Short answer: Gemini SEO comes down to two things you control and one thing you can only measure. You control the declaration: Google documents Google-Extended as a usage control rather than a crawler, deciding whether content Googlebot has already fetched may be used to train Gemini models and to ground answers in Gemini Apps and Vertex AI, and Google states it does not affect Google Search inclusion or ranking. You control the page Googlebot reads. Everything in between happens out of sight, so the effect has to be measured in Gemini's answers, not in your server logs.

That makes Gemini the simplest engine to configure and the hardest to verify from your own side. The audience is large enough to justify the effort: in its Q2 2026 earnings remarks Google said the Gemini app has 950 million monthly active users. This article covers what the declaration does and does not cover, how it sits next to Google's separate controls for Search and AI Overviews, how it compares with the other vendors' switches, what you can still improve on the page, and how to see the effect.

Gemini SEO starts with a declaration, not a crawler

Google-Extended is a token you address in robots.txt, but no agent by that name ever requests your pages. The fetching is done by Googlebot, the same crawler that serves Google Search. Google-Extended only governs what may happen to that content afterwards. The source is Google's list of common crawlers, which is worth reading in the original because pages like it change.

Two practical consequences follow. First, you will not find Google-Extended in your access logs, so there is nothing to confirm on your side after you change the rule. A log search that returns nothing is the expected result, not a sign that the rule failed. Second, access for Gemini is access for Google Search: if Googlebot cannot fetch a page, no Google-Extended setting will put that page into a grounded answer. A site that blocks Googlebot on a section of its site, or returns errors to it, has made the Gemini decision for that section already.

Because the decision leaves no trace in your logs, write it down where the next person will find it, next to the rule:

# Gemini: usage control, not a crawler.
# Decision: ALLOW use for training and grounding in Gemini Apps / Vertex AI.
# Reason: being in grounded Gemini answers outweighs our training concerns.
# Owner: web team. Reviewed: 2026-10. Google Search is unaffected either way.
User-agent: Google-Extended
Allow: /

# To decline instead, replace the group above with:
# User-agent: Google-Extended
# Disallow: /

An explicit group is not strictly required: without one, Google-Extended simply follows your User-agent: * rules. It is still worth writing, because a file with the decision and its reason written out is much harder to change by accident than a file where the decision is an absence. A future editor who sees no Google-Extended line cannot tell whether nobody thought about it or somebody chose it.

Three Google switches over one crawl

The most common Gemini SEO mistake is treating Google-Extended as the switch for everything Google does with AI. Google actually documents three separate decisions, all applied to the same content Googlebot fetched once. Mixing them up produces edits that change nothing you intended and something you did not.

Googlebot fetches the page oncethe only Google crawler involved in any of thisGoogle SearchInclusion and rankingTHE SWITCHGooglebot rules inrobots.txt, or noindexSIDE EFFECTSBlocking Googlebot alsoempties the other twoGeminiTraining + groundingTHE SWITCHGoogle-Extended tokenin robots.txtSIDE EFFECTSNone on Search inclusionor rankingAI Overviews, AI ModeAppearing in and groundingTHE SWITCHSearch Console opt-out(first: some UK sites)SIDE EFFECTSNot a ranking signal forordinary search results
Figure 1. One Googlebot fetch feeds three uses, and each has its own documented switch. Google-Extended governs only the Gemini column. Source: Google's common crawlers documentation and Google's June 2026 post on new controls for website owners.

The third switch is newer than the first two. In June 2026 Google added a Search Console control that lets a site opt out of appearing in, and helping ground, AI Overviews, AI Mode and AI Overviews in Discover. Google says sites that opt out receive no traffic or impressions from those features, that the control is not used as a ranking signal for results outside them, and that it was first rolled out to a subset of website owners in the UK. The announcement is on Google's blog. Nothing in it describes Google-Extended as reaching those features, and nothing in Google's crawler documentation describes Google-Extended as reaching AI Overviews either.

So the question "how do I get out of Google's AI?" has no single answer. It depends on which surface you mean:

If your goal isThe documented routeWhat happens to Google SearchWhat happens in Gemini
Decline Gemini training and groundingUser-agent: Google-Extended + Disallow: /Nothing; inclusion and ranking are explicitly unaffectedRemoved from grounded answers in Gemini Apps and Vertex AI
Stay out of AI Overviews and AI ModeThe Search Console opt-out, where it is available to youNot used as a ranking signal for ordinary resultsNot described as affected; Google's post names only the Search features
Stay out of everything Google showsBlock Googlebot itselfRemoved from SearchNothing left to ground on, since nothing is fetched
Be eligible everywhereAllow Googlebot, allow or omit Google-Extended, no opt-outNormalEligible for grounding

The last row is the one most brands want, and it is worth stating plainly that it is a starting condition, not an outcome. Being eligible means Google may use the page; whether a particular answer names you is decided per question.

How Google's control differs from other vendors' switches

OpenAI, Anthropic, Apple and Google each give you a way to decline training, but they draw the line in different places. That matters when you copy a "block AI training" list written for one vendor and apply it to another.

TRAINING USEAPPEARING IN ANSWERS OR SEARCHOpenAIGPTBotcrawlerOAI-SearchBotcrawler, ChatGPT search resultsAnthropicClaudeBotcrawlerClaude-SearchBotcrawler, Claude search resultsAppleApplebot-Extendedtoken, not a crawlerApplebotcrawler, Spotlight, Siri, SafariGoogleGoogle-Extended: one token for both Gemini usesGemini training and grounding in Gemini Apps and Vertex AI; Google Search is separateTwo boxes: you can decline training and stay in answers. One box: for Gemini, you cannot.
Figure 2. Where each vendor separates training from answers. Google-Extended is the only control that covers training and grounding with one switch. Source: crawler documentation from OpenAI, Anthropic, Apple and Google, checked 2026-09-09.

