What Speed Brain does
Every page visit sets off a chain of request and response cycles. Network transmission itself is no longer the main culprit: data crossing the 11,000 km round trip between New York and London on Cloudflare takes about 76 milliseconds, faster than a blink. The remaining delay comes from processing requests and responses and applying configuration. Speed Brain attacks that delay by anticipating the visitor's next navigation rather than waiting for it.
The feature is built on the Speculation Rules API. Speed Brain prefetches the content of likely next navigations into the browser cache, so that when the actual navigation occurs, the page loads almost instantly.
The initial implementation is deliberately cautious: it prefetches static content for the next page once the user starts a touch or click event. Through the fourth quarter of 2024 and into 2025, Cloudflare plans more aggressive speculation models, including prerendering — fetching the page before navigation and rendering it completely. The eventual goal is a static site with no latency and no configuration, with the browser cooperating to load pages as fast as possible.
A clothing ecommerce site illustrates the model. Global request logs show that a visitor viewing "Mens > Clothes" is highly likely to click "Shirts", so static content such as images can be delivered before the click. In recent lab testing of the aggressive loading model, this approach produced up to a 75% reduction in Largest Contentful Paint (LCP), the time for the largest visible element — an image, video or text block — to load and render.
Availability and measured results
Speed Brain is available on all plan types at no cost. It can be switched on per website from the dashboard or via the API. It is already on by default for all free domains, where successful prefetches show a 45% reduction in LCP. Pro, Business and Enterprise domains must enable it manually.
Cloudflare also recommends enabling Real User Measurements (RUM) from the dashboard to observe page performance. Enabling RUM additionally feeds improved, customized prefetching and prerendering rules for a domain in the near future.
Why navigation still costs time
Loading a page in a browser requires a series of request and response cycles. The browser first establishes a secure connection with the server, then sends an HTTP request for the base document; the server processes it, builds the HTML and returns it in the response.

On receiving the HTML, the browser parses it immediately. Parsing surfaces references to external resources — CSS, JavaScript, images and fonts — which the browser fetches with additional HTTP requests. Resources already in the browser cache can be read locally, cutting network latency.
As HTML, CSS and JavaScript are processed, the rendering engine displays content. Once the visual elements are on screen, an interaction such as a link click restarts much of this work to fetch the next page. This cycle repeats throughout a session: each navigation fetches and renders new or uncached resources, adding delay before the new page is fully loaded.
On the shopping site, moving from the homepage to the men's section, then clothing, then shirts, accumulates that retrieval time on every step — time that can contribute to a shopper abandoning the site before completing a transaction. Having prefetched and prerendered pages already in the browser when those links are clicked would remove most of the network latency impact and produce a smoother experience.
Why earlier prefetching attempts fell short
Speculative fetching is not a new idea, and browsers and site owners have tried it before. The recurring problems were a lack of data about real user behavior and few APIs for acting on it.
An early example was the Google Web Accelerator, a client-side application introduced in 2005 to speed up browsing for broadband users. It was innovative but short-lived, shelved over privacy and compatibility issues. Predictive prefetching at the time had no data or API support for capturing behavior around sensitive actions such as deletions or purchases.
Manual resource hints and static lists
The usual mechanism for prefetching has been the <link rel="prefetch"> attribute, part of the Resource Hints family. Developers had to add it to each page for each resource the browser should fetch ahead of time and hold in memory. That manual approach is laborious, and developers rarely had insight into which resources were worth hinting, so hint quality suffered.
Cloudflare's URL prefetching feature, available since 2015, works differently: instead of filling the browser cache, it pulls a static list of resources into the CDN cache during idle time or favorable network conditions. Customers still have to choose which resources are good candidates on every page they own. A misconfigured static prefetch can hurt rather than help page load time.
Server Push and its blind spots
HTTP/2 server push aimed to cut latency by sending resources before the client requested them, avoiding extra round trips. But servers had no context about what the browser already had cached, so pushing resources wasted bandwidth and could delay critical assets such as base HTML and CSS through race conditions during rendering. Cache digests, the proposed fix that would have told servers what clients had cached, never saw widespread implementation. Chrome removed Server Push support in October 2022, and Firefox followed in September 2024.
What Early Hints solves, and what it doesn't
Early Hints, specified in 2017, was largely unused until Cloudflare and partners deployed it with browsers and key customers in 2022. The server sends a 103 Early Hints HTTP status code listing key page assets so the browser can start loading them while the main response is prepared. The browser prioritizes based on what it needs, and assets already cached are not redundantly fetched.
Early Hints does not yet adapt to user behavior or dynamic page conditions, and its main use is preloading specific assets rather than whole pages, particularly when the server takes a long time to produce HTML.
Cloudflare points to Argo Smart Routing, Smart Tiered Cache, and Smart Placement as earlier performance work built around user behavior. The company frames its new speculative system as a further step toward an adaptable framework for serving content quickly.
Inside the implementation
Speed Brain moves resource prediction to the client. Rather than sitting idle until a user asks for the next page, the browser reads the signals of the current interaction and requests the next document before the click completes. When the feature is switched on, Cloudflare adds a Speculation-Rules HTTP response header whose value is a URL pointing to an opinionated rules configuration. The browser fetches and applies that configuration to prefetch candidate navigations. The first visit to a site gains nothing from this; later navigations within the same site can.

