Connection hints
dns-prefetch, preconnect and preload links the document declares.
At a glance the headline numbers for Connection hints
dns-prefetch, preconnect and preload links the document declares.
Connection hints on the typical site: 0 dns-prefetch, 0 preconnect, 0 preload.
Distribution & median LCP site count and median LCP at each level of connection hints — dns prefetch
Passing LCP by connection hints — dns prefetch which level passes the LCP most often
Connection hints — dns prefetch 0. p75 1. p99 8. Spearman with LCP r = +0.32. At the low end (0): LCP 1.5s. At the high end (>p98): LCP 2.0s. computed
Distribution & median LCP site count and median LCP at each level of connection hints — preconnect
Passing LCP by connection hints — preconnect which level passes the LCP most often
Connection hints — preconnect 0. p75 1. p99 6. At the low end (0): LCP 1.5s. At the high end (>p98): LCP 1.8s. computed
Distribution & median LCP site count and median LCP at each level of connection hints — preload
Passing LCP by connection hints — preload which level passes the LCP most often
Connection hints — preload 0. p75 1. p99 18. At the low end (0): LCP 1.5s. At the high end (>p98): LCP 2.1s. computed
Why this matters for the Core Web Vitals, and where to start fixing it
Hints move network work earlier. dns-prefetch resolves a domain name before it is needed. preconnect goes further and completes the TCP and TLS setup too. preload fetches one specific file that the scanner would otherwise find late, like a background hero image or a critical font.
Hints are easy to overuse. Every preconnect spends sockets and CPU on a connection you might never use. A preload that is not critical steals bandwidth from something that is. A handful of deliberate hints beats a wall of them.
How does this affect the Core Web Vitals?
Connection hints correlate with the LCP. Preloads separates passing sites from failing sites the most. With few preloads, 82% of sites pass the LCP. With many, 63% do. The decline is gradual. There is no point where sites suddenly start failing.
Chrome field data from 94,910 sites, representing millions of real page loads. How we measured.