Between a request departing and HTML arriving, the connection sits idle; resource hints exist to spend that idleness. Used precisely they remove round trips from the critical path; sprayed indiscriminately they congest it, and the difference is entirely in the discipline.
The hint taxonomy, with contracts
preconnect opens a connection (DNS, TCP, TLS) to an origin you will definitely use: cheap, bounded, do it for your two or three critical third-party origins. preload declares a specific resource the current page needs urgently but discovers late, fonts behind CSS, the LCP hero behind JS: it competes for current-page bandwidth, so every preload taxes something else. prefetch requests likely next-page resources at idle priority: a bet on navigation, spending spare bandwidth now to save round trips later. The taxonomy matters because the failure mode of each is the success mode of another.
103 Early Hints, the server’s turn
Early Hints lets the server (in practice, the edge) respond before the response: a 103 carrying Link headers fires while origin still thinks, so the browser preconnects and preloads during the server-wait segment our waterfall article dissected. CDNs are the natural emitters, they know the hints from config or from observing past responses, and support has matured across the major platforms. For origin-heavy TTFBs it recovers hundreds of otherwise-dead milliseconds, which is why it is the most consequential hint most estates still have not enabled.
The speculation frontier extends this logic: speculation rules push prefetching to full prerendering of likely next pages, and edge platforms increasingly participate, serving speculated navigations from cache with the page already constructed. The economics are the interesting part: speculation converts spare bandwidth and edge capacity into perceived-instant navigation, a trade that improves as delivery gets cheaper, which our pricing series documents it relentlessly doing. The discipline transfers unchanged: speculate from measured navigation probabilities, cap by device and network class, and audit the hit rate, because a speculation that is usually wrong is just a bandwidth leak with good intentions.
The misuse patterns
Preloading everything (the priority article’s lesson inverted: undifferentiated urgency is no urgency), preloading fonts without crossorigin (double-fetch bug, still everywhere), prefetching heavy resources on metered mobile connections, hint lists that fossilize while the pages evolve (preloads for assets no longer used, silently wasting bandwidth on every view), and third-party tag managers injecting their own hint spam. Every hint is a standing claim about your critical path; stale claims are anti-performance.
In practice
Derive hints from measurement: the waterfall shows what is discovered late (preload candidates), analytics shows where users navigate next (prefetch candidates), and both lists get revisited quarterly with the same discipline as the cache-key audit. Enable Early Hints at the edge for your slow-origin routes and measure the LCP delta in RUM cohorts. Hints are the rare optimization with instant rollback and per-page measurability; there is no excuse for cargo-culting them.
Hint audits ship inside our performance engagements: derived-from-measurement lists, Early Hints enablement, RUM-verified deltas.