We crawled 14,764,966 websites to find out what the web
actually looks like. Platform market share, real performance data, security risks, and which
technology choices make a measurable difference.
We crawled the homepage of every primary domain from Google Chrome's User Experience Report dataset. The list of domains comes from real Google Chrome users, and reflects sites that are actually receiving visits. This is what a slice of the active web looks like.
Not every domain in the dataset runs an established website. Some are domains that are for sale, others are default pages waiting for a real site, spam, or just broken. 7.3% of sites blocked us from loading a page at all.
Filtering out the noise and errors leaves us 13.5 million real websites.
WordPress comes in two forms. The open source software from wordpress.org is downloaded and run on hosting of your own choosing, and it accounts for 4,654,967 of the sites we crawled. WordPress.com is the hosted service Automattic sells and it accounts for 75,869. The open source version is 61.4 times the size of the paid one.
WooCommerce is a plugin that can run on either version and is installed on 22.1% of every WordPress site we found.
Each square denotes one in every hundred websites we crawled. The ten market leaders are shown; another 1,049 platforms we can identify run the remaining 1,967,330 sites between them.
# The busier a website is, the less likely it is to run off-the-shelf software
As website traffic increases the most popular CMS platform (WordPress.org) decreases in market share and custom website builds dominate. WordPress is most popular on sites that people visit less frequently.
Traffic rank comes from Google's Chrome User Experience Report, which sorts the sites it observes into popularity bands.
Top 100k
62.4%15.7%21.8%
Top 1M
47%21%32%
Top 10M
29.7%38.3%32%
Custom code (no CMS)
WordPress.org
Other CMS
#64.1% of the platforms we identified are self-hosted
Self-hosted software runs on a server the site's owner controls or pays for directly: they choose the host, the configuration, and when any of it changes. On a hosted platform the company runs both the software and the server.
Shares of the 8,712,650 sites where we identified the platform.
8,712,650
#60.4% of the platforms we identified are open source
Open source content management systems mean the code is open and public and often written and maintained by a community. Source available is similar, but comes with more restrictions. Closed source is private and you are not allowed to edit, view or republish the code.
We map a system to its licence when their code is available on services like npm and GitHub. When a licence can't be identified we assume closed source.
We detect ecommerce on 339 platforms. For multi-purpose CMSs like WordPress.org we detect installations of 6 ecommerce plugins (including WooCommerce and Easy Digital Downloads). For platforms like Wix and Squarespace we also detect the activation of their commerce features.
# Domestic platforms lead in South Korea, Argentina, Russia, Czechia and Saudi Arabia
WooCommerce leads in most countries including Britain, Germany, France, Spain and Italy. Shopify leads in Australia, India and Canada. Everywhere else the market-leader is usually domestic: Cafe24 87.6% in South Korea, Tiendanube 56.8% in Argentina, Bitrix 49% in Russia, Shoptet 37.9% in Czechia, Salla 37% in Saudi Arabia and Prom.ua 35.5% in Ukraine.
Geographic location comes from the site's location schema (where defined) and the domain's TLD, excluding .com and some country domains like .co that have become trendy.
# Shopify stores clear all three Core Web Vitals twice as often as WooCommerce stores
Google's Core Web Vitals measure a webpage's user experience. They assess loading speed, interactivity, and visual stability. We mapped a site's pass rate back to the platform it runs on. Of the two most popular platforms, stores running on Shopify pass Core Web Vitals on 86.9% of sites, nearly double the rate of 44.4% for WooCommerce stores.
Tiendanube89.0% · 45,272 storesShopify86.9% · 761,477 storesWix85.8% · 318,641 storesOpenCart83.0% · 54,201 storesSquarespace82.8% · 107,015 storesBitrix75.2% · 53,996 storesCafe2473.3% · 26,319 storesPrestaShop71.4% · 69,703 storesMagento59.6% · 30,653 storesWooCommerce44.4% · 1,043,221 storesThe 10 largest platforms, each with at least 1,000 live stores.
#192,701 ecommerce sites (10%) send no HTTP security headers
HTTP security headers enforce specific rules for what a site can load and do and help mitigate security vulnerabilities.
On average SaaS / hosted ecommerce sites do a better job of setting security headers on requests, setting 3.03 of these 5 headers on average, versus 1.28 on self-hosted stores. 48.8% of self-hosted stores send none of them at all, against 3.9% of SaaS ones.
Shopify is the most secure platform, sending four of the five headers to nearly every store it runs, and no platform sends all five.
