You use a bunch of subdomains -- a.example.com, b.example.com, etc. -- each configured so that a particular URL (call it the 'set' URL) sends an HSTS header. A different URL (the 'get' URL) doesn't.
You generate an ID for the user, and encode it as a bit pattern using the subdomains to indicate positions of '1' digits. Say your ID is 101001 -- you serve a page which includes images loaded from the 'set' URLs for subdomains a, c, and f. On later page loads, you serve a page including images loaded from the 'get' URLs of every subdomain, and you pay attention to which ones are requested via HTTPS. Since the 'set' URL sent an HSTS header, subdomains a, c, and f get requested over HTTPS, and now you reconstruct the ID from that: 101001.
I feel like this changes all the time; I was recently surprised to discover 'TLS Client Channel ID' (my nomenclature is a bit fuzzy - an RFC for automatic client certs "for security") and would love to learn more about the extent of its current implementation in Chrome.
https://news.ycombinator.com/item?id=15753648
>londons_explore: In Chrome, it also uses the TLS Client Channel ID, which is a persistent unique identifier established between a browser and a server which (on capable platforms) is derived from a key stored in a hardware security module, making it hard to steal. Ie. if you clone the hard drive of a computer, when you use the clone, Google will know you are a suspicious person, even though you have all the right cookies.
https://en.wikipedia.org/wiki/Transport_Layer_Security_Chann...