The difference is the merged box. OpenAI and Anthropic each run a training crawler and a separate search agent, so you can disallow GPTBot or ClaudeBot and keep OAI-SearchBot or Claude-SearchBot. Apple's Applebot-Extended is, like Google-Extended, a token rather than a crawler, but it covers only training Apple's foundation models, and Apple's documentation says pages that disallow it can still appear in Apple's search results. Google-Extended alone covers training and grounding together: you cannot decline one Gemini use and keep the other.

On the other hand, Google offers the cleanest separation from classic search of any vendor. Declining Gemini does not touch your Google rankings, while blocking OAI-SearchBot removes you from ChatGPT search results. Apple's documentation is at Apple's Applebot article; the OpenAI and Anthropic side is covered in how to block GPTBot without disappearing.

Google also documents a second AI-related agent, Google-CloudVertexBot, which crawls a site at the site owner's own request when building Vertex AI Agents. It only matters if you are building an agent on your own content; it is not a general retrieval agent, and it does not change the Gemini decision above. The Google-Extended glossary entry keeps the short definitions together.

Making the Gemini decision

Because the switch is bundled, the decision comes down to one question: is appearing in grounded Gemini answers worth more to you than declining Gemini training? A few points help answer it honestly.

  • Decide per business, not per article you read. A publisher whose revenue depends on its content being exclusive may weigh this differently from a brand that wants to be named when buyers ask about its category. Neither answer is wrong; an unexamined answer copied from someone else's robots.txt is.
  • Know what you are not deciding. Google Search inclusion and ranking are outside this control. A drop in search traffic after a Google-Extended change is not explained by that change, and reverting the rule will not bring the traffic back.
  • Do not stretch the scope. The control as documented names Gemini model training and grounding in Gemini Apps and Vertex AI. AI Overviews are a separate subject with their own control, covered in how to appear in Google AI Overviews.
  • Write the review date into the file. These controls are new and Google has added one in the last few months. A decision recorded with a date is easy to revisit; one without a date ages silently.

Gemini SEO on the page: what you can still improve

Once the declaration is settled, the remaining work is on the pages Googlebot already fetches, and none of it is specific to Gemini. That is an advantage: the same page serves every engine, so work done here is not lost if your priorities shift between assistants.

  • Keep the substance in the HTML. Most AI crawlers do not execute JavaScript. Whatever Googlebot does with scripts, the same page is also read by ChatGPT, Perplexity and Claude agents, so content that appears only after client-side rendering costs you elsewhere.
  • Answer first, with the subject named. Put the direct answer in the first sentence under a heading phrased as the question, so the passage survives being lifted on its own.
  • Write facts, not adjectives. Assistants lift sentences that answer a constraint: a dimension, a material, a return window, a compatibility. Copy written for brand tone gives them nothing to lift, however good it reads.
  • Date anything time-bound. Questions that imply a current answer favour dated, recently updated pages. State a real published and modified date, and stop your build pipeline from bumping it on every deploy.
  • Make the entity unambiguous. Organization markup with your official name, logo and sameAs links to the profiles you control lets you state your identity in a form machines already trust. Treat it as disambiguation, not as a ranking lever. Entity and sameAs explains which links to include.

One more reason not to treat Google's surfaces as one: an Ahrefs comparison across 540,000 query pairs found AI Mode and AI Overviews cited the same URLs only 13.7% of the time, even though their answers were semantically very similar. Two Google products answering the same question can pick different sources, so Gemini deserves its own row in your measurement rather than being assumed from the others.

Measure the effect where it is visible

With Gemini you control the declaration and the page, and you cannot inspect what happens in between. The only place the effect is observable is in the answers, so build the measurement there:

GeminiControlGoogle-Extended rule changed (dated)Run 1Run 2n/aRun 3Run 4Run 5Run 6Baseline: same questions, samemethod, brand-free wordingCounts only if it holdsin two runs in a rowRun 2 timed out: no data, not a 0%
Figure 3. A schematic of an honest before-and-after comparison: fixed questions, a dated change, a control engine, a failed run left blank, and a change confirmed over two runs. Illustration, not measured data.
  1. Use a fixed set of buyer questions that do not contain your brand name, and keep it fixed. Every percentage has the question set as its denominator; adding easier questions raises the average while nothing improves.
  2. Report Gemini on its own before averaging it with other engines. In our own benchmark one model named 3.5 providers per answer and another 2.1, so a single cross-model number describes neither situation.
  3. Treat a change as real only when it appears in two consecutive runs, because answers vary between runs even when nothing changed.
  4. Do not render a failed run as zero. A model that timed out produced no evidence either way; counting it as "not mentioned" invents a decline.
  5. Keep the method fixed across the comparison. Enabling web search once moved a site from 3% to 13% while the site did nothing. If the way you query Gemini changes, start a new baseline rather than drawing one line across both.

If you change the Google-Extended rule, mark the date and compare the runs either side of it, ideally against an engine where nothing changed. Measuring change over time covers how to set that comparison up, and the Gemini engine page shows how Gemini behaved across the categories in our own benchmark.

Sources