Accessibility Testing Tools: What AI Catches, and What it Still Misses

When accessibility testing vendors claim “90% of issues resolved automatically,” it’s tempting to assume they automatically make your site 90% compliant with WCAG guidelines. But issue resolution isn’t compliance. And there hasn’t been a study published that proves AI can deliver a compliant website on its own. Compliance still requires manual testing.

Regulators responded to a stronger version of the claim. In April 2025, the FTC finalized a $1 million order against accessiBe over claims that its AI tool could make any website compliant with WCAG. The FTC order bars the company from making that claim without supporting evidence. That outcome should make buyers skeptical of any AI testing tool that promises full automation of compliance without an independent evaluation to back up the claim.

AI enhances automated testing by improving issue-detection accuracy, reducing false positives, and scanning pages faster than rule-based predecessors. But a passing scan is only part of compliance. It can’t tell you whether a person using keyboard navigation or a screen reader can successfully navigate your site.

What’s actually improved: AI-enhanced automated testing

Automated accessibility testing exposes issues that fail to pass WCAG success criteria by scanning code and rendered pages. Using AI finds common issues like missing alt text, color contrast issues, unlabeled form fields, and empty links more consistently. 

AI and machine learning models perform visual analysis to identify low-contrast separation in images and text. AI reduces the number of false positives reported, which means less time chasing the wrong fixes. Newer tools find issues that older rule-based scanners couldn’t, like alt text that exists but doesn’t accurately describe an image.

That last point, semantic understanding, was the subject of research presented at the ACM ASSETS 2025 accessibility conference. University of Stuttgart researchers used an LLM in combination with the accessibility testing engine axe-core to identify semantic violations. 

Identifying these violations goes beyond telling them if alt text was present or not. It tells them if any text doesn’t accurately describe an image. Testing the same 16 pages, rule-based tools reported zero violations. The LLM layer flagged 104. 

Results vary widely from one tool to the next. In a 2024 test by Equal Entry, researchers built a 31-page site with 104 added defects across 47 WCAG 2.1 AA success criteria. They tested six different scanners against the site, and those scanners only found between 3.8% and 10.6% of the defects, with false positives ranging from 2 to 474. 

Each defect only appeared once on the site because the test measured each tool’s ability to identify distinct issues. On a real site, where templates cause issues to repeat, those same issues would span multiple pages.

AI delivers faster, more accurate scans than a few years ago. Though it still needs a human review to verify the website works.

Where AI still falls short: keyboard navigation and judgment calls

AI accessibility testing will tell you whether an element exists on the page and is coded correctly. It can’t tell you if assistive technology navigates the page successfully. When a person needs to take specific actions, like filling out four fields in a form, an automated test won’t tell you if assistive technology sees the fields in the correct order. Keyboard access is where the gap is most obvious.

In a WordPress Accessibility Meetup hosted by Equalize Digital, accessibility specialist Tanveer Khan demonstrated two failures that automated tools frequently miss:

• Illogical focus order: On a page whose header looked correct, keyboard focus jumps from the first link to the fourth, then back to the second and third.
• Focus escaping dialog: Focus doesn’t stay inside an open dialog and could move onto the page behind it, leaving keyboard and screen reader users unsure where they are.

In both cases, keyboard navigation is broken, even though the sites passed an automated accessibility scan.

AI hasn’t solved this gap yet. The University of Stuttgart researchers found their model worked from static screenshots, so it missed dropdowns, pop-ups, and other content that requires interaction. Those are the same places keyboard problems crop up.

Some things caught by AI still need a human judgment call:

• Whether alt text accurately describes an image in the context it appears.
• Whether a custom menu, image carousel, or data picker behaves the way a screen reader user expects.
• Whether error messages and instructions are understandable when read aloud.
• Whether content remains usable as zoom level increases.

“One-tool compliance” was a misconception before AI arrived, and it remains that way. Someone still needs to use the site the way your visitors do to verify it works correctly.

The accessibility tools enterprises actually combine

Meeting accessibility requirements typically involves a combination of testing tools, each specialized for different stages of the publishing lifecycle. Preventive tools catch issues before they reach production. Reactive tools monitor the live site. The following table provides examples of each category of tools.

How to compare accessibility testing tools

In each tool category, compare tools by checking these criteria:

• Coverage vs. accuracy: How many issues does it find, and how frequently is it wrong? Too many false positives hide real issues and waste time chasing the wrong fixes.
• Fit for each role: Content and QA teams will use visual tools to find and fix issues, while developers want something that integrates into the CI/CD pipeline.
• Workflow integration: How does the tool integrate into your content workflows and release pipeline?
• Audit trail: Can the tool show what changed, when it changed, and who made the change?
• The vendor’s VPAT: Does the company selling you the tool meet the accessibility standards it tests for?
• Actionability: Does the tool include the details required to fix the issue?
• Cost and overhead: How much does the tool cost and how much time will your team spend triaging alerts?

Why a governed system beats a single AI scan

Most content teams aren’t allowing AI to publish content without supervision. The same level of caution applies to accessibility. In WordPress VIP’s Future of the Web 2026 research on AI content governance, 85% of enterprise leaders said AI content published without human review erodes brand trust. 

AI can triage and flag accessibility issues faster than ever. But human review remains a valuable step for verifying the experience actually passes before declaring a page compliant.

WordPress VIP applies this governance approach to its own properties. We had a featured agency partner run a full accessibility audit of wpvip.com, docs.wpvip.com, and the VIP Dashboard. 

Gary Jones, Engineering Lead at WordPress VIP had this to say about the experience, “We want our teams to understand why web accessibility is beneficial, how that can manifest in code, and why it’s not just a checkbox activity to avoid legal repercussions.”

What your CMS changes

Your CMS determines how realistic a governed system is to maintain:

• Fix it once at the source. Reusable blocks and templates simplify issue resolution. When a fix in a template flows through to every page, it can resolve hundreds of issues at once.
• Control who changes what. Roles and permissions are effective tools for controlling who edits components with accessibility impact. Tracking those changes creates an audit trail for compliance reviews.
• Enforce each workflow. Required review steps put a human check before a draft becomes a published page.
• Start with the most visited pages. Content analytics tell you which pages get the most traffic. That’s the place to start when you’re choosing what to fix first.

Migrations, redesigns, and major releases are the moments when accessibility is most likely to drift, so plan for extra testing as part of any major change.

Choosing a tool is easier than building a system 

AI makes accessibility testing tools faster and more accurate. But achieving WCAG compliance and site usability still requires someone to own the testing, real people to interact with the pages, and a plan to fix issues at the source. 

An accessibility scanner nobody checks isn’t a compliance program. It’s a false sense of security with a green checkmark.

Before you add a new accessibility testing tool, ask three questions:

• Does it fit the way our teams already work?
• Does it give each role (design, development, content) feedback they can act on?
• Does a person review what the tool finds before the page is considered done?

See how enterprise leaders are putting human review at the center of AI governance in Future of the Web 2026, Chapter 4. Then put it into practice by downloading The Website Accessibility Implementation Toolkit.

Frequently asked questions