What does Gemini look at when citing a source?

For AI Overviews and AI Mode, Google's own documentation says a page must be indexed and eligible to be shown in Google Search with a snippet — and that there are no additional technical requirements. Google-Extended is a separate control for the Gemini assistant, and Google states it does not affect inclusion or ranking in Google Search. Every documented lever on this surface removes you; none adds you.

Almost every argument about Gemini is two people talking about two different products. One is the Gemini app and the Gemini API — an assistant that pulls pages in at the moment you ask. The other is Google Search itself: AI Overviews and AI Mode, sitting on top of the ordinary search index. Same brand, different plumbing, different controls. This matters in Answer Engine Optimization (AEO) — engineering a site so AI assistants quote it — because the advice you were given for one surface is often being applied to the other.

Watch the walkthrough

Why does the failure with Gemini feel so quiet?

Because nothing tells you it happened. With ChatGPT the failure is loud: a crawler is blocked and you are simply not in the running. With Gemini you do everything the internet told you to do — add the markup, build the entity, publish more — and your dashboard still shows one flat, low number. It shows that number because the dashboard you own cannot separate an AI answer from an ordinary blue link. Nothing in it tells you that you optimised for a mechanism that was never there.

What does Google-Extended actually control?

The assistant surface, and only that. Google's crawler documentation describes Google-Extended as a standalone product token publishers use to manage whether content Google crawls may be used for training future generations of Gemini models and for grounding — providing content from the Google Search index to the model at prompt time. That is the Gemini app and the API. And on the same page comes the sentence that ends most arguments: Google-Extended does not impact a site's inclusion in Google Search, nor is it used as a ranking signal in Google Search. So blocking it does not take you out of AI Overviews, and allowing it does not put you in. If somebody sold you that toggle as an AI-visibility fix, they sold you the wrong row.

What does allowing Google-Extended actually buy you?

A door left open, not a result. Our own robots.txt allows Google-Extended — we verified that again on 25 August 2026 — and we want to be precise about what that means: we have not opted our pages out of the pool the assistant grounds on. It does not mean we get cited. That is the honest description of every access check, including the OAI-SearchBot check in the previous episode. Access is a precondition. It is never a promise.

What is the official requirement for AI Overviews?

Two conditions, and that is the whole list. Google's own page about AI features says that to be eligible to be shown as a supporting link in AI Overviews or AI Mode, a page must be indexed and eligible to be shown in Google Search with a snippet, fulfilling the Search technical requirements. Read it as a checklist: indexed, snippet-eligible. The very next sentence in the same paragraph is the one that saves you money — there are no additional technical requirements. Not a longer list you have not found yet. None.

Do you need special schema markup to appear?

No, and Google says so on that same page: you don't need to create new machine-readable files, AI text files, or markup to appear in these features, and there's also no special schema.org structured data that you need to add. Be careful what you do with that, though — this is not permission to rip your markup out. Structured data is still worth having: Google's own best-practice list on the same page asks you to make sure your structured data matches the visible text, and markup still earns rich results in ordinary search. What it is not, is the thing that gets you into an AI answer. Hygiene, yes. Cause, no.

What is query fan-out and what does it change about writing?

It is the mechanic on that page worth more than every piece of markup advice put together. Google states that both AI Overviews and AI Mode may use a query fan-out technique — issuing multiple related searches across subtopics and data sources — to develop a response. So the question a person typed is not the query your page has to win: the engine splits it into sub-questions you never see and searches those instead. That changes the shape of the work. One page straining to rank for one big phrase is the wrong shape. Take the question your customer actually asks, write down the five smaller questions hiding inside it, and answer each one explicitly, in its own short section, in plain words. You are not writing for one query any more. You are writing for the fan.

Why does fan-out matter more outside English?

Because those sub-questions get asked in the language of the person asking them, and the non-English fans are visibly less answered. On our last tracker run the one Gemini answer that named us at all came from a Polish query, not an English one — that is one cell out of fifty-six, and we are not going to dress up one cell as a trend. But the pattern behind it is checkable by anyone: Polish and German sub-questions in our category have thinner competition than the English ones. For a small business selling in Poland or the DACH region, that is a real opening — and it is a claim about the market, not about us.

Can Search Console tell you whether any of it worked?

Not separately, and Google says so plainly. On the same page: just like the rest of the search results page, sites appearing in AI features such as AI Overviews and AI Mode are included in the overall search traffic in Search Console, and in particular they are reported on in the Performance report within the Web search type. Included — not separated. Your own dashboard folds the AI answer into ordinary web traffic and hands you one number for both. That single fact explains most of the frustration people have with this surface: you can genuinely improve and be unable to prove it from the report you own.

