit, the content of the files you download etc. Examples of what would be of interest, and only comprises a tiny portion of all web traffic, is

* All URLs you request (will include your search history)

* The contents and headers (and maybe attachments) of your e-mails

* Your online IM conversations

* _Concievably_ the video from your video calls, but this is too much data to store

* Arguably any data transmitted through a text form online

Again, the interesting data is only a tiny portion of all web traffic, because the rest is transmitted repeatedly and can be recreated on demand. Much of the interesting stuff is, in fact, in the "metadata".