You get cited in Google AI Overviews by ranking for the sub-queries Gemini generates and writing passages it can lift whole. Most citations do not go to the highest-ranking pages, so ranking in AI Overviews is a passage-level job that sits inside SEO rather than beside it, and it pays to understand how the summaries are built before targeting them.
What are AI Overviews and how do they work
An AI Overview is the answer block Google generates on the results page, with inline links to its sources. Google reported more than 2.5 billion monthly AI Overview users at Google I/O 2026, so those citation slots reach a large audience. Google generates AI Overviews with a customized Gemini model that runs on top of its ordinary ranking systems. Google's Liz Reid described the arrangement on May 30, 2024: "While AI Overviews are powered by a customized language model, the model is integrated with our core web ranking systems and designed to carry out traditional 'search' tasks, like identifying relevant, high-quality results from our index." Google said on January 27, 2026 that Gemini 3 is the default model for AI Overviews globally.
The pipeline starts with a trigger decision. Overviews "often don't trigger," Google's documentation says, because one must be "additive to classic Search." When one does, the model may use query fan-out, which Google's Search Central docs define as "issuing multiple related searches across subtopics and data sources" to develop a response. Google adds that its advanced models identify more supporting web pages while generating a response, which widens the cited set beyond the original SERP. This resembles a retrieval-augmented generation loop, although Google has not disclosed the full production implementation. The model retrieves pages, then writes from them. Reid wrote in May 2024 that Google designed AI Overviews to show only information backed by top web results.
What determines whether a page gets cited
Indexing and snippet eligibility put a page into consideration. Organic ranking and coverage of fan-out sub-queries then improve its citation odds.
The organic ranking correlation
Google's only stated requirement is that Google index the page and allow it to appear in Search with a snippet. Google adds that "there are no additional technical requirements." Past that bar, citations favor pages that rank for the query or one of its sub-queries, and the unit that competes is the passage, not the page. Deep pages beat homepages. BrightEdge's analysis of AI Overview citations found 82.5% pointed to pages two or more clicks from the homepage, and 0.5% pointed to homepages.
Top-10 ranking raises citation odds and still leaves most citations unexplained. Conductor's vendor-published September 2026 analysis of 167.9 million U.S. citations (June 15 to August 15, 2026) found 43% came from pages in the organic top 10 and 57% from outside it.
seoClarity's vendor-published October 2025 sample of 362,000 U.S. desktop queries found 94% of Overviews contained at least one URL from the top 20, so organic fundamentals still gate most citations. Pages outside the top 20 still get cited, so treat top-20 presence as improving the odds rather than determining eligibility. Audit your ranking coverage for the head term and its sub-queries before you spend a sprint on page formatting. Gary Illyes said in July 2025 that normal SEO practices apply to AI Overview visibility.
Query fan-out and topical coverage
Fan-out accounts for much of the citation traffic you can't trace to a top-10 rank. When Gemini decomposes "how to rank in ai overviews" into technical and measurement sub-queries, a page ranking for one of those becomes a candidate without ranking for the head term. A vendor-published Surfer test from December 2025, covered by Search Engine Land, ran 10,000 keywords through Gemini to generate 33,000 fan-out queries and studied 173,902 cited URLs against both SERP sets. The test found that pages ranking for fan-out queries had a 161% higher citation likelihood than pages ranking only for the main query, and the Spearman correlation between the number of fan-out queries a page ranked for and its citation likelihood was 0.77. Even so, about 68% of cited pages ranked in the top 10 for neither the main query nor any fan-out query, so coverage increases the candidate set without explaining every citation.
That favors a content cluster over one exhaustive page. A pillar page for the head query, surrounded by spokes that each rank for a sub-query, puts more of your URLs into the positions Surfer tied to that 161% lift. Reid described the Gemini 3 fan-out upgrade in November 2025, before Gemini 3 became the AI Overviews default the following January, saying it "can perform even more searches to uncover relevant web content" and, because Gemini understands intent better, can surface content the system previously missed.
