Lovable's AI-Built Website Scores 100% on Accessibility Tests—But Fails Real-World Screen Reader Testing
Key Takeaways
- ▸Automated accessibility testing tools like Axe provide incomplete validation—they can give perfect scores while missing real-world usability issues for assistive technology users
- ▸AI-generated sites need deliberate focus on managing focus states and component layering for screen reader users, not just template accessibility
- ▸Third-party integrations (security, payments, tracking) can silently break accessibility without appearing in automated audits
Summary
A real-world accessibility test of a website built entirely by Lovable's AI revealed significant gaps between automated testing scores and actual usability for people with disabilities. The site for a developer conference, built using Lovable on a smartphone during commutes, achieved a perfect 100% accessibility score using Lovable's built-in testing tool (which runs Axe), but struggled when tested with a screen reader user navigating on an iPhone.
The testing uncovered several critical issues that automated tools missed. Most significantly, when the site's mobile menu was open, the screen reader continued reading content beneath it rather than staying focused on the menu itself—a "ghosting" effect that confused the user about their position in the interface. Similar problems occurred with the ticket purchase modal, where the user initially got stuck behind the modal when using swipe navigation and had to switch to touch navigation to access it.
Beyond UI component issues, the site also suffered from unexpected interruptions from security features. Third-party Cloudflare verification tools automatically announced "Verify if you are a human" repeatedly, creating what the tester described as "a megaphone going off in a library" for someone trying to navigate the site. These problems stemmed from a mismatch between the visual interface and the underlying code structure, as well as poor integration of third-party security and accessibility features.
- Testing with actual assistive technology users remains essential; even fully AI-built sites require human validation by people with disabilities
Editorial Opinion
While Lovable's ability to generate functional, production-ready websites from voice commands is impressive and democratizes web development, this test reveals a troubling blind spot: perfect automated accessibility scores are meaningless if the site doesn't work for people actually using screen readers. AI builders must make assistive technology testing a core part of their development loop, not an afterthought, or they risk perpetuating digital exclusion at scale.