Share of each platform's stores whose homepage arrived carrying the header, for platforms running at least 15,000 stores. Where a hosted platform is not at 100%, it is usually a store running a custom frontend the platform never rendered, or a domain served from somewhere other than the platform's own edge. Presence is not quality: we record that a header arrived, not whether its value was strict enough to stop anything.
# WordPress.org is a very flexible open source CMS running 34.5% of sites
Developers can edit the code and site owners can install and change site themes and plugins. That same flexibility means the user experience, quality and speed of a WordPress site can vary a lot depending on how it's set up, where it is hosted, and what theme and plugins are installed.
Running a version with a known security hole · 409,981 sites
# Three themes run a quarter of WordPress, and all three sit in the bottom half on Google's Largest Contentful Paint (LCP) metric
Hello Elementor, Astra and Divi run 1,157,496 sites between them. GeneratePress leads on LCP at 66.7%.
Caching matters more than the theme. We count a site as cached when a caching plugin, a cache on its server, or a CDN is storing a finished copy of the page. Cached, every theme here responds in 267 to 355ms. Uncached, the same themes take 964 to 1,890ms, so the difference between them only shows on a site that does not cache.
# WordPress editors and page builders, by speed and Core Web Vitals
53.4% of WordPress sites run a page builder plugin we detect.
We looked at WordPress sites from version 5.0 onwards, and measured their median TTFB when the site is cached (server, CDN and/or plugin) versus uncached. The Classic Editor plugin leaves no trace in the page, so we infer it when no block markup and no other page builder is found.
# Some cache plugins perform wildly differently depending on the server they're optimised for
We compared self-hosted WordPress sites running the latest major version, split by web server and by caching plugin. Each panel shows the 10 most-installed plugins with at least 500 sites on that server.
LiteSpeed Cache is built for the LiteSpeed server. On nginx and Apache the same plugin is slower than running no cache plugin at all. WP Rocket holds up wherever it lands.
4,098,928 sites sit behind one provider, Cloudflare
#Cloudflare serves 70.5% of the sites that use a CDN
A CDN sits between a visitor and a website and serves the site's pages and files from a server near the visitor. We detected a CDN on 43.1% of sites, and 70.5% of those are served by Cloudflare.
70.5% of CDN-served sites · 4,098,928 · Standalone
#Cloudflare answers DNS for more sites than the next 3 operators combined
Which nameservers a domain points at is a separate choice from who hosts the site, and most sites inherit it from whoever sold them the domain. Cloudflare's grip here is far looser than the 70.5% of CDN-served sites it delivers, but this is still the first answer every browser waits on.
#Hostinger hosts more sites than any other provider we identified
We identify 555 hosting providers from 455 signals: the network a site's server sits on, the headers the server sends back, and the names its DNS points at. Hostinger is the largest, with 685,708 sites, 6.87% of every site whose host we could name.
#50.9% of sites tell you what server software answered
Every HTTP response can carry a server header that names the underlying webserver that handled the request. 99% of sites behind Cloudflare hide their origin server.
The Herfindahl-Hirschman Index squares every supplier’s share and adds them up, so one dominant player counts far more than many small ones. A pure monopoly scores 10,000. Regulators read below 1,500 as competitive and above 2,500 as highly concentrated.
A market split between more suppliers scores lower, so the four layers are not perfectly like for like. The suppliers column shows how many each index covers. Each layer counts only the sites whose supplier we could identify.
# Seven companies, and how much of the web each one sits inside
Counted by company rather than by product, across every layer above: the nameserver, the CDN, the hosting, the firewall, and the scripts, fonts and embeds a page loads. A site counts once for a company however many of its products it carries, so the parts under a row add up to more than the row.
#78.6% of sites on SaaS platforms pass Core Web Vitals, against 55.6% on WordPress.org
When one company controls the whole stack, sites tend to be fast by default. The options sold on speed are no better than average: headless CMS sites pass at 58.8% and enterprise CMS sites at 57.1%, level with the 60.5% of sites whose platform we could not identify. AI site builders pass at 30.7%.
Core Web Vitals are Google's measure of what real visitors experienced: how quickly the main content appeared, how fast the page responded, and how much it moved while loading.
Under 200ms 67.9% · 2,750,788200 to 800ms 68% · 4,220,091800 to 1800ms 55.4% · 1,228,040Over 1800ms 26.3% · 593,854
# Where caching sits, and how common each layer is
Caching keeps a finished page ready so the server does not rebuild it for every visitor. It can be generated by a plugin or the CMS on the site, on the server itself, or at a CDN in front of both, and 58.8% of sites have at least one of the three in place.
