On July 31, PageSpeed Insights gave our own website a 61 performance score on desktop. Yes, the website of the studio that sells speed. We could have kept quiet; we'd rather tell it with numbers, because this is what our maintenance looks like from the inside: you measure, you find the real cause, and you fix it without sacrificing the design. A few hours later, the same tool read 100.
Three figures told the story: 910 milliseconds of main-thread blocking on desktop, 18 seconds of accumulated CPU work, and a headline that took 3.9 seconds to settle on mobile. The page looked fine; underneath, it was overspending.
Our hero's aurora runs on the graphics card and costs almost nothing. But its code asked the browser "how big am I?" on every frame, and that question forces the whole page's layout to be recalculated 30 times per second. The effect wasn't expensive; the question was.
90% of the problemOn a 4G connection, Google Analytics' 130 KB downloaded in parallel with the headline's font. The result: the main text waited its turn behind a script nobody sees. The visitor paid for that queue in seconds.
The mobile LCPPageSpeed's servers have no graphics card: they emulate WebGL on the CPU, where our effect cost hundreds of milliseconds no real visitor ever pays. Knowing where you're measured, and what hardware the measurer runs, is part of the craft too.
The curious findNo effect was removed and the design wasn't touched. Every piece now runs where it belongs.
It now measures itself once, and again only if the window truly changes. And it learned to bow out gracefully: if it detects a machine with no graphics card (a measurement bot, a virtual machine), it doesn't turn on. Our philosophy always said animation layers are optional and separable; now it keeps that promise even with bots.
Analytics now waits 3 seconds after the page finishes loading: same measurement, same consent, but out of the content's way. And the fonts were subset to the characters the site actually uses: 20% less weight without losing a single accent, arrow, or typographic quote.
| Metric | Before | After |
|---|---|---|
| Desktop performance | 61 | 100 |
| Mobile performance | 86 | 100 |
| Main-thread blocking (desktop) | 910 ms | 0 ms |
| Headline settled on mobile (LCP) | 3.9 s | 1.7 s |
| Mobile Speed Index | 4.4 s | 0.9 s |
| Layout shifts (CLS) | 0 | 0 |
None of this asks for an act of faith: open PageSpeed Insights, type index01.net, and run the test. The numbers in this note are from the early hours of August 1, 2026 and vary a couple of points between runs, like every honest measurement. That's the difference between a case study and a boast: the case can be repeated.
A site that's fast on launch day can stop being fast six months later without anyone touching it: scripts pile up, meters change, fonts grow. That's why our maintenance includes auditing regularly and fixing the real cause, not the symptom. This note is, literally, that service applied to ourselves.
Run PageSpeed on your website. If you don't like the number, tell us what it read and we'll tell you, no strings attached, where the cause usually hides.
Let's talk →