The request path, step by step
- For a page with Speed Brain enabled, the Cloudflare server returns the extra
Speculation-Rulesresponse header alongside the page. - As the browser parses the header, it retrieves the rules configuration and loads it with the page.
- The configuration tells the browser when to prefetch the next likely Cloudflare page, driven by how the visitor is engaging with the current page.
- A user action — a mouse-down on a link, for instance — triggers the rules, and the browser issues a prefetch request carrying the
sec-purpose: prefetchHTTP request header. - The server parses the header to recognize the request as a prefetch. If the content is in our cache it is returned; otherwise the request is denied with a 503 status code. This prevents unsafe side effects from reaching origins or Cloudflare Workers that have no awareness of prefetching. Only content held in the cache is served.
- On success the browser stores the prefetched content in memory, and a later navigation to that page renders straight from the browser cache.
Design choices: document rules and eagerness
The two failure modes that limited earlier speculation efforts — stale prefetch configuration and incorrect prefetching — drove the rule design.
Static prefetch configurations fall out of date as sites change, producing prefetches that are inefficient or useless. The rel=prefetch attribute and static CDN prefetch URL sets both require developers to hand-maintain a list of prefetchable URLs per page, usually built on intuition rather than real navigation data. Speed Brain instead uses the API's document rules, expressed as "source": "document" plus a "where" key, so prefetch candidates come from the live structure of the page and no predefined list exists to go stale. Adding "relative_to": "document" to the where clause restricts prefetching to same-site links, which also means no cross-origin prefetches and therefore no privacy implications from following users around the web.
Incorrect prefetching is the more subtle problem. A prefetch is an ordinary request apart from the sec-purpose header, so it carries the same client, network and server cost, yet the response may never be used. Prefetch accuracy — the share of prefetched pages the user actually views — therefore matters most, especially on metered mobile networks or in low-bandwidth environments.
The configuration this first release ships:
{
"prefetch": [{
"source": "document",
"where": {
"and": [
{ "href_matches": "/*", "relative_to": "document" },
]
},
"eagerness": "conservative"
}]
}
The eagerness setting controls how aggressively the browser acts. immediate starts prefetching as soon as the browser sees the rule, generally on page load. eager matches that behavior but also waits on slight interaction events such as the cursor heading toward a link (coming soon). moderate fires after the pointer rests on a link for more than 200 milliseconds, or on the pointerdown event if that comes first, and on mobile where no hover event exists. conservative prefetches on pointer or touch down on the link.
Speed Brain starts at conservative, trading away the gains of more aggressive settings to keep incorrect prefetching rare. More dynamic eagerness settings for sites that would benefit from a more liberal value are planned, and the rules will be extended to cover prerendering.
Cache-only prefetching
Prefetch requests are answered only if the static content is already in Cloudflare's CDN cache; anything else gets a 503. Serving from cache sidesteps any question of cache eligibility: a page that isn't cacheable shouldn't land in the browser cache either, because that can increase origin load and cause unintended effects. Prefetching a logout page, for example, could sign a user out before they finish what they were doing, and stateful prefetching or prerendering can alter server state for actions the client never confirmed. Restricting prefetch to cached pages keeps those outcomes off the table.
Measured on Cloudflare's documentation
The baseline conservative deployment was measured across all pages of Cloudflare's developer documentation in early July 2024. The correct content was prefetched — a page the user did in fact navigate to — 94% of the time, while navigation improved by cutting LCP at the p75 quantile by 40%, with no unintended side effects observed.
The prefetched response is delivered from a network spanning more than 330 cities and within 50 milliseconds of 95% of the Internet-connected population, so cache hits arrive with almost no latency. Because that position sits in front of the customer's origin, Speed Brain is enabled by flipping a switch, with no origin server configuration changes. The first version is now live.