# Only 17.5% of homepages behind a CDN were served from its cache
Where the CDN edge was not caching the HTML at all, the largest group here by some distance, the median page arrived slower than the median site with no CDN at all. That is not the CDN failing; it is a CDN doing nothing except adding a hop.
Where the page is cached, the CDN vendors are within 71ms of each other. Where it is not, they span 884ms.
Compressing a response is the oldest speed win there is, and 6.8% of sites still skip it. Brotli packs the same page smaller than gzip; Zstandard packs and unpacks it faster, and browsers only began accepting it in 2024.
Image optimisation reduces the number of bytes sent on page load, whilst hints tell the browser which order to fetch assets in. Charts show adoption only, how common a technique is, not its effectiveness.
jQuery was released in 2006 to paper over the differences between browsers of that era. Those differences are gone, the browsers that caused them are gone, and 6,611,511 sites still load it.
React is the largest of the modern libraries, on 10.6% of the live sites we visit.
# Webflow includes jQuery on 97.6% of the sites it builds. Wix includes it on 0.2%.
WordPress bundles jQuery with the software, which explains a large part of its reach: 84.9% of WordPress sites carry it. It does not explain Webflow, a visual builder launched in 2013 that puts jQuery on almost everything it makes. Wix is the same kind of product and has all but removed it, so this is a decision each platform takes rather than something the web is stuck with.
# Every modern meta-framework combined reaches 6.3% of the homepages we read
A meta-framework builds a site out of components and renders it on the server, the browser, or both. They are what the last decade of frontend writing has mostly been about. Next.js is the category: 544,513 sites, more than the other nine put together.
Headless sites have a separate system from the rendered page you view. They account for 0.76% of the live sites we visit.
When we detect a site looks like a headless install we take a closer look at the site's JavaScript. We map environment variables, function names, API endpoints and script namespaces to headless CMS providers. Even this is imperfect as code can be abstracted or pre-rendered, so these market share numbers are a minimum.
#Hugo builds 60.9% of the static sites we can identify
A static site generator turns folders of content into finished HTML before anyone visits. There is no database and no application serving the page, so the output is indistinguishable from any other HTML file. We identified 17,871 sites built this way, which is 0.14% of everything we crawled.
#117,961 sites keep their data in Supabase and 59,171 in Firebase
A backend service supplies the database, the sign-in and the file storage behind a site, so the code that runs in the browser is the whole application. Firebase has sold that since 2011 and Google has owned it since 2014. Supabase, built on Postgres and launched in 2020, has caught it up in five years.
We find a backend service in the page for every site, and in the JavaScript only for the sites whose code we read. A site whose platform we had already identified rarely has its JavaScript read, so these counts are a minimum market share.
#Laravel runs 264,354 sites, more than any other backend framework
A backend framework is the code a developer builds a site on when no platform does the work for them. We find one on 668,594 sites, 4.95% of everything we crawled, from the session cookie, the form token or the header it sets for itself. A platform built on a framework counts here too, so Statamic sites count towards Laravel and Discourse forums towards Rails.
#13.0% of sites still include markup for a browser that was retired in 2022
HTML5 became a W3C recommendation on 28 October 2014, and 96.1% of sites now declare it. The rest are XHTML, HTML 4, or the 134,615 pages that declare no doctype at all.
The second chart counts markup written for software that has since been withdrawn. Microsoft retired Internet Explorer on 15 June 2022, and 1,595,195 sites still carry a conditional comment addressed to it. Adobe ended Flash on 31 December 2020, and 21,922 sites still embed it.
A security header is one line in a server's response telling the browser what to refuse: which scripts may run, whether the page can be framed, whether to insist on encryption. Adding one is a single line of configuration. Keeping a strict one working as a site changes is the harder part, which is where most sites stop.
#196,914 sites cannot serve a working HTTPS connection
We request every site over HTTPS first and fall back to plain HTTP only when the handshake fails, 1.46% of sites.
Most of those failed sites already have a certificate. Only 31,203 of them have no HTTPS at all. The rest serve one the browser refuses: issued for a different domain, signed by an authority the browser doesn't trust, or long expired. A visitor gets a full page warning before they get the site.
#48.1% of sites that reveal their PHP version run a version that is no longer getting security patches
We detect PHP is running on 3,080,957 sites. Of those sites 1,838,448 leak or expose the version of PHP the site is running on.
