The question was “How many users browse the main Internet Archive site with a default font size other than the common value of 16 pixels?” By knowing this, we would determine how many users would be affected by sizing with relative units like rems/ems.
Using the methodology I describe below, we found that the answer is 3.08% of our users.
So if you set type in pixels, and your traffic is anything like the Internet Archive's, 3% of your users won't have their explicitly-asked-for font-size alteration accommodated.
It's true. I made two little reduced test cases. First I left the default medium font-size preference on and set one with pixels and one with ems and sized them to match. Then bumped up the preference to large, and only the ems changed, the pixel-set one stays the same.
As ever, there is more to think about. How does page zooming factor in? How annoyed are those 3%? How annoyed are the 3% on other sites that use pixels but set type pretty large anyway? How much do we care about people who use the preference to size type down?