The pour 100,000.

By failures

Every failing element on every site was counted, and the sites grouped by that count. Most sites have a handful to a few dozen; a small group run into the hundreds, and they pull the average up. Each row is one group, and the bar is how many sites fall in it:

• Failing elementsSites
• none2,586 (5.1%)
• 1 to 510,142 (20%)
• 6 to 2013,934 (27%)
• 21 to 5010,684 (21%)
• 51 to 1006,992 (14%)
• 101 to 2504,882 (9.6%)
• more than 2501,670 (3.3%)

By rule

Six rules make up most of what was found: text must have sufficient contrast against its background on 73% of sites, links must have an accessible name on 45% of sites, every form field needs a label on 42% of sites, every <img> needs a text alternative on 32% of sites, text inside form controls must have sufficient contrast on 28% of sites, and interactive targets must be at least 24×24 pixels on 25% of sites. A rule is one check. The criterion is the part of WCAG it tests. Elements per site is the average over all 50,890 sites, including the ones that passed.

Behind each rule is a person locked out:

• Low-contrast textpeople with low vision, and anyone reading on a dim screen or in sunlight
• Links with no namea screen reader hears "link" with no idea where it goes; voice control cannot click what it cannot name
• Unlabelled form fieldsa screen reader reaches the field and cannot say what it asks for
• Images with no text alternativepeople who cannot see the image get the filename, or nothing
• Faint text inside form fieldspeople with low vision typing into fields they can barely read
• Small touch targetspeople with tremor or limited dexterity, and every thumb on a small screen

By success criterion

The same failures, grouped by the WCAG 2.2 criterion they break. A site counts once per criterion. A rule that covers two criteria counts in both.

By severity

How much a failure gets in the way: critical stops a task, serious makes it hard, moderate and minor slow it down.

By page shape

The average site also has 84 images, 60 headings, 283 links, 113 buttons and 40 form fields. Bigger pages have more to fail:

By popularity

The Chrome UX Report groups sites in tiers: the top 1,000, then the top 5,000, 10,000, 50,000 and 100,000. Each row is the sites in that tier and not in a smaller one.

By breadth

Elements say how much is wrong. Criteria say how many kinds of thing are wrong: forty low-contrast links is one criterion. Sites by the number of criteria they fail:

• Criteria failedPages
• none2,586 (5.1%)
• one5,674 (11%)
• 2 to 312,324 (24%)
• 4 to 623,696 (47%)
• 7 or more6,610 (13%)

By consent

A consent layer covering the page was accepted before the audit, so the page was measured, not the layer. Sites with and without one:

By language

Languages with at least ten pages, each the one the page declares; 6,019 pages (12%) declare none.

Site scores

It can be done: canva.com, bbc.co.uk, support.microsoft.com, yandex.com and whatsapp.com are among them.

Nothing the engine could prove wrong is a floor, not a pass; the human checks still apply.

The hundred pages with the most failing elements:

Click a heading to sort. Rank is the page's place in the list; failing is failing elements; rules is how many rules failed; elements is the size of the page.

Month by month

The run repeats monthly on the freshest list. The months so far:

Method

The list is the Chrome UX Report's ranking of sites for July 2026, built from what real Chrome users load and published in tiers, taken from the crux-top-lists (opens in a new tab) mirror. The rank here is the site's place in that list; within a tier the order is the list's own. Sites were tried in order. One counts when it answered with an HTML page; a bot-challenge page in place of the site is a skip, not an audit. Each site was loaded once in headless Chromium at 1440 by 900 pixels, given 2 seconds to settle, then audited: the top document only, no scrolling, no clicking, a consent layer accepted where one covered the page. The engine ran at WCAG 2.2 AA, the same as the extension and the command line. A later pass retried 300 entries, 289 that did not finish inside the time allowed and 11 that failed during the audit; 130 answered with a page and count as audited.

39,660 entries could not be audited: 11,635 redirected to a site already in the report; 8,944 were refused by the site's firewall after every try; 6,693 did not resolve or refused the connection; 5,969 showed a bot challenge instead of the page; 5,191 did not finish inside the time allowed; 881 answered with an HTTP error; 340 rendered too few elements to be a page; 5 failed during the audit; 2 served something other than an HTML page. A further 9,450 were tagged unavailable by an earlier pass and not tried. Most are CDN, DNS and advertising domains with nothing to visit. Firewall blocks and bot challenges are never tagged; they are tried again next time.

Not sure

Nothing above is a guess. When the engine cannot be sure, it hands the element to a person, 82.7 per site on average.

Limits

Automation checks part of WCAG. A site with no failures here can still be hard to use; only a person can find the rest. One failing element is one count, so a menu of forty low-contrast links counts forty times. Sites change by the hour. The window size matters too: sites show different content at different widths, and these numbers are for this one. And a site is judged on the page it chose to serve: where that was a sign-in form or a redirect stub rather than a home page, a clean result speaks for that form, not the site behind it.