How to structure a page for AI extraction
Put the answer in the first 100 words, then earn the rest of the page. Roast.page's April 2026 correlational comparison of 200 cited pages against 197 uncited controls across 142 Overviews found that 87% of cited pages gave a direct answer in the first 100 words. The comparison also associated FAQ-style or Q&A headings with 3.4 times as many citations.
Ma, Qin, Xu, and Tan's 2025 Gemini RAG experiment across more than 10,000 websites is the closest controlled evidence: a one-standard-deviation drop in perplexity (a measure of how predictable the text is to the model) lifted citation probability from 47% to 56%. So cut ornamental phrasing before you cut word count.
Hand a writer this template:
- Direct answer capsule: Open each H2 with a two-to-four sentence answer that makes sense quoted alone, with the subject named rather than pronoun-referenced.
- Question-shaped headings: Write H2s and H3s as the conversational sub-queries a buyer would type, so each section maps to a plausible fan-out query.
- Lists for enumerable facts: Break enumerable information into list items rather than a running paragraph. Surfer's August 2026 study of 405,576 English-query Overviews found 78% of the answers contain lists, which describes the answers rather than the source pages.
- Key-takeaway blocks: Where a section runs long, close it with a short self-contained paragraph the model can lift.
Length on its own adds little. Ahrefs' vendor-published December 2025 review of 174,048 cited pages found a 0.04 word-count correlation with citation and 53.4% of cited pages under 1,000 words. Google's May 2026 guide cautions against overfitting: "You don't need to write in a specific way just for generative AI search."
E-E-A-T, entity signals, and off-site authority
Google runs no separate trust system for AI Overviews. Its documentation says you "can apply the same foundational SEO best practices for AI features as you do for Google Search overall," and Danny Sullivan said in December 2025 that "Generative Engine Optimization (GEO) isn't separate from SEO," calling it a subset of SEO. The rater guidelines updated September 11, 2025 added AI Overview examples without changing rating guidance, and they keep Trust at the center: "untrustworthy pages have low E-E-A-T no matter how Experienced, Expert, or Authoritative they may seem."
Trust still shows up in outcomes. Xu, Iqbal & Montgomery's independent 2026 study of 55,393 U.S. trending queries found cited domains carried a mean credibility score of 0.732 against 0.645 for matched first-page URLs, a statistically significant skew toward more credible sources. On-page, that means:
- Named author and About page: Publish a byline with credentials a reader can verify. State who the company is and what it sells.
- Traced claims: Link every figure to its origin.
Sullivan's caveat from February 2024 applies: "Having an expert write things doesn't magically make you rank better."
Google says AI Overviews work "in tandem with our existing Search systems," naming its quality and ranking systems and the Knowledge Graph, but no Google statement names Knowledge Graph alignment as a source-selection factor, and no published research has tested it. Remove ambiguity instead by naming your company and products consistently across your site and third-party profiles, including social accounts.
Off-site authority warrants attention alongside on-page work. Digital PR coverage from editorially reviewed publications and YouTube explainers put your brand in off-site contexts related to the credibility skew Xu, Iqbal & Montgomery measured, though that study did not test whether earned coverage moves a domain's score. Google said in May 2024 it had "updated our systems to limit the use of user-generated content in responses that could offer misleading advice," a limit on user-generated sources that could mislead.
Schema markup and structured data
Google does not require schema for generative search, and one matched-panel study found a small citation decline after pages that already had citations added it. Google's AI optimization guide says structured data "isn't required for generative AI search, and there's no special schema.org markup you need to add." A Search Engine Land report covered the Ahrefs matched-panel study published May 11, 2026, which compared 1,885 pages that added JSON-LD between August 2025 and March 2026 against 4,000 controls. AI Overview citations fell 4.6% in relative terms, small but statistically significant.
