TL;DR
- Cloudflare published BEACON on 28 September 2026, a public dataset of anonymized real-user LCP, CLS and INP data from the 10,000 largest sites on its network, refreshed daily in BigQuery.
- Domains and URL paths are removed and groups under 5 data points are dropped, so individual sites cannot be looked up.
- In 46 countries where WebKit has over 10% of traffic it trails Blink by 10% or more; in Cambodia WebKit is 17.5% of page views with LCP 50% worse.
- Landing pages hit a P75 LCP of 2,681 ms; Cloudflare says finding the LCP resource and unblocking render matter more than download time.
Cloudflare has released BEACON, a public dataset of anonymized real-user speed measurements from the 10,000 largest sites on its network. The company announced it on 28 September 2026 in a blog post by Ryan Townsend and Nic Jansma. It covers Largest Contentful Paint (LCP), Cumulative Layout Shift (CLS) and Interaction to Next Paint (INP), refreshed daily in Google BigQuery, and it gives marketers a Core Web Vitals benchmark that does not come from Google's own CrUX data.
What BEACON is and how it is built
BEACON stands for Browser Experience Across Cloudflare's Observed Network. Cloudflare describes it as billions of daily performance records collected from real visitors, drawn from the 10,000 largest sites on its network. The data is public in Google BigQuery under the project cf-open-web-performance, and Cloudflare says it is updated daily. It follows the RUM Archive community standards and, by Cloudflare's account, expands the RUM Archive footprint 100-fold. The post gives no absolute record count beyond "billions".
The three metrics are LCP, CLS and INP. They are published as full histograms, so a reader can calculate any percentile rather than relying on a single average. The post mentions P50, P75, P90 and P95 in its own analysis.
To keep the sample balanced, Cloudflare says it normalized every site to the traffic volume of the 10,000th-ranked site. This stops the busiest sites from dominating the numbers. Records are then grouped by country, operating system, browser and connection protocol.
Anonymization: why you cannot look up your own site
Domain names and URL paths are stripped from the dataset, and any group with fewer than five data points is discarded. The result is useful for industry and market benchmarks but not for checking a named competitor. Anyone expecting a per-site league table will not find one here. The data answers questions such as how a browser, a country or an industry category performs, and it leaves a single website unidentifiable.
The dataset also does not replace measurement of your own site. Field data for your own pages still has to come from your own real-user monitoring, from Search Console, or from another source that is tied to your domain.
The main findings in the post
Cloudflare reports several results from the data. The table below lists the ones that carry specific figures.
| Finding | Figure reported by Cloudflare |
|---|---|
| Countries where WebKit (iOS) trails Blink browsers by 10% or more on LCP and/or INP, among countries where WebKit has over 10% of traffic | 46 countries |
| Cambodia: WebKit share of page views, and LCP gap versus Blink | 17.5% of page views; LCP 50% worse |
| Single-page app hard navigation, P75 | 1,421 ms |
| Single-page app soft navigation, P75 | 582 ms |
| Landing pages (hard navigation), P75 LCP | 2,681 ms |
| Landing pages (hard navigation), P95 LCP | 8,940 ms |
The browser result is the most surprising. WebKit, the engine used by iOS browsers, is described as best overall, yet in 46 countries where it carries more than 10% of traffic it lags Blink browsers by at least 10% on LCP, INP or both. Cambodia is the one example given: WebKit accounts for 17.5% of page views there and its LCP is 50% worse than Blink. The post names Cambodia as its example, and the source summary of it does not list the other 45 countries.
Industry categories
By industry category, Cloudflare reports that Government and Politics, Health, and Safe for Kids perform best, while Ads, Religion and Weather perform worst. The post does not, in the text read, publish the numbers behind that ranking in a form that can be quoted here, so the ranking is the whole fact. Treat it as a direction and not a benchmark to hit.
Where LCP time actually goes
One of the more practical sections breaks LCP into sub-parts and lists three values for each: good, needs improvement and poor. The figures below are reported as they appear in the post.
| LCP sub-part | Good / needs improvement / poor |
|---|---|
| Document TTFB | 598 ms / 1,015 ms / 1,891 ms |
| Load delay | 76 ms / 1,049 ms / 1,485 ms |
| Render delay | 157 ms / 437 ms / 2,002 ms |
The post lists a fourth sub-part, load duration, with values of 119, 199 and 119 ms. The poor value equals the good value, which looks inconsistent, so it is left out of the table instead of being repeated without a way to check it. Cloudflare's conclusion is stated plainly: downloading the resource contributes the least to perceived load delay, while discovering the LCP candidate and unblocking rendering contribute more.
For a site owner, that points the work toward how early the browser finds the main image or text block, and toward what blocks it from painting. Compressing a file that the browser discovers late helps less than making the browser discover it sooner. This is an interpretation of Cloudflare's conclusion, and the post presents it as a finding about the dataset and not as a prescription for any one site.
