How to rank in Gemini: what Google says decides the sources it shows

How to rank in Gemini, in short

To rank in Gemini, rank in Google Search for the questions your buyers ask, and do not block Google from reading you. Gemini’s web answers are built from Google’s own search index, so there is no second index to submit to.

Google says this in plain words. Its page on AI features and your website says a page must be “indexed and eligible to be shown in Google Search with a snippet” to be a supporting link in AI Overviews or AI Mode. The same page says there are “no additional requirements” and no “special optimizations necessary”.

That sounds like nothing has changed, and much has not. Two things have. The word Gemini now covers several products that read the web in different ways. And there is a robots.txt switch, Google-Extended, that touches some of them and not others. Most confusion about ranking in Gemini comes from mixing those up.

Which Gemini you mean

Gemini answers reach people through at least four Google products. They share the Google Search index, but they are controlled by different settings.

The Gemini app says little about how it picks links. Google’s help page on sources in Gemini Apps says the app “may provide links to sources and related content”, and that “not all responses include related links or sources”. It also says that if Gemini quotes a large amount of text from a webpage, that page appears in the sources list.

The developer version is better documented. Google’s page on Grounding with Google Search says the model “analyzes the prompt and determines if a Google Search can improve the answer”, then writes one or more searches, runs them, and returns its answer with citations to the pages it used. That matters because many tools that track AI answers ask Gemini through an API like this one, and its answer can differ from the app’s.

What Google-Extended controls, and what it does not

Google-Extended is a name you can use in robots.txt, not a separate robot. It tells Google whether your pages may train future Gemini models and ground answers in the Gemini app, and Google says it has no effect on Google Search.

Robots.txt is the plain text file at the root of a site that tells robots what they may read. Google’s common crawlers page says Google-Extended controls whether content “may be used for training future generations of Gemini models” and “for grounding” in Gemini Apps and in Grounding with Google Search on Vertex AI. Grounding, in Google’s words, means “providing content from the Google Search index to the model at prompt time”.

The same page says Google-Extended has no visitor of its own. Google’s ordinary robots do the reading, and the name only acts as a control. It also says Google-Extended “does not impact a site’s inclusion in Google Search nor is it used as a ranking signal”.

So blocking Google-Extended is a decision about training and the Gemini app, not about AI Overviews. A publisher who does not want its pages used to train Gemini can block it and stay in Search. The cost is that the Gemini app may no longer ground its answers in those pages. Make that choice on purpose, and write down why, because a later reader of your robots.txt will not know.

Our Gemini page quotes Google’s crawler documentation in full. Our page on AI crawler access explains how a robots.txt rule can block a robot without anyone meaning to.

What decides whether Gemini picks your page

Google says its AI answers are rooted in the same ranking systems as Search. The answer is written from pages that Search retrieves for the question, and for several smaller questions the model asks on its own.

Google’s guide to optimizing for generative AI features names two techniques. The first is retrieval-augmented generation, also called grounding: Search finds relevant pages, and the model writes its answer from what they say. The second is query fan-out, which Google describes as “a set of concurrent, related queries generated by the model”. Its own example is a question about a lawn full of weeds, which fans out into searches about herbicides, weeding without chemicals and preventing weeds.

Fan-out explains something people notice in AI answers. The cited pages are not always the ones that rank first for the question as typed. A page can be retrieved because it ranks for one of the smaller questions behind it. A buyer asking which payroll tool suits a small shop may set off searches on prices, on setup time and on one rival’s reviews. A page that answers one of those clearly can be cited even if it never ranked for the original question.

Google also warns against the obvious overreaction. Its guide says that making separate pages for every variation of a query, mainly to manipulate AI answers, “violates Google’s scaled content abuse spam policy”. One good page that covers the main question and its follow-ups beats twenty thin ones.

On what makes a page worth picking, Google’s guide says unique, compelling and useful content “will likely influence your website’s presence in generative AI search in the long run more than any of the other suggestions”. It contrasts commodity content, such as a generic list of tips anyone could write, with first-hand content that adds something only you know.

Gemini can only cite a page that Google can index and show with a snippet. Check that first, before you rewrite a single word.

The steps, in order

Start with access, then answer the questions, then measure. The steps below follow Google’s own guidance, linked in the sources.

1. Confirm your key pages are indexed. Verify your site in Google Search Console, which Google recommends for finding technical problems quickly.
2. Read your robots.txt. Make sure Googlebot is allowed, and decide about Google-Extended on purpose rather than by default.
3. Look for snippet settings left on by accident. A nosnippet or a small max-snippet on a template can quietly limit every page that uses it.
4. Write down the questions a buyer asks before they buy, in their words. Those are the questions to win, not the keywords you already rank for. Our guide to finding buyer questions shows where to look.
5. Answer each question near the top of a page, in plain sentences, with the facts a model can lift: a price, a limit, a date, a name.
6. Cover the follow-up questions on the same page. Think of what fan-out would search for next, and answer it in its own short section.
7. Add what only you know. A test you ran, a number you measured or a case you handled is the opposite of commodity content.
8. Keep Merchant Center and your Business Profile current if you sell products or serve a local area. Google says these feed AI answers too.
9. Earn real mentions on other sites. Google says inauthentic mentions are not as helpful as they might seem, because its spam systems sit underneath AI answers.
10. Measure, change one thing, and measure again.

None of this is new advice. Google’s May 2025 post on succeeding in AI search says that “meeting the technical requirements covers you for search generally, including AI formats”. What is new is that the question set is wider, because buyers ask assistants longer questions and follow-ups.

What you can skip, according to Google

Google lists several popular tactics that it says Google Search ignores. That is a statement about Google, and it may not hold for other assistants.

The guide also warns about tools. It says no third-party tool “has access to our internal ranking or AI systems”, and asks site owners to be wary of any that claim to. That applies to every tool in this market, including the one that wrote this post. What a tool can honestly do is ask the questions and record what came back.

How to see whether it worked

For AI Overviews and AI Mode, use the generative AI performance report in Search Console. For the Gemini app, nothing reports it to you, so you have to ask the questions yourself and record the answers.

Google announced generative AI performance reports in Search Console on 3 June 2026, and a note on that post says they reached all websites on 31 August 2026. The announcement says the reports show how often your URLs appeared in generative AI features in Search and Discover, which pages appeared, and the country, device and date. It does not mention the Gemini app.

For the Gemini app, the honest method is slow. Write your list of buyer questions. Ask each one in Gemini, with no history and no signed-in context if you can. Save the whole answer and every link. Then ask again the next week, because answers change from one run to the next. One answer is a sample of one, and our method page explains why a single answer should never be read as a ranking.

Agentcites is one way to do this on a schedule. Today it asks Gemini, with Google Search switched on, alongside ChatGPT, Perplexity and DeepSeek. It asks up to 50 buyer questions for each website every week, keeps every answer word for word with the links it gave, and reports questions that name you apart from ones that do not. It costs $149 a month per website, as the pricing page shows. It asks through the API, so its answer is Gemini’s answer to the bare question, and a person in the app with their own history can see a different one.

What this guide cannot tell you

Google publishes the rules for access, not the weights for ranking. Nobody outside Google can tell you exactly why Gemini chose one page over another, and a guide that claims to is guessing.

What you can know is narrower and more useful. You can know whether Google can read you, whether your page answers the question a buyer asks, and what Gemini said this week compared with last week. Those three things are enough to act on.