This has to do with the way that the way the algorithm works. When you click like, you like a page. Then, a server-side script on webperflab's server contacts Facebook's opengraph, gets a list of users, looks up information on the last user that liked the page, makes that user "unlike" the page (the page can do that, it's like kicking a user out of a group), and then returns information about that user. If, for example, another user likes the page before the script can return information about you, it will return information about the other user. The demo could probably fix this issue by having multiple groups.