accessiBe Finds Only 36.8% Check AI Content Every Time

accessiBe released a six-report research series on July 28 after surveying 304 US ecommerce, retail, and direct-to-consumer decision-makers about how they create and review AI-generated customer experiences.

The adoption number is large: 78% of surveyed brands said AI generates at least a quarter of their customer-facing content. The control number is smaller. In accessiBe’s governance report, only 36.8% said they evaluate every AI-generated item for accessibility before publishing; another 42.1% do so frequently, but not for all content.

For ecommerce teams, the AI-generated ecommerce content accessibility gap is not an argument against AI. It is a production-control problem. When IVRIS covered Canva’s AI-content trust findings, the warning was that output can scale before judgment and policy. accessiBe’s data adds a harder requirement: customer-facing content also needs an accessibility release gate.

What accessiBe’s Research Actually Found

The series maps AI use across product copy, images, alt text, adaptive layouts, chat, forms, checkout, and promotions. In the usage report, 69.4% of respondents said they use AI for product descriptions and embedded chat, 60.1% for alt text, and 58.5% for adaptive layouts.

Those are not neutral market-wide estimates. accessiBe sells accessibility technology, and the findings are based largely on self-reported responses from a defined US sample. Brands should read the percentages as vendor research that identifies an operational pattern, not as proof that every ecommerce program has the same risk profile.

The pattern still matters because AI is being applied to parts of the storefront that determine whether shoppers can understand products, move through interfaces, and complete purchases. Accessibility therefore belongs inside the core ecommerce operating model, not in a cleanup queue after a catalog or interface has already gone live.

The Risk Is a Release-Gate Failure

The strongest finding is the distance between written policy and routine execution. accessiBe says 66.4% of respondents have a documented plan for AI accessibility, but only 36.8% evaluate every item before publication. Just 12.5% reported a dedicated accessibility team or lead.

The press-release headline describes the content as almost entirely unchecked. The underlying report supports a more precise conclusion: many brands do review, but coverage is inconsistent. A frequent review process can still miss the exact product page, chatbot response, modal, or layout variation that creates a barrier.

Our read: the control should attach to the asset, component, or experience before release. A policy document cannot show which version was tested, what failed, who approved an exception, or whether a later AI edit invalidated the earlier review. That is the same evidence problem addressed by an evidence layer for AI-assisted workflows.

Where AI-Generated Ecommerce Content Breaks

accessiBe’s content-failure report says 73% of brands that audited their AI-generated content found accessibility issues. It highlights generic or inaccurate alt text, layouts that disrupt predictable navigation, and chat experiences that produce dense or poorly structured responses.

The commissioned storefront test makes the interface risk more concrete. accessiBe had two Lovable-built ecommerce sites professionally tested against WCAG 2.2 AA with screen readers, keyboard navigation, and automated scanning. The audits found 24 issues, including eight critical failures. A more detailed prompt reduced the total issue count, but both sites shared the same critical-failure profile.

Examples included dialogs without accessible roles or labels, focus remaining on the page behind an open popup, unlabeled controls, and contrast failures. These are not cosmetic defects. W3C guidance requires user-interface controls to expose a programmatically determinable name, role, and value, while keyboard focus order must preserve meaning and operability.

What Ecommerce Teams Should Change Before Publishing

• Classify AI output by risk. Treat alt text, forms, checkout, dialogs, navigation, adaptive layouts, and live chat as higher-risk than low-interaction promotional copy.
• Name an accountable owner. Assign one person or team to define the release rule, approve exceptions, and resolve gaps between marketing, engineering, design, legal, and accessibility specialists.
• Automate repeatable checks. Run structural tests for labels, landmarks, contrast, keyboard traps, and component states whenever AI changes a page or reusable template.
• Use human testing where automation is weak. Test high-risk flows with keyboard navigation and screen readers, and review whether generated alt text and chat responses are accurate, concise, and useful.
• Keep an asset-level record. Store the content version, generating tool, source material, test result, reviewer, approval, and exception. Reopen the gate whenever AI materially changes the experience.

This record should follow the content into production. IAB’s GenAI asset-provenance framework applies the same logic to video: teams need to know what AI changed, which source supported it, who accepted it, and where it ran. Ecommerce accessibility adds one more field that cannot be inferred later: whether the exact customer-facing version was tested before release.

accessiBe’s research does not establish that every AI-generated asset is inaccessible. It shows why brands should stop treating accessibility as an assumed property of generated output. AI can accelerate production, but the publish button should remain behind evidence that the experience works for the people expected to use it.

Frequently Asked Questions