This was actually my primary role at Google from 2006 to 2010. One of my first test cases was a certain date range of the Wall Street Journal's archives of the

This was actually my primary role at Google from 2006 to 2010. One of my first test cases was a certain date range of the Wall Street Journal's archives of their Chinese language pages, where all of the actual text was in a JavaScript string literal, and before my changes, Google thought all of these pages had identical content... just the navigation boilerplate. Since the WSJ didn't do this for its English language pages, my best guess is that they weren't trying to hide content from search engines, but rather trying to work around some old browser bug that incorrectly rendered (or made ugly) Chinese text, but somehow rendering text via JavaScript avoided the bug. The really interesting parts were (1) trying to make sure that rendering was deterministic (so that identical pages always looked identical to Google for duplicate elimination purposes) (2) detecting when we deviated significantly from real browser behavior (so we didn't generate too many nonsense URLs for the crawler or too many bogus redirects), and (3) making the emulated browser look a bit like IE and Firefox (and later Chrome) at the some time, so we didn't get tons of pages that said "come back using IE" er "please download Firefox". I ended up modifying SpiderMonkey's bytecode dispatch to help detect when the simulated browser had gone off into the weeds and was likely generating nonsense. I went through a lot of trouble figuring out the order that different JavaScript events were fired off in IE, FireFox, and Chrome. It turns out that some pages actually fire off events in different orders between a freshly loaded page and a page if you hit the refresh button. (This is when I learned about holding down shift while hitting the browser's reload button to make it act like it was a fresh page fetch.) At some point, some SEO figured out that random() was always returning 0.5. I'm not sure if anyone figured out that JavaScript always saw the date as sometime in the Summer of 2006, but I presume that has changed. I hope they now set the random seed and the date using a keyed cryptographic hash of all of the loaded javascript and page text, so it's deterministic but very difficult to game. (You can make the date determistic for a month and dates of different pages jump forward at different times by adding an HMAC of page content (mod number of seconds in a month) to the current time, rounding down that time to a month boundary, and then subtracting back the value you added earlier. This prevents excessive index churn from switching all dates at once, and yet gives each page a unique date.)