Web Accessibility as a Foundation for Supporting AI Agents

The Growing Need to Support AI Agents

It’s already well understood that not all website access is coming through human-operated web browsers. Instead, in today’s era of rapidly advancing AI technologies, the presence of AI agents is becoming increasingly prominent.

The term “agent” is defined as a proxy or intermediary. As the name suggests, an AI agent refers to software or systems in which AI autonomously performs tasks and achieves goals on behalf of humans.

In a Column from last September titled “The Distance Between AI Technology and Web Browsers Is Being Questioned” (in Japanese), I described web browsers integrated with AI technologies as “agent browsers.” In a broad sense, these types of browsers can also be considered a form of AI agent.

AI agents may already be visiting and interacting with the websites that are managed and operated by readers of this Column. Additionally, some are even referring to an emerging web, where AI agents access and operate content autonomously alongside humans, as the “Agentic Web.” Such developments raise important questions – such as: will websites need special accommodation for AI agents? If so, what can web professionals do to support such agents? As you might guess from the title of this Column, my answer is: improve content accessibility.

Machine Readability - and the Accessibility Tree - as the Keys

All web content is ultimately delivered to users through machines. Web browsers, screen readers used by visually impaired users, as well as systems that evaluate and rank web pages for search services are all machines. Content that is easier for machines to read and process – that is content with high machine readability - is also easier to communicate effectively to users.

This is why, in our company’s Accessibility seminars, I’ve often explained that search engine optimization (SEO) and accessibility improvement are quite similar in practice as both aim to enhance machine readability.

As AI agents are also a type of machine, content that is highly machine-readable is, therefore, certainly more suitable for them.

One specific mechanism that supports machine readability is the accessibility tree. While somewhat technical, the accessibility tree is an internal information structure generated by browsers from HTML elements and ARIA attributes. Many assistive technologies, such as screen readers, rely on this tree to convey information to users.

Recently, I came across a paper titled “A11y-CUA Dataset: Characterizing the Accessibility Gap in Computer Use Agents.” It highlights that AI agents are not yet fully utilizing accessibility information relied upon by visually impaired users, and suggests that leveraging the accessibility tree could, therefore, be one way to improve performance.

In my view, the key to making web content compatible with AI agents lies in machine readability and more specifically, in the accessibility tree.

Preparing for the Agentic Web Era

At present, not all AI agents make use of the accessibility tree. This creates a “chicken-and-egg” situation: will AI agents begin using accessibility trees first, or will websites improve accessibility and provide better accessibility trees first?

However, some AI agents, such as Browserbase’s Stagehand and Vercel’s agent-browser, are already utilizing accessibility trees. There are also movements to incorporate accessibility trees into MCP server implementations, with examples including Playwright MCP and macos-use for macOS.

For these reasons, I believe that web accessibility will become increasingly important as a foundation for supporting AI agents. In addition to responding to the increasing diversity of users and browsing environments, as well as compliance with legislation, a new motivation is emerging: preparing for an Agentic Web so that content can be understood by AI. By doing so, this is expected to improve accessibility across the entire web.