Leaving the PHP version declared in HTTP headers isn't the best idea from a security perspective and updating the version of PHP on your hosting is normally a simple task.
#8.8% of WordPress sites run a version with a known security hole
WordPress leaks its version number in asset paths by default; we detect the version on 62.8% of installs.
WordPress has shipped 959 releases; the latest when we crawled was 7.1.2. Most sites trail it by a version or two. The rest sit on an end-of-life release that will never be patched again.
# Where a WordPress site is hosted changes its odds of being updated
We charted the median TTFB of WordPress sites hosted on one platform against the share of those installs on an out-of-date release. Bottom left are hosting providers that have faster sites that are up to date. Top right are hosts with slower sites and outdated WordPress installs.
↑ Share on an out-of-date releaseMedian time to first byte, ms →
Chart only includes hosts with at least 2,000 WordPress installs.
#19,986 sites print a server error into their own homepage
When something breaks on the server a site can show a plain error page, or it can print the failure into the page it serves. The second kind names the software, the file paths and sometimes the database behind the site, which is the reconnaissance an attacker would otherwise have to do themselves.
Nearly all of it is one setting. PHP, Laravel, Django, Rails and ASP.NET all ship a debug mode that prints the failure to the browser, and all of them expect it to be off before the site goes live.
A web application firewall inspects requests before they reach the site and blocks the ones that look like attacks. Most of what we can see sits at the network edge, in front of the host, and the rest runs on the server itself as a plugin or a module.
We don't tend to detect firewalls as our crawler tries to be polite and not break rules or make intrusive probes. The numbers here are an undercount of the true market share for this reason. Some vendors name themselves on normal requests and are shown here.
4,213,824 sites are accessible and include a skip link
# The least adopted SEO signal reaches 52.4% of sites. The least adopted accessibility signal reaches 34.2%.
Both groups are measured on the same sites and drawn on one scale, so a bar in either can be read against a bar in the other. The SEO signals are what a page declares about itself for a search engine. The accessibility signals are what it declares for assistive software.
Title tag 98.7%XML sitemap 80.3%Meta description 73.9%Schema.org 56.2%og:image 52.4%
Accessibility
HTML lang 92.8%Semantic HTML 90.1%Image alt text 73.9%Form labels 71.1%Skip nav link 34.2%
# A site running a CMS is up to 6.5 times more likely to include a skip link
Accessibility falls as traffic rises: the busiest sites score worse than the long tail. That is not effort, it is composition. Most of the head of the web is hand-built and most of the tail runs a CMS, and a skip link is something a theme provides rather than something an author writes.
A site counts as running a CMS where we identified one. Everything else, from a bespoke build to a platform we could not name, sits on the other side.
# Platforms are optimised for SEO and inconsistent about accessibility
We score every site we crawl across several quality metrics. Here those sites are mapped to the platform they run on, and accessibility and SEO are each scored out of 100.
For SEO we score meta description, canonical URL, robots.txt, XML sitemap, open graph tags and structured data markup. The accessibility score checks for an <h1> tag, a language attribute, pinch to zoom being enabled, a skip link and the use of semantic HTML.
#51% of sites link to a privacy policy, and 25.5% show a consent banner
To detect cookie banners at scale we map 107 consent management platforms, recognising each by its scripts, file paths, cookies or markup. We also check for 35 generic markup patterns and CSS classes, such as 'cookie-banner' and 'gdpr-cookie'.
For privacy policy links we check link URLs for 39 translated variants, and link text for 56. We also exclude third-party links, such as Google reCAPTCHA's, so the figures are not artificially inflated.
These figures should be taken as a minimum, as full detection is difficult at scale.
A consent platform is the banner that asks whether a site may set cookies, and the record of what you answered. We detect 107 of them by name, from the plugins that dominate WordPress to the enterprise suites, plus the banners Wix, Squarespace and Shopify build into their own platforms.
CookieYes is the largest at 11.9% of the sites that show a banner, and everything below the top three is under 6%. 29.6% is a banner we detect by markup or styling, but is either custom coded, or unattributed to a vendor.
# Two companies collect almost all of the web's advertising data
An advertising pixel is a few lines of code a site adds so an ad platform can watch what visitors do after they arrive. It reports the purchase, the sign-up, the page viewed, and ties it back to the ad that sent them. The site gets attribution and an audience it can advertise to again. The platform gets the behaviour of everyone who visits, on a site it does not own.