Structured data labels what a field on the page means, which is how Google says it makes pages eligible for certain search features and rich results. So if you already run Article or VideoObject JSON-LD, keep it accurate against the Search feature docs that define its use. Google has not identified either type as an AI Overview citation factor.
FAQPage no longer earns a rich result either. Google restricted FAQ rich results in August 2023 and stopped showing them entirely on May 7, 2026. Keep the JSON-LD accurate wherever it maps to a live feature, and don't budget schema work as AI Overview optimization.
Technical foundations for crawlability, Core Web Vitals, and freshness
Indexing and snippet eligibility are prerequisites. No Google document ties Core Web Vitals directly to citation selection. Google warns that recrawling "can take anywhere from several days to several months," so preview-control changes register slowly. A blocked or noindexed page is invisible to fan-out retrieval no matter how well it's written.
Core Web Vitals sit in the same category. Google's "good" thresholds are LCP within 2.5 seconds, INP under 200 ms, and CLS under 0.1, and no Google document ties any of them to AI Overview selection. Search Engine Land's January 2026 analysis found "no strong positive correlation" between CWV and AI visibility, while "severe performance failures are associated with poorer AI outcomes." Fix a page that fails badly.
AI-cited content tends to skew fresher than standard organic results, so freshness is worth maintaining on the pages you care about. Set an update cadence for those clusters, revise rather than bumping a date, and show a last-updated timestamp with matching dateModified metadata.
What does not work
These AI Overview optimization patterns spend effort on signals Google's own documentation says it doesn't use, and none on the passage that would get quoted. Each one is common enough that you can find it in most content briefs teams wrote last year.
- Narrative-first intros: A page that opens with 300 words of context before answering competes against pages that answered in the first sentence.
- Thin topical coverage****: One page targeting a head term with no spokes for its sub-queries gives fan-out one chance to find you instead of several.
- Missing credentials: The evidence here is domain-level credibility, not bylines, and an anonymous site with no About page is betting against that skew.
- Slow or broken pages: Indexing errors and snippet blocks remove a page from consideration outright, and a page that fails Core Web Vitals badly is the one case where a speed sprint is worth booking.
- Special AI files: Google's May 2026 guide says you don't need to create new machine readable files, AI text files, markup, or Markdown to appear in Google Search.
Two caveats sit outside content quality. nosnippet and max-snippet:0 remove a page from AI Overviews entirely, since Google's robots meta spec says nosnippet "will also prevent the content from being used as a direct input for AI Overviews and AI Mode." A low max-snippet value caps how much of the page can be used. Audit templates for inherited directives before diagnosing anything else. And YMYL topics are volatile by design. Google applies "an even higher bar" to money-or-health queries and aims not to show Overviews on hard news, while Conductor's tracking of 274.5 million U.S. searches showed the overall trigger rate, not the YMYL-specific rate, swinging more than 24 points between September 2025 and February 2026, so a finance or health page can lose its citation with nothing changing on it.
How to measure whether you were cited
Google Search Console now reports AI Overview impressions. It will not tell you which queries produced them. The dedicated Generative AI performance report, announced June 3, 2026 and available globally since August 31, 2026, covers AI Overviews and AI Mode.
- No query or click data: Impressions break out by page, country, device, and date, and AI feature traffic has sat under the Web search type since June 16, 2025 with no filter to separate it.
- One shared position: An Overview occupies a single position, and every link inside it shares that number.
- Deduplicated impressions: A URL appearing in both the Overview and the blue links counts as one impression.
GA4 misattributes a chunk of this traffic. A study of 50,000+ tracked events reported by Search Engine Land found an average 22.4% landing in the Direct channel. Featured Snippets and People Also Ask use the same text-fragment pattern (#:~:text=), which complicates detection. Treat GSC impressions as the citation signal and referral traffic as a lagging, noisy proxy.
The GrowthX platform measures AI visibility across Presence (whether the brand appears in answers), