Gemini answers that lean on Google Search are assembled from pages Google has already crawled and indexed. That single fact does most of the work in Gemini SEO: you are not optimizing for a separate Gemini index, because for grounded answers there generally isn't one. You are trying to be the page Google retrieves and then quotes.
What grounding actually means
When Gemini answers with search grounding, it runs a query, pulls candidate pages from Google's index, reads them, and writes a response that cites some of them. The retrieval step looks a lot like a search results page. The generation step looks a lot like a summarizer deciding which sentences are worth repeating.
That has a practical consequence. If a page is not indexed, it cannot be grounded. If it is indexed but thin, it can be retrieved and still not cited, because the model has nothing quotable to lift. Gemini citations reward pages that say something specific in a form that survives being copied out of context.
This is different from how other assistants assemble answers. ChatGPT search has its own crawler story, which we covered in How ChatGPT Search Picks the Sources It Cites. Google AI Overviews sit inside the ordinary search results page and follow normal Googlebot rules, a contrast worth understanding if you are allocating effort across surfaces: Google AI Overviews vs ChatGPT: Two Different Games.
Google-Extended is narrower than most people assume
This is the most misunderstood control in the stack. Google-Extended is a robots.txt user-agent token that lets you decide whether your content is used to train future Gemini models and for grounding in Gemini apps and the Vertex AI Gemini API. Per Google's documentation, it does not affect inclusion in Google Search, and it is not a ranking signal.
So blocking Google-Extended does not remove you from search results. It does not remove you from AI Overviews, because those are part of Search and governed by Googlebot. What it does is opt you out of a specific set of Gemini training and grounding uses. If your goal is Gemini citations, blocking Google-Extended works against that goal for the grounded surfaces it covers, while doing nothing to protect or improve your ordinary rankings.
Read the token's scope directly rather than trusting secondhand summaries: Google's documentation on Google-Extended.
The trap in "just block the AI crawlers"
A common move is to block every AI-related user agent in robots.txt as a blanket policy. That is a legitimate choice, but it should be a deliberate one. Google-Extended, GPTBot, ClaudeBot and PerplexityBot sit in different systems with different consequences. Blocking training collection is not the same as blocking retrieval for live answers, and some vendors split those into separate tokens precisely so you can choose.
Anthropic, for instance, separates ClaudeBot (training), Claude-User (fetching pages when a user asks a question) and Claude-SearchBot (indexing for search). Perplexity separates PerplexityBot (surfacing sites in results) from Perplexity-User (fetching for a user question, which generally ignores robots.txt because the user asked for it). If you want to appear in answers, you need the retrieval and indexing tokens allowed, not just the training token blocked.
How Gemini picks which pages to cite
There is no published ranking formula for grounded Gemini answers, and anyone who claims to have reverse-engineered one is guessing. What is publicly documented is the retrieval substrate: Google's index. What is observable is the shape of the answers.
In practice, cited pages tend to share a few traits:
- They answer the question in the first screen. A grounded answer needs a sentence it can lift. If your answer is in paragraph nine, another page supplies it.
- They are unambiguous about the entity. The page names the product, company or concept plainly, in text, not only in an image or a logo.
- They carry structure. Headings, tables, short lists and definition-style sentences give a summarizer clean units to quote.
- They are crawlable and indexable. No accidental noindex, no JavaScript-only rendering of the core answer, no paywall in front of the part that would be cited.
- They are current. A page with a visible, accurate date and recent substance competes better than one that looks abandoned.
Notice that none of this is Gemini-specific trickery. It is the same set of properties that make a page quotable by a person skimming search results.
Grounding versus training
It helps to keep two questions separate. "Will my content be used to train future models?" is a training question, governed by tokens like Google-Extended and GPTBot. "Will my page be retrieved and cited when someone asks a question right now?" is a retrieval question, governed by Googlebot and the indexing pipeline for Gemini, and by other vendors' search and user tokens elsewhere.
You can answer yes to the second and no to the first. Many publishers do exactly that: allow search crawling, decline training. That position is coherent, and it is the one most brands that want citations should be considering.
A practical checklist
Work through these in order. Most of the value is in the first four.
- Confirm the page is indexed. Search for a distinctive sentence from the page in Google. If it does not appear, fix indexing before anything else.
- Check robots.txt against the right tokens. Verify that Googlebot is allowed. Decide deliberately about Google-Extended, and remember it does not govern Search inclusion.
- Put the answer at the top. One or two sentences that directly answer the likely question, before the background.
- Add quotable structure. A short definition, a comparison table, a numbered process. Each is a candidate for extraction.
- Name things plainly. Entity names in body text, consistently, so a summarizer can attribute the claim.
- Show recency honestly. Update the page when the facts change and reflect that in the content, not just a timestamp.
- Test with real prompts. Ask Gemini the questions your buyers ask. Note which pages get cited and what those pages do that yours does not. This is the loop that actually moves Gemini citations.
If you want to track that loop over time rather than checking by hand, PromptRanking monitors whether assistants including Gemini recommend a brand, which turns spot checks into a trend line.
What to do next
Pick one page that should be cited for a question your buyers actually ask. Run that question through Gemini, see whether you appear and who does, then rewrite the opening of your page so the answer is in the first two sentences and the supporting detail is in clean, quotable blocks. Recheck in a week. One page, one query, one rewrite is a better use of an afternoon than a sitewide audit.