Browser support and operational caveats
Speed Brain is built on the Speculation Rules API, an emerging web standard. As of September 2024, browser support for it is limited to Chromium-based browsers at version 121 or later, including Google Chrome and Microsoft Edge. Wider adoption by other vendors depends on the web community reaching consensus on standardizing the API.
Because prefetching a dynamic page risks delivering stale data to visitors and adding load to the origin, Speed Brain acts only on non-dynamic pages that are cached on Cloudflare's network; it has no effect on the loading of dynamic pages. To get the most out of the feature, Cloudflare recommends cache rules that make all static content on a site — HTML especially — eligible for caching.
Why prefetch requests log 503s
If the browser receives a 503 status code in response to a speculative prefetch (identifiable by the sec-purpose: prefetch header), it abandons the attempt. The error in the console can look alarming, but the cancellation is harmless. The specification guidance recommends a 503 as the simplest way for the browser to safely discard a speculative request, and early beta testers raised the issue with Cloudflare. The company says it is discussing with Chrome a dedicated non-error response code, which it believes would be more appropriate and less confusing. Until then, 503 logs tied to Speed Brain prefetches are harmless; operators whose tooling makes those requests hard to ignore can temporarily disable the feature.
When a site runs its own custom Speculation Rules alongside Speed Brain, both rule sets operate at once. Cloudflare's guardrails restrict its speculation rules to cacheable pages, a limit that may surprise sites with existing implementations, and disabling one of the two is the suggested remedy. A Speculation-Rules header from the origin is not overridden, so conflict potential applies mainly to predefined inline rules.
Measuring the effect
Cloudflare recommends running Speed Brain alongside its RUM performance measurement tool, which reports on loading (how long content takes to become available), interactivity (responsiveness to user input) and visual stability (how much the page shifts during load). With RUM on, the Web Analytics dashboard shows how Speed Brain reduces core web vitals metrics such as Largest Contentful Paint (LCP) and load time.

Example RUM dashboard for a website with a high amount of prefetchable content that enabled Speed Brain around September 16.
Results from the rollout
The feature is on by default for all free plans. Cloudflare reports the following.
Domains and LCP
Tens of millions of domains use Speed Brain today. Measuring LCP at the 75th percentile (p75) across these sites shows an improvement of 40% to 50%, averaging around 45%. The comparison was made by contrasting navigational prefetches with normal, non-prefetched page loads on the same set of domains.

Latency saved per prefetch
Before Speed Brain, free websites on Cloudflare had a p75 LCP of about 2.2 seconds. Each successful prefetch now saves roughly 0.88 seconds at the low end and up to 1.1 seconds.
Reach across the network
Chromium remains the only browser family supporting the Speculation Rules API. Cloudflare Radar data shows about 70% of requests come from Chromium browsers (Chrome, Edge and others). Cloudflare handles hundreds of billions of HTML requests daily; roughly half are cached, and about 1% of those are navigational prefetches from visitors. In total, Speed Brain can save more than 82 years' worth of latency every 24 hours.

Roadmap
A machine-learning predictor
Cloudflare's vantage point on Internet traffic gives it browsing-pattern data it can use to sharpen prefetch prediction without compromising individual user privacy. It is developing an adaptive speculative model that trades the current conservative approach for a data-driven one. The model builds a privacy-preserving user traversal graph per site from same-site Referrer headers, then predicts the likelihood that a typical user moves between any two pages connected by a navigational hop, using aggregated traffic insights.
Rule sets are then tailored with custom eagerness values for each relevant next-page link. Pages with high predicted confidence are aggressively prefetched or prerendered, while pages the model doesn't cover fall back to the current conservative rules. In lab tests the model cut LCP latency by 75%, predicted visitor navigation with roughly 98% accuracy — helping ensure only the right pages are prefetched and resources aren't wasted — and stayed within existing safety guardrails. Scaling it means periodically retraining to follow shifting user behavior and site changes; an online learning approach would largely eliminate manual updates and content drift while keeping accuracy high.
Observability and prerendering
Cloudflare plans to let RUM filter by navigation type, so operators can compare browser rendering of prefetched and non-prefetched content. Prefetching today fetches only the main document; it doesn't tell the browser to prerender the page or download subresources. Later versions will prefetch into the CDN cache and coordinate with browsers on the best prerender candidates, bringing static content closer to instant rendering.
Argo Smart Browsing
Cloudflare is investigating a product that bundles the aggressive ML implementation — 75% lower LCP in lab testing — with Argo Smart Routing under the name "Argo Smart Browsing." Customers will still be able to use Speed Brain on its own; those wanting more enable Argo Smart Browsing with a single click. Alongside up to 75% faster loading of cacheable static content, requests that must reach an origin travel over the most performant network path, for an average 33% performance increase. Optimizations thus apply to nearly every segment of the request lifecycle, whether content is static or dynamic, cached or not.
Enabling it
Speed Brain can be switched on in the Cloudflare Dashboard under Speed > Optimization > Content Optimization > Speed Brain, or through the API. Domains on free plans already have it enabled by default.