Where INP time goes
INP is split into three sub-parts in the same way.
| INP sub-part | Good / needs improvement / poor |
|---|---|
| Input delay | 18 ms / 32 ms / 84 ms |
| Processing time | 55 ms / 112 ms / 284 ms |
| Presentation delay | 56 ms / 111 ms / 217 ms |
Cloudflare notes that for the slowest interactions, JavaScript execution dominates, and that presentation time, which covers work such as layout recalculation, is also significant. In practice the post points toward reducing the amount of script that runs in response to a tap or click, and toward limiting layout work that follows it.
Single-page apps and landing pages
Cloudflare compares hard navigations, which load a full page, with soft navigations in single-page applications built with frameworks such as React, Vue, Angular and Svelte. Soft navigations render two to three times faster at every percentile reported.
| Percentile | Hard navigation | Soft navigation |
|---|---|---|
| P50 | 791 ms | 274 ms |
| P75 | 1,421 ms | 582 ms |
| P90 | 2,636 ms | 1,169 ms |
| P95 | 4,122 ms | 1,816 ms |
Landing pages are slower than the general hard-navigation figure. For hard navigations to landing pages, the post reports P50 of 1,370 ms, P75 of 2,681 ms, P90 of 5,397 ms and P95 of 8,940 ms. The post also notes that Google Chrome has added support for the Soft Navigations API, which allows accurate measurement of single-page apps.
Wealth, bandwidth and the gap in devices
Cloudflare joined BEACON with World Bank GDP per capita figures and reports a correlation between a country's economic conditions and its web performance. It says European bandwidth is notably higher than that of other continents, and that Africa shows smaller transfer sizes, which Cloudflare reads as content being adapted for constrained networks. The release is framed around the gap between developers working on fast laptops and users on older, budget phones.
The post also mentions a new Web Performance section on Cloudflare Radar, tied to its Internet Quality Index, and says LCP and INP sub-part metrics will arrive in the Cloudflare RUM dashboard in the coming weeks. It points to Smart Hints and Zaraz as Cloudflare tools related to resource discovery and third-party scripts. These are Cloudflare products, so any decision to use them should rest on testing against your own pages.
What the source did not say
- Exact record counts. The post says "billions" and gives no total.
- Any Thailand-specific figure. Cambodia is the only Southeast Asian country the post names in the text read.
- Which 45 other countries, beyond the Cambodia example, show the WebKit gap.
- Whether the sample is representative of small sites, since it is built from the 10,000 largest sites on one network and normalized to the 10,000th-ranked site.
- Any link between BEACON figures and Google rankings. The post is about measurement and benchmarks.
What this means for Thai marketers
The post has no Thai numbers, so this section is reasoning. Thai audiences are often reached on phones, and iPhone share varies by campaign, so the WebKit finding is a reason to test pages in iOS Safari rather than assume Chrome results carry over. Landing pages used for paid search and campaign traffic are a sensible place to start. The finding that discovery and render-blocking matter more than download size also suggests looking at technical SEO items such as how hero images and fonts are requested before looking at file compression.
The dataset itself offers benchmarks, not a ranking signal. It lets a team ask whether its own LCP sits near the typical value for the same market or industry, and it gives an independent comparison alongside Google's CrUX data.
What to check on your own site
- Measure LCP, INP and CLS in the field on iOS Safari and on Android Chrome separately, not only on desktop.
- Split LCP into the four sub-parts and see which is largest for your landing pages. Fix the biggest one first.
- Check how early the LCP image or text is discoverable in the HTML, and what blocks rendering before it.
- For slow interactions, look at JavaScript running after a tap or click and at layout work that follows.
- If you run a single-page app, measure soft navigations as well as first loads.
- Compare your own results against the BEACON percentiles for the relevant country or industry category, remembering that individual sites cannot be identified in it.
Frequently asked questions about Cloudflare BEACON
What is Cloudflare BEACON?
BEACON is a public, anonymized dataset of real-user LCP, CLS and INP measurements from the 10,000 largest sites on Cloudflare's network. It is hosted in Google BigQuery under the project cf-open-web-performance and refreshed daily.
Can I look up my own site or a competitor in BEACON?
No. Cloudflare removes domain names and URL paths and discards any group with fewer than five data points. The data is grouped by country, operating system, browser and protocol, so it supports benchmarking but not site-level lookups.
Is iOS faster or slower than Chrome according to the data?
It depends on the market. WebKit is described as best overall, but in 46 countries where it carries more than 10% of traffic it lags Blink browsers by at least 10% on LCP and/or INP. In Cambodia its LCP is 50% worse than Blink.
Where does LCP time go according to Cloudflare?
Mostly to discovering the LCP resource and unblocking rendering, not to downloading it. Cloudflare says resource download contributes the least to perceived loading delay, so earlier discovery and fewer render blockers matter more.
Bottom line
BEACON is a new public benchmark, and its most usable results are the LCP and INP sub-part breakdowns, the soft versus hard navigation gap, and the slow landing page figures. It cannot show how your own site performs, so your own field data remains the starting point.
