The Short Answer
When Gemini answers with information from the web, Google grounds the answer in the Google Search index: it gives the model content from Search at the moment the question is asked. So the route into Gemini answers runs through Google Search. Your page has to be indexed and eligible for a snippet, written so a single passage answers a single question, consistent with what the rest of the web says about you, and current. One extra control decides whether Gemini apps may use your pages at all: the Google-Extended token in robots.txt. Block it and you opt out of training and of grounding in Gemini apps, without any effect on your Google rankings.
The one sentence version
Rank in Google Search with pages that answer plainly, keep Google-Extended allowed if you want Gemini apps to use you, and measure the agents that fetch for Gemini users in your own logs.
Which Gemini? Four Surfaces, Different Rules
Gemini is a model family and several products, and the rules depend on which one is asking. People say cited by Gemini about the Gemini app, about AI Overviews and AI Mode inside Google Search, and about notebook tools where users add sources by hand. Google documents each of these separately, and they are controlled by different settings.
| Surface | Where it gets web content | What controls your inclusion |
|---|---|---|
| Gemini app answers | Grounding with content from the Google Search index | Google Search eligibility, plus the Google-Extended token |
| AI Overviews and AI Mode | The Google Search index, as supporting links | Indexing and snippet eligibility; nosnippet and related controls |
| Gemini notebooks | Pages a user adds as sources | Nothing in robots.txt; the fetch is on a user request |
| Agents acting for a user | Live browsing by Google-Agent | Nothing in robots.txt; the visit is on a user request |
| Vertex AI grounding for developers | Grounding with Google Search | The Google-Extended token |
AI Overviews and AI Mode have their own guides on this site, ranking in AI Overviews and optimizing for AI Mode. This guide is about the Gemini app and the fetchers that serve Gemini users, which is where most of the confusion sits.
What Google-Extended Really Controls
Google-Extended is the most misunderstood name in this area. It is not a crawler. Google describes it as a standalone product token that web publishers use in robots.txt to manage whether content Google crawls from their sites may be used to train future Gemini models and for grounding in Gemini apps and in Grounding with Google Search on Vertex AI. Crawling is still done by the ordinary Google user agents. Google-Extended only tells Google what it may do with what it already fetched.
| Question | Answer from Google documentation |
|---|---|
| Does blocking Google-Extended remove me from Google Search? | No. It does not affect inclusion in Search |
| Is it a ranking signal? | No |
| Does it have its own user agent? | No. Crawling uses existing Google user agents |
| What does blocking it stop? | Use of your content to train future Gemini models and for grounding in Gemini apps and Vertex AI grounding |
| Does it control AI Overviews? | AI Overviews and AI Mode are governed by the Search controls: indexing, nosnippet, data-nosnippet, max-snippet |
The trade off nobody mentions
Blocking Google-Extended to keep your content out of training also removes it from grounding in Gemini apps. If you want Gemini to cite you, leave the token allowed. You cannot opt out of training and stay in grounding with this one switch.
Because Google-Extended never appears in a request, any request in your logs that calls itself Google-Extended is fake by definition. We checked ours. Across the two weeks of logs each of our sites keeps, 885 requests carried that name, and 394 of them asked for files such as environment files and private keys. Not one could have been Google. Our crawler list covers how to verify the real Google agents against their published address ranges.
A Robots File That Keeps You in Gemini
Most sites need no Gemini specific lines at all. The default, with no rule naming Google-Extended, allows both Search and grounding. The decisions only arise when a template or a plugin adds AI rules for you, which is common now, so read your file rather than assuming.
| What your robots.txt says | Google Search | Grounding in Gemini apps | Training future Gemini models |
|---|---|---|---|
| Nothing about Google-Extended | Included | Allowed | Allowed |
| Disallow for Google-Extended | Included, rankings unaffected | Not allowed | Not allowed |
| Disallow for Googlebot | Pages cannot be crawled, so no content to show | Nothing from Search to ground on | Not crawled |
| Disallow for every AI token from a template | Included | Not allowed | Not allowed |
The last row is the easiest one to end up with by accident. Lists of AI user agents get pasted into robots files to block training, and Google-Extended is on every such list. The site owner wanted to keep their writing out of model training and, without knowing it, also asked Gemini apps not to use the site when answering. If that was not the intent, remove the line and keep the rest.
No special files required
For the Search side, Google states that you do not need new machine readable files, AI text files or special markup to appear in AI Overviews and AI Mode. Effort spent on a Gemini only file is effort not spent on the page itself.
