This doesn't surprise me. The PageSpeed Insights site has a very atypical use case. Anyway, the PSI 'score' is only a rough guide to help point you to problems

This doesn't surprise me. The PageSpeed Insights site has a very atypical use case. Anyway, the PSI 'score' is only a rough guide to help point you to problems. It can be kind of useful, in that if you get something like 97-100 then you can be pretty sure it's a very fast page... but it's also probably a page that doesn't do much beyond displaying something. With a more serious web app that does more interesting things, the score becomes a less useful indicator – it's often possible to make code changes that improve perceived performance but actually reduce the PSI score. For example, it's impossible for PSI to really know if inlining a script is going to generally improve or worsen performance, because it doesn't know your site's usage patterns, or which parts of your UI need to render first for a user to feel that the your UI 'is fast', whatever that means. UI performance is a subtler art than, say, algorithmic performance, and much less quantifiable. That's why many people prefer webpagetest.org, which comes with much smarter tools to record and analyse how the page loads, so you can actually improve UI performance.