Google Tests Agentic Browsing: A Challenge for Websites

Google is exploring new ways to enhance website interaction with artificial intelligence agents through an experimental feature called "Agentic Browsing." Integrated into its analysis tool Lighthouse, this new category aims to test how well websites are prepared to accommodate agents capable of performing tasks such as filling out forms or comparing products. It is important to note that this category is based on proposed standards and is not yet final.

Unlike traditional Lighthouse assessments, which assign a score from 0 to 100, Agentic Browsing simply provides a ratio of successful checks. This audit focuses on several technical aspects, including the integration of Google's WebMCP API. This API allows developers to make logic and forms specifically accessible to agents. The audit also checks the accessibility tree, which serves as a central data model for machines, the visual stability of pages via Cumulative Layout Shift (CLS), as well as the presence of a llms.txt file.

A striking example of these tests is that of Airbnb, which only manages to pass one of the three Agentic Browsing checks. The accessibility tree of the site is not properly structured, the retrieval of the llms.txt file fails, and the WebMCP audits are deemed inapplicable.

Google emphasizes that, while some aspects of Agentic Browsing may seem unnecessary for current AI search, they are crucial in the context of optimization for generative search engines. To prepare for this new era, Google recommends that developers prioritize the use of semantic HTML, integrate appropriate ARIA labels, and minimize layout shifts.