I'm just guessing here, but I'm fairly sure that they use a model that updates dynamically as the "user" or victim changes his or her web browsing settings, and even when the user tries to hide. It easily sounds like some kind of Bayesian filtering going on, or some sort of Markov Chain or decision tree. That is to say that their model tracks the likelihood that you're the same unique user that reloads the page based on the information it can glean from you.
This makes it exceedingly hard to hide from such a filter, because in communicating with these sites, you are bound to reveal at least some information about yourself. And then the "likelihood-machine" does the rest by connecting the dots, even if you gave them "fewer dots."
It's also quite interesting - or perhaps chilling - to see how fingerprinting through NLP and other language tracking algorithms can also track just about any forum post you do, even if you're using a pseudonym.
This makes it exceedingly hard to hide from such a filter, because in communicating with these sites, you are bound to reveal at least some information about yourself. And then the "likelihood-machine" does the rest by connecting the dots, even if you gave them "fewer dots."
It's also quite interesting - or perhaps chilling - to see how fingerprinting through NLP and other language tracking algorithms can also track just about any forum post you do, even if you're using a pseudonym.