Which controls actually exist on this surface?

Four, and every one of them takes something away. To limit the information shown from your pages in Search, Google's list gives you nosnippet, data-nosnippet, max-snippet and noindex. Look at what that list is made of. There is no opposite list on that page — no tag anywhere on it puts you in. That is the whole design of the surface, and it is why the standard advice to add something is so often wrong here. The practical instruction is the boring one: make sure none of those four is set anywhere you did not mean it, because a max-snippet left over from an old template can quietly disqualify you from the very appearances you are trying to earn.

What does our own measurement show — and where does it stop?

One finding we trust, and one boundary we have to name in the same breath. Across four runs on our own account this summer, youtube.com appears in Gemini's cited sources every single time: 9 of 222 citations, 7 of 209, 11 of 224 and 32 of 713 — between three and five percent, four times running. On the same runs, in OpenAI's citations, zero out of 223, and in Anthropic's, zero out of 426. That difference between the engines is real, and we re-ran the numbers on 25 August 2026 before publishing this. Now the boundary: all 59 of those citations are the bare domain. Zero of the 59 carry a path — not one names a video, a channel or a creator. So we cannot tell you that a video got cited, and we cannot tell you a channel got cited. We can tell you one string keeps appearing in one engine's source set and never in the other two, and that our own extraction cannot resolve it any further. If you have heard anybody say "post videos and Gemini will cite you", this is the paragraph that should stop it. We publish these videos anyway. We just do not get to call it a measured lever.

Where did our own tracker mislead us?

It picked the flattering table, and we are naming it rather than quietly patching it. On that same run our own tracker wrote us a task: publish a YouTube video, because this leverages the number one most cited source, YouTube, 32 times. Open the other table in the very same file, though — ranked by host, youtube.com is third, at 32, behind sostav.ru at 39 and clutch.co at 38. One file, one dataset, two rankings, and our recommendation text silently chose the one that supported the recommendation. It is number one only when you rank by the canonical URL form. That goes on the board as a defect, because a tool that writes you tasks has to be honest about which table it read.

What is the Gemini checklist?

Five checks, and notice how little of it is about Gemini. One: the page is indexed and snippet-eligible — that is the entire official requirement and it is the one almost nobody checks. Two: no nosnippet, no max-snippet, no data-nosnippet and no noindex anywhere you did not mean it, because those are the only switches on this surface and they only ever remove. Three: decide Google-Extended deliberately — allow it if you want to be in the assistant's grounding pool, block it if you do not, and know it changes nothing in Search either way. Four: write for the fan-out, with the five sub-questions inside the question each answered in its own short section, in the language your customer asks in. Five: measure per engine, and say out loud where the measurement stops.

Frequently asked questions

Does blocking Google-Extended remove me from AI Overviews?

No. Google states that Google-Extended does not impact a site's inclusion in Google Search, nor is it used as a ranking signal there — and AI Overviews sit on the Search index. Google-Extended governs the assistant surface: training future Gemini models and grounding at prompt time. Decide it on whether you want to be in that grounding pool, not as an AI-visibility fix.

Do I need special schema.org markup for AI Overviews?

Google's own page says no — no new machine-readable files, no AI text files, and no special schema.org structured data. Keep your markup anyway: it earns rich results in ordinary search, and Google asks that it match the visible text on the page. Treat structured data as hygiene rather than as the cause of an AI citation.

How do I write for query fan-out?

Stop writing for the one phrase you want to rank for. Take the question a customer actually asks, list the five smaller questions inside it, and answer each explicitly in its own short section with a plain heading. Write them in the language the customer asks in, because the fan-out searches happen in that language too.

Can I see AI Overviews traffic separately in Search Console?

No. Google says appearances in AI features are included in the overall search traffic and reported in the Performance report within the Web search type. They are folded in, not split out. If you need per-engine visibility you have to ask the engines your buyers' questions directly and keep the list of sources each one cited.

Does publishing YouTube videos get you cited by Gemini?

Nobody can honestly tell you that from data like ours. Across our four runs youtube.com appeared in every Gemini source set and in neither of the other two engines, which is a real cross-engine difference. But all 59 of those citations are the bare domain with no path, so none of them names a video or a channel. Publish videos for the audience; do not sell it as a measured Gemini lever.

Want these five checks run across your site?

We run all five — index and snippet eligibility, accidental preview controls, the Google-Extended decision, fan-out coverage in your customers' language, and per-engine measurement with its limits stated — and the first AI-visibility audit is free. If you want to see what ChatGPT, Gemini and Claude currently answer about your category, request a free audit at webappski.com.