I saved the paper for reading later, so this may be already discussed in it, but enough information leaks through the Referer HTTP header when browsing the web in a traditional browser.

I've never inspected Recaptcha (on gstatic.com), but it does some degree of tracking, ostensibly to detect unusual usage patterns and pick who gets to help train Google's ML models with distorted street objects, and who's never shown the captcha window.