What is Google-Agent?
Google-Agent is the identity used by agents running on Google's infrastructure when they navigate a website for a user. Google names Project Mariner as an example of the kind of product that can use it. The person starts the task, then the agent requests the pages needed to carry it out.
That makes its traffic different from a scheduled search crawler. A session may move through several pages as the agent follows the user's instruction and interacts with the site. One request in a log is evidence of an agent visit, but it does not reveal the user's full task or prove that the task finished.
Google publishes desktop and mobile user-agent formats. Both contain the stable Google-Agent token inside a Chrome-like browser string, and the browser version placeholders can change. Rules should match the token rather than a particular Chrome version.
A Google-Agent visit does not add the page to Google Search and is not a Gemini training crawl. Googlebot controls Search crawling. Google-Extended is the separate robots.txt product token for the use of Google-crawled content in Gemini model training and grounding.
The practical effect is on the user's agent session. If a firewall blocks Google-Agent, an agent may be unable to inspect a page or continue the requested workflow even though the same page works in a person's browser. Allowing it improves access for that workflow but does not raise search rankings or create an AI citation by itself.
Google lists the agent in its user-triggered fetcher documentation and publishes its network ranges in user-triggered-agents.json. Cloudflare also classifies it as a Google-operated AI assistant. Check both the token and Google's current network data when distinguishing a genuine request from a copied user-agent string.
