Cloudflare published billions of real-user web performance measurements as a free public dataset on Google BigQuery

Anyone who has argued about a P75 Largest Contentful Paint number has been arguing without a reference sample. Cloudflare published one on 2026-09-28: BEACON, an anonymized dataset of billions of Real User Monitoring (RUM) records from 10,000 of the largest websites on its network, updated daily on Google BigQuery and free to query. The dataset is named for Browser Experience Across Cloudflare's Observed Network, and Cloudflare says it expands the community-led RUM Archive standard "100-fold" in coverage.

What is in the dataset

BEACON reports all three Core Web Vitals, Largest Contentful Paint, Cumulative Layout Shift and Interaction to Next Paint, as full histograms rather than single averages. Any percentile can be derived from a query, including the long tail past P75 that a single number always flattens. The dataset also includes LCP sub-parts (Document TTFB, Load Delay, Load Duration, Render Delay) and INP sub-parts (Input Delay, Processing Time, Presentation Delay), so a slow score can be traced back to which stage owns the time. Records are cut by country, operating system, browser and connection protocol, and joined to industry categories from Cloudflare's Intel API.

The findings Cloudflare called out

Two numbers in the launch post are worth quoting on their own terms. Cloudflare reports that in 46 countries where WebKit represents at least 10 percent of traffic, its LCP, INP or both are at least 10 percent worse than Blink-based browsers such as Chrome, Edge and Opera. In Cambodia, WebKit is 17.5 percent of page views and its LCP is 50 percent worse than Blink's. By industry, Government and Politics, Health and Safe for Kids score best on the Core Web Vitals; Ads, Religion and Weather score worst. Soft navigations, which frameworks such as React, Vue, Angular and Svelte use for client-side routing, render two to three times faster than hard navigations at every percentile Cloudflare published, though the initial landing page they build on is often the slowest request in a session.

How the numbers are cleaned

Cloudflare strips domain names and URL paths from the records, aggregates by shared dimensions, and discards any bucket with fewer than five data points. Records are also normalised down so the 10,000th-largest site is not drowned out by traffic from the largest ones. Data is updated daily, and the launch post ships with the exact BigQuery queries behind each table it discusses, saved under names like Global LCP Sub-parts and Country-level Performance Aggregates.

A site owner watching a poor LCP score now has an outside reference to compare against, cut by the same country, device and industry as their own visitors. That is the thing that did not exist before: a public dataset of billions of real-user browsing sessions across 10,000 sites, queryable in a browser tab, at no cost.

Source

• Cloudflare: How fast is the web? Explore billions of real-user measurements with BEACON, 2026-09-28.
• Reference: RUM Archive, the community-led standard BEACON extends.