How do I get found on Gemini and AI Overviews?

You have read that AI Overviews are absorbing your clicks, so you go looking for the thing you are supposed to do about it. A file to publish. A markup type to add. A toggle in Search Console. Google's documentation has an answer, and it is the least satisfying answer in this whole series: there isn't one.

That is not evasion. It is a fairly precise statement about where AI Overview supporting links come from. Take it at face value and the actual leverage becomes visible — it just isn't where most people are looking.

Google's stated eligibility rule

Google's AI features documentation puts it flatly, and the sentence is worth reading rather than the summary of the sentence.

There are no additional requirements to appear in AI Overviews or AI Mode, nor other special optimizations necessary... 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.

Google Search Central, AI features and your website

The same page closes the obvious follow-up before you can ask it: “You don't need to create new machine readable files, AI text files, or markup to appear in these features. There's also no special schema.org structured data.”

So the gate is binary and old-fashioned. Indexed and eligible, or not. A stray noindex, an inherited Disallow on Googlebot, a page Google crawled and declined to index — each is a hard disqualification that no amount of AI-flavored optimization repairs. Confirming index status is the dullest possible first move, and it is genuinely the first move.

Query fan-out is where the leverage actually is

Google's AI surfaces do not run the user's query once. They use query fan-out: the system issues multiple related sub-searches across subtopics, then assembles an answer from what those return. That is Google's own description of the mechanic, and it explains something you have probably noticed — AI Overviews and AI Mode cite a broader, stranger set of links than a classic results page. A forum thread. A spec sheet. A comparison page you have never seen rank for anything.

Restated as an operating fact: you are not competing for one query anymore. You are competing for a cloud of sub-queries you never see. They do not appear in Search Console, and no keyword tool returns them, because they are generated at answer time from the user's phrasing. That is the single most actionable idea in this post and also the least comfortable, because it dissolves the tidy keyword-to-page mapping most content plans are built on.

A different engine offers corroboration. A Feb 2025 Ahrefs analysis of 1.4 million ChatGPT prompts found cited page titles scored higher cosine similarity against fan-out-style sub-queries (0.656) than against the original prompt (0.602). That study is about ChatGPT, not Google, and predates mid-2026 ranking changes — but it is evidence that decomposed sub-questions are what retrieval matches against.

What to do about a query you cannot see

The response is not a new tactic. It is a change in what counts as coverage.

• Write the sub-question list yourself. For your most important commercial page, list every question a buyer asks on the way to it: price, prerequisites, alternatives, failure modes, comparisons, who it is wrong for. That list is your best proxy for the fan-out set.
• Answer each one somewhere real. A sub-question that gets a passing clause inside a 3,000-word pillar page competes worse than one with its own page, or at minimum its own headed section.
• Resolve it near the top of the section. Retrieval rewards passages that answer directly rather than building to an answer over four paragraphs.
• Cover the unglamorous subtopics. Fan-out reaches questions with no commercial keyword volume at all, which is exactly why competitors skipped them.
• Stop treating one tracked keyword as the unit of success. A page can be cited for a sub-question it does not rank for, and invisible on a term it ranks well for.

This is the honest reason “just do good SEO” is true but incomplete. Classic SEO optimizes a page against a query you can measure. Fan-out asks for breadth against queries you cannot.

Google-Extended, and how to opt out without disappearing

The second question people arrive with is whether they can keep their content out of Gemini. You can, and there is a right token for it. Google runs several, and they do different jobs.

Google-Extended does not fetch pages of its own. It is a control token, read from robots.txt, that governs how content Googlebot already has may be used. That separation is the point: it lets you decline the training use without touching the indexing AI Overview eligibility depends on.

# Keep full Google Search indexing. AI Overview and
# AI Mode eligibility is derived from Search eligibility,
# so this line is what keeps you in those surfaces.
User-agent: Googlebot
Allow: /

# Decline Gemini model training and Gemini Apps /
# Vertex AI grounding. Google states this token does not
# affect Search inclusion or ranking.
User-agent: Google-Extended
Disallow: /

Make the call deliberately rather than by inheritance. A publisher whose archive is the product has a real argument for disallowing Google-Extended. A services business hoping Gemini recommends it does not — the same token also blocks Gemini Apps grounding, which is a live retrieval path, not just a model-weights question.

The five myths Google itself debunks

In May 2026 Google Search Central published “Optimizing your website for generative AI features on Google Search” (2026-05-15, updated 2026-06-24). Its framing sentence is “optimizing for generative AI search is optimizing for the search experience, and thus still SEO,” and it names five things people do that Google says do not work.

