Cloudflare dropped a dataset this week: web speed measurements based on billions of real users, covering 10,000 top websites and updated daily. The takeaway is stark — developers and ordinary users live in two completely different internets.

What this is

Cloudflare published the BEACON (Browser Experience Across Cloudflare's Observed Network) dataset, making years of real-user web speed data collected on behalf of enterprise customers freely available.

Three core metrics, the industry's accepted standard for evaluating web experience:

  • LCP (Largest Contentful Paint — when the main page content finishes loading)
  • CLS (Cumulative Layout Shift — whether page elements jump around as they load, measuring visual stability)
  • INP (Interaction to Next Paint — the latency from a click or swipe to a visible UI response)

The data covers 10,000 top websites across all major browser engines and refreshes daily on Google BigQuery. Crucially, it provides the full distribution (a histogram of how many users fall into each bucket), not a simple average — any percentile is queryable.

Industry view

The most counter-intuitive finding: Safari's WebKit engine actually performs best overall, but in 46 countries where WebKit accounts for over 10% of traffic, LCP or INP is at least 10% slower than Chrome. Cambodia stands out: WebKit accounts for 17.5% of page views, but LCP is 50% slower than the Chrome family (Chrome/Edge/Opera all use the Blink engine).

Supporters see this as a good thing — previously, the "pages are fast" conclusions developers drew from testing on their own laptops over gigabit Wi-Fi were pure self-deception. Now there's finally objective data, so product decisions no longer have to serve just the wealthiest 10% of users.

There are also calmer voices. The data is based on Cloudflare's network, with traffic concentrated on mainstream sites in Europe and the US; coverage is weaker in mainland China and parts of Africa. The Core Web Vitals framework itself is Google-dominated, and its reference value for other search and recommendation systems is debatable. A cross-border e-commerce practitioner told us privately: "Knowing Cambodia is slow is useless. What matters is how much real data our SDK can capture there — Cloudflare's local coverage remains a mystery."

Impact on regular people

For enterprise IT: teams building AI Agents (AI programs that can automatically operate browsers), automated testing, or overseas businesses will find BEACON's "long-tail distribution" more useful than averages. AI systems are most vulnerable to crashes under edge network conditions, and management rarely notices.

For individual careers: there will be a more objective reference for evaluating whether a website or AI tool feels laggy. Next time a product manager says "our load times are fast," you can push back: "What about at P95 (the level covering 95% of users)?"

For consumer markets: now that the experience gap for users in emerging markets like Cambodia and Indonesia has been quantified, going-global teams can no longer hide behind "the user's network is bad." Developers must rethink: design for the fastest network, or design for the most ordinary phone?