OpenAI, ARIA, and SEO: Making the Web Worse (adrianroselli.com)

OpenAI launched Atlas, a browser with ChatGPT built in, and is encouraging website owners to add WAI-ARIA roles, labels and states so its ChatGPT agent can better interpret page structure and interactive elements. The Atlas developer FAQ and roadmap also flag app discoverability features for Apps SDK developers, signaling OpenAI wants structured metadata it can use to classify and interact with sites. In plain terms: OpenAI is asking authors to annotate pages so its agent can “understand” and surface content more reliably. That request is alarming to accessibility and web-pros for technical and ethical reasons. ARIA exists as an accessibility fallback—not a primary content API—and many practitioners warn that treating ARIA as scrapeable metadata invites misuse: keyword-stuffed aria-labels and aria-roledescription, over-roling of elements, and compounding inaccessibility. It also creates fresh incentives for SEO gaming and automated content extraction, amplified by LLM-driven agents, while exposing gaps in OpenAI’s apparent understanding of HTML/APG best practices. For the AI/ML community this is a cautionary example of how agentic models change web incentives and data pipelines: model designers, web devs, and standards bodies need to anticipate adversarial markup, preserve semantic integrity, and protect content and user privacy rather than normalize ARIA as a content-discovery channel.