According to a post on DEV Community, originally from the NEXUPIRA engineering blog, a SvelteKit marketing site served in English, Traditional Chinese and Japanese had weak mobile Lighthouse results on its non-English pages. The /zh page scored 86 with a Largest Contentful Paint (LCP, the time the biggest visible element appears) of 3.84 s. The /ja page scored 77 with a Cumulative Layout Shift (CLS, a measure of page jumping) of 0.179.
The team first suspected large CJK web fonts. The main cause turned out to be a bug. During early hydration, when the page's JavaScript takes over the server-rendered HTML, the locale store had no page data yet and answered 'en'. That briefly set the html lang attribute to English, which switched fonts and headline size through CSS, then switched back. The fix was to read the language from the URL when page data was missing. According to the post, that alone brought /ja CLS to zero.
Other changes followed. The CJK font stylesheet, 135 KiB, had been loaded with a preload at top priority, so it competed with the CSS needed for first paint. The authors self-hosted the Latin fonts and moved the font stylesheets to load after first paint, with display=optional for CJK fonts. A localStorage flag lets repeat visitors get cached fonts on first paint. They also dropped about 50 modulepreloads and delayed hydration until after first paint.
Two smaller changes helped too. Sixteen logo files became WebP at display size, falling from about 438 KB to about 37 KB. Removing article bodies and a Markdown parser from the home bundle cut home route JS from 470 KB to 409 KB. Measured on a local production build, the medians moved to 100. A single production spot check gave 97 on both pages, with LCP of 1.3 s and CLS of 0.
Why it matters
The post shows that a slow LCP is not always a download problem. On /zh, 3.39 s of the 3.84 s LCP was render delay, meaning the browser had the HTML but was busy with other work before painting. Checking the LCP breakdown first points you at the real cause.
It also warns that defaults can differ between server and browser. On any multilingual site, a value that briefly says the wrong language can change fonts and shift layout. The authors also note that lab numbers are not real-user data, and one fix that seemed simpler, inlining font rules, made first-visit LCP worse when measured.
Some limits apply. The blocking behaviour for repeat visits works only in Chromium, and the CJK fonts still come from Google Fonts.
Source: DEV Community. This brief summarises their reporting; read the original for the full detail.