Gemini Agent is Google's agent inside the Gemini app. It takes multi-step actions for a user, such as preparing a car rental or travel booking, handling Gmail and Calendar tasks, filling forms and browsing sites, and it asks before it sends email or buys anything. Google launched it for AI Ultra subscribers in the US on Nov 18, 2025. Gemini Spark, announced at Google I/O on May 19, 2026, is the always-on version that keeps working in the background on dedicated Google Cloud virtual machines, and it reached AI Pro subscribers in the US and more than 160 other countries on Jul 30, 2026.
Across sites that use Verisoul, the google_agent id made up about 1% of the personal-agent sessions we named between Oct 29 and Nov 1, 2026. Google's assistive traffic shows up in much larger numbers elsewhere: Google Read Aloud, a text-to-speech fetcher and not a personal agent, produced about twice as many sessions as Meta Muse in the same window, which is a reminder to keep Google's agents, fetchers and crawlers in separate rules.
Google retired its earlier agent experiment, Project Mariner, on May 4, 2026 and folded its capabilities into Gemini Agent and Chrome auto browse.
How Gemini Agent shows up on your site
When Gemini works from Google's cloud, it uses a remote browser that identifies itself. Google's Google-Agent documentation describes a Chrome on Linux user agent with a compatible; Google-Agent token and a link back to the docs, plus a mobile variant, for agents hosted on Google infrastructure. Google added Google-Agent to its list of user-triggered fetchers on Mar 23, 2026 and publishes the IP ranges in user-triggered-agents.json.
Google also signs some Google-Agent requests with Web Bot Auth as agent.bot.goog, which it describes as experimental and limited to a subset of requests. Spark is different when it drives the user's own desktop Chrome, a mode added on Jul 30, 2026: those sessions use the person's browser and logins and carry no Google-Agent token.
| Check | What Gemini Agent does |
|---|---|
| User agent | Chrome on Linux with a Google-Agent token |
| Web Bot Auth signature | On a subset of requests, as agent.bot.goog |
| IP ranges | Published in user-triggered-agents.json |
| robots.txt | Generally ignored, since a user asked for the visit |
Why Gemini Agent is easy to detect in the cloud
Cloud sessions line up three ways: the user agent token, a published IP list with reverse DNS ending in gae.googleusercontent.com or google.com, and a partial signature. Google itself warns that user agent strings can be spoofed, so the IP or signature check is what turns a claim into proof. A missing signature doesn't prove a fake, because Google only signs a subset.
The blind spot is local mode. When Spark drives a person's own Chrome, there is no Google-Agent token and no Google IP, so the session looks like that person browsing. Google's docs also say Google-Agent covers agents on its infrastructure without naming Gemini Agent or Spark specifically, so treat the id as Google's agent traffic as a whole.
How Verisoul detects Gemini Agent
Verisoul treats Google-Agent as self-declared and then verifies it. The network signal confirms the source IP belongs to Google's published ranges, the device signal confirms a cloud-hosted browser with a virtual machine profile, and the behavioral signal confirms automated clicks, typing and page flow. A Google-Agent string from outside Google's network fails that check.
"agent": {
"id": "google_agent",
"name": "Google Agent",
"provider": "Google",
"type": "ai_agent",
"detection_method": "self_declared",
"detection_signals": [
"network",
"behavioral"
]
}idisgoogle_agent, which covers Gemini's cloud agent sessionstypeisai_agent, an agent taking actions in its own browserdetection_methodisself_declared, meaning the agent announced itself and Verisoul verified itdetection_signalslistsnetwork,deviceandbehavioral
The Agent Atlas entry for Google Agent has the sample response, and the AI agent detection docs explain self-declared verification.
Google Agent was 1% of the personal-agent sessions Verisoul named, Oct 29 to Nov 1, 2026
Share of sessions with a named personal agent on sites that use Verisoul, %. Oct 29 to Nov 1, 2026
How to detect Gemini Agent step by step
- Add Verisoul to sign-up, checkout and account pages so each session returns an agent field and bot score
- Match
agent.idequal togoogle_agent, and keep Googlebot and Google's other fetchers in their own rules - Allow verified
google_agentsessions to browse and buy, and use bot score for Spark sessions in a local Chrome that Verisoul can't name - Review
google_agentsessions on sign-up and payouts, as described in the AI agent detection docs
Should you block Gemini Agent?
Gemini acts for a signed-in Google user and asks before purchases, so most teams allow verified Google-Agent sessions to browse and buy. Blocking Google-Agent also doesn't affect search indexing, because it is separate from Googlebot, so the decision is purely about the agent.
Review sign-up and payouts, where an always-on agent can repeat actions at scale. Spark sessions in a user's local Chrome are not named and fall back to Verisoul's bot score, so keep a score-based rule for those steps too.