1. You need llms.txt or AI-specific markup. Not required, per Google. Outside evidence agrees: an Ahrefs study of 137,210 domains using May 2026 data found 28% published a valid llms.txt, and 97% of those files received zero requests during the study period.
2. Content must be chunked into small pieces. Google says its systems handle long-form content fine.
3. You need an AI-specific writing voice. Also not required. No house style reads better to a retriever than clear writing does.
4. Planting brand mentions across the web lifts AI citations. Google says its spam systems filter inauthentic mentions.
5. Structured data is a required generative-AI ranking factor. Google's wording: structured data isn't required for generative AI search. It stays useful for classic rich results, which is a separate reason to keep it.

How often do AI Overviews actually appear? Nobody agrees

You will see a prevalence number quoted as though it were a fact about the internet. Here are four, all current, all from credible trackers.

• BrightEdge: AI Overviews on 48% of tracked queries as of Feb 2026, up 58% year over year.
• Advanced Web Ranking: 60.32%.
• Conductor: 25.11%.
• Semrush: 15.69%.

A spread of nearly four times, and none of them is lying. They track different keyword sets, locales, device mixes and sampling cadences, and AI Overview triggering is sensitive to query type — informational questions fire them far more than transactional ones.

The lesson is not to pick the right number. It is to distrust any post, this one included, that hands you a single AI Overview prevalence figure without naming the keyword set behind it. The only rate with operational meaning is the one for the queries that feed your business.

The click math, stated honestly

Prevalence matters because the click behavior underneath it has genuinely shifted.

Seer Interactive's series is the most useful of the three because it has a second act. That 61% collapse in organic CTR partially recovered to roughly 2.4% by Feb 2026 — above where it started. Paid CTR fell harder over the same window, 19.70% to 6.34%. Do not read the recovery as the problem resolving itself; read it as a results page settling into a shape users have learned to navigate.

The zero-click figure carries a caveat SparkToro states itself: methodology shifted across the years in the series (2019 50.33%, 2020 64.82%, 2024 60.45%, 2026 68.01%). The direction is solid; the deltas are softer than they look.

The number to plan around is the third. One analysis attributed to Surfer SEO, covering 46 million citations across 36 million AI Overviews, found cited brands earn about 120% more organic clicks per impression than uncited competitors on comparable queries. It is a vendor analysis, so hold the magnitude loosely. The direction inverts the old hierarchy: presence inside the AI Overview is now worth more than the blue link beneath it.

Keep the base rate in view. Chartbeat and Axios data published 2026-03-17 put Google Search referral traffic to publishers down 34% year over year — small publishers down 60%, large ones down 22% — while chatbot referrals still accounted for under 1% of total publisher pageview referrals. Gemini visibility is not a replacement traffic channel yet. It is a defense of the channel you already have.

Gemini is now an Apple surface too

One development changes the size of the prize. Apple confirmed at WWDC in June 2026 that its next-generation Siri is powered by Google Gemini under a commercial deal — first reported by TechCrunch on 2026-01-12, confirmed publicly that April. Work you do for Gemini visibility now reaches a surface you previously had no path into.

Two caveats. Apple's own tokens, Applebot and Applebot-Extended, still govern Spotlight, Siri, and Safari indexing and Apple foundation-model training separately, and Apple has published no guidance on how they interact with a Gemini-backed Siri. And Siri's answer layer being Gemini's does not mean its citation behavior matches AI Mode's — nobody outside the two companies has data on that yet.

The checklist

1. Confirm you are actually eligible. Open the Search Console Pages report. Anything in Excluded that you meant to be indexed is a page that cannot be cited, full stop. Then read robots.txt for a Googlebot rule nobody has looked at in years.
2. Write the sub-question list for your three most important pages. Twenty questions each, phrased the way a buyer says them out loud. Find where each is answered on your site. The gaps are the work.
3. Give the unglamorous sub-questions a home. Own page or own headed section, answer in the first sentence. Zero search volume is not a reason to skip one.
4. Make the Google-Extended decision on purpose. Decide whether you want your content training Gemini and grounding Gemini Apps, set robots.txt to match, record the reasoning. Do not block Googlebot.
5. Measure prevalence on your own queries. Run your twenty most valuable queries by hand, record whether an AI Overview fires and who it cites, repeat monthly. That beats every published percentage.
6. Stop chasing special files. No llms.txt, no AI schema type, no AI-voice rewrite. Every hour there is an hour not spent closing the coverage gap fan-out rewards.

The disappointment at the top of this post is also the good news. No secret is being withheld from you. There is a wider surface to cover than there used to be, the queries on it are invisible, and the people who map them out by hand will be cited by the people who don't.