The Google Fetchers That Serve Gemini Users
Three user triggered fetchers matter for Gemini, and Google says these generally ignore robots.txt, because a person asked for the fetch. That makes them different from crawlers, and it means you cannot block them with robots rules if you wanted to.
| User agent token | What triggers it | Obeys robots.txt? |
|---|---|---|
| Google-Agent | An agent on Google infrastructure browsing and taking actions at a user request | Generally no, user triggered |
| Google-GeminiNotebook | A user adding a page as a source in a notebook project | Generally no, user triggered |
| Google-NotebookLM | The former name of the notebook fetcher, supported until August 2026 | Generally no, user triggered |
| Google-Read-Aloud | A user asking for a page to be read aloud | Generally no, user triggered |
| Google-CloudVertexBot | Crawls requested by a site owner building Vertex AI agents | Yes, and no effect on Search |
Google publishes the address ranges for user triggered agents, and is experimenting with a signed identity for Google-Agent, so these visits can be verified the same way as Googlebot. Verification matters, because the names are copied: requests calling themselves Google-CloudVertexBot reached our servers asking for a web server configuration file, which no agent builder needs.
What We Saw in Our Own Logs
A real agent used one of our tools
On 14 October 2026 a Google-Agent session from inside the published Google range opened a tool page on one of our sites, waited, ran the tool and fetched the result. Thirty nine requests, one task completed for a person we never saw.
Notebook fetchers read like browsers
Requests under the notebook fetcher name, 348 in two weeks, included calls to an endpoint that only the page script makes, which suggests the fetcher runs page scripts rather than reading raw text.
The famous name was always fake
Every one of 885 requests calling itself Google-Extended was an impostor, because the real token never sends a request. Most were scanners.
The first card is the one to plan for. An agent acting for a user does not read your page to quote it. It uses your page. If your tool, form or checkout breaks for a browser that is not a human, the agent fails quietly, and the person it works for never knows you were the problem. The same shift is behind our guide to MCP servers for business, the other route by which agents use a product directly.
How to Get Cited by Gemini, Step by Step
Get indexed and snippet eligible
Grounding draws on the Google Search index, so check the page is indexed in Search Console and carries no nosnippet or tiny max-snippet on the content you want quoted.
Leave Google-Extended allowed
If your robots.txt disallows Google-Extended, Gemini apps cannot use your pages for grounding. Decide deliberately.
Answer in the first lines
Put the direct answer, the number and the date in the opening sentences of the page and of each section, where a passage can be lifted whole.
One question per section
Headings phrased as the question a person asks, followed by a self contained answer, give grounding clean passages to use.
State facts with sources and dates
Specific figures with a method and a date are easier to ground on and harder to contradict than adjectives.
Keep your entity consistent
Use one name, one description and the same key facts across your site, profiles and directories, so the model connects them to you.
Keep key pages current
Refresh pages that answer questions about the present and show the date of the last check near the top.
Make tools work for agents
Test your forms and tools in a clean browser with no saved login, because an agent arrives exactly like that.
Steps three to six are the same discipline that gets a page quoted anywhere, described in more detail in our guide to writing for AI extraction and our guide to entity SEO and the knowledge graph. Gemini does not need a different page from Google Search. It needs the page Google Search would rank, written so a passage can stand alone.
How to Measure Gemini Visibility
There is no Gemini report in Search Console. Google says AI Overviews and AI Mode traffic is included in the Performance report under the Web search type, mixed with ordinary results. For the Gemini app there is no impression count at all, so measurement has to be assembled from three places.
| What you want to know | Where to look | What it can and cannot tell you |
|---|---|---|
| Is the page eligible at all? | Search Console indexing and the URL inspection tool | Indexed and snippet eligible, yes or no |
| Did people click from Gemini? | Analytics referrals from the Gemini domain | Clicks only; answers with no click are invisible |
| Did an agent use the page for someone? | Server logs, Google-Agent and notebook fetchers, verified by address | Fetches and actions, not the answer given |
| What does Gemini say about us? | Asking Gemini the questions your customers ask, on a schedule | The answer today, which changes over time |
Asking is still the most direct check. Pick ten questions a customer would ask, run them monthly, and record which sources Gemini shows. Our guide on getting cited by Claude uses the same method for another assistant, and the comparison is often more useful than either result alone.
Common Mistakes
Blocking Google-Extended by accident
A robots.txt copied from a template that disallows every AI token also removes you from grounding in Gemini apps.
Trusting the name in the log
Requests named Google-Extended or CloudVertexBot are often scanners. Verify by address before counting anything.
Hiding the answer
A nosnippet on the main content, or answers only inside images and scripts, leaves nothing to ground on.
Writing a page for Gemini
There is no Gemini markup and no special file. The page that ranks and answers plainly is the page that gets used.
The Bottom Line
Getting cited by Gemini is mostly getting found by Google Search and being easy to quote once found. Keep pages indexed and snippet eligible, answer early and plainly, keep your facts and your name consistent, and leave Google-Extended allowed if you want Gemini apps to use your work. Then watch the newer visitors. Agents and notebooks fetch pages on behalf of people, ignore robots.txt by design, and judge your site by whether it works, not by how it reads. Make sure it works.