Ultimately, you can’t assess anonymity without knowing how your adversary is trying to identify you.
Personally, I feel the below is a more accurate way to say this:
Ultimately, you
can’t assess[can’t assess to a high degree of confidence] anonymity without knowing how your adversary is trying to identify you.
No test will be a perfect reflection of reality (this is especially true of things like browser fingerprinting which is probably still in its adolescence and where there is a lot that is hard to know). A test can be at best a reflection of the test conditions and the input data. But dismissing testing or modeling as a concept because it falls well short of perfect seems like throwing the baby out with the bath water.
Any test like this should be assumed to be at best a rough indication of uniqueness. Getting a result of 1 in 775 doesn’t tell you that your browser actually looks identical to 1 in 775 browsers in the real world against all types of naive or advanced fingerprinting, but you can probably reasonably infer that the browser config that gets you that 1 in 775 test result is less unique than the browser config with a result of 1 in 9,500,000 result. And you can dig into the raw data see what changes made the biggest difference. Basically my point is that so long as you understand the limitations and fuzziness of this sort of testing (and understand it is a reflection of selection bias/sampling), and don’t put too much faith or weight on it, I think it has value.