Meta's pixel is on 16.3% of sites and Google's conversion tag on 14.8%. Nothing else is close, and the two are rarely a choice between one or the other: 83% of the sites carrying Meta's pixel carry a Google tag as well.
#58% of sites carry a tracking script (analytics, advertising, CRM or lead generation), and 47.9% carry Google Analytics
Analytics scripts tell site owners what pages are being viewed and where visitors have come from. Sometimes these can be privacy conscious and anonymise the data that is recorded or shared.
Analytics and tracking scripts can be loaded by tag managers like Google Tag Manager. 21.5% of sites use it, and when they do we take a look inside to see what scripts are loaded.
No scripts 40.7% · 5,012,3101 script 26.2% · 3,222,3882 scripts 14.5% · 1,783,3463 scripts 8.9% · 1,098,8874 scripts 5.2% · 636,8375 or more scripts 4.6% · 565,406
# Consent banners lead in Europe. Ad trackers do not follow the same pattern.
The 27 highest banner rates are all European countries. Italy is highest at 65.8%. Israel is 28th, the first country outside Europe. The bottom of the list is under 4%.
Germany, Austria, Switzerland and France are among the lowest on the ad tracker layer. Bulgaria, Hungary and Romania are among the highest, alongside Vietnam and Ukraine. EU countries range from 9.7% to 38.3% under the same law.
Geographic location comes from the site's location schema (where defined) and the domain's TLD, excluding .com and some country domains like .co that have become trendy.
#63.4% of sites running Google Analytics or Tag Manager have no consent mechanism. Of those with a banner, 25.8% pass the answer to the tag.
A banner is the consent interface the visitor sees. Consent Mode is a line of JavaScript telling the Google tag whether it may collect. They are independent: a site can have either, both or neither.
24.1% show a banner but send no signal, so the tag fires whichever the visitor chooses. 8.4% do both.
No banner, no Consent Mode 63.4% · 3,909,339Banner only 24.1% · 1,484,080Banner and Consent Mode 8.4% · 515,850Consent Mode only 4.1% · 255,346
#71.5% of sites load files or a service from Google
A site is counted as loading a Google service when its HTML asks the browser to fetch something from a Google host: a script, a stylesheet, a font file, an embedded frame.
Each request sends the visitor's IP address along with other data, but Google states it doesn't always use this data for advertising profiles, like with Google Fonts. Regardless, its reach across the corpus is impressive: 71.5% of sites touch Google in some form.
An llms.txt file sits at the root of a site and lists the pages it wants an AI assistant to read. It was proposed in 2024 and has never been ratified by anyone. No major model provider has committed to acting on it, Google has confirmed that no Search system reads it, and crawlers do not request it in any meaningful volume.
55.2% of the files we found sit on platforms that switched one on for every customer, and Shopify alone accounts for 743,970. Where someone has to add it themselves, adoption collapses.
Blocking rises with traffic. The two kinds of crawler are not blocked equally: bots that collect text to train models are disallowed around three times as often as bots that fetch a page in response to a question.
Percentages of that tier's sites disallowing that bot in robots.txt
# AI builders load slowest of any category we identify
You trade the speed of building the site for a heavy front end that takes longer to load. An AI builder sends a near-empty document and assembles the page in the browser afterwards: the median Lovable site sends 12KB of HTML, against 161KB for WordPress. Nothing appears until the JavaScript has run, and that wait is what Google's field data records. The two builders that hold up, Hostinger Horizons and v0 by Vercel, are the two that send a real page.
# A standard for refusing text and data mining exists, and 1,113 sites use it
The W3C's TDM Reservation Protocol lets a server say, in one header, that its content is reserved from text and data mining. 1,113 sites send it. Another 135 send the same header set to allow mining. Almost all of both are academic journals and university research portals.
1,080,408 domains blocked our crawler from loading a page. Some blocks come from WAFs, some from hosting providers and some are a plain 403 Forbidden error which we can't attribute.
# We can still identify who serves a blocked domain, usually not what CMS platform it runs on
We can still detect the CDN from the response headers, the nameserver from the DNS records, and the host from who owns the IP address. None of those needs the page. Detecting the platform usually does.
#1.71% of the sites we visit are spam, and 0.17% declare themselves as adult content
The spam sites we detect share a common build signature: the same HTML, tracking scripts or asset hosts. They're custom builds and rarely run on a CMS platform we detect, but are on thousands of domains.
We detect adult sites when they declare an RTA label or a rating tag. CMS detection rates for them are also lower than on an average site.