This doesn't surprise me. The PageSpeed Insights site has a very atypical use case.<p>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.<p>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 <i>how</i> the page loads, so you can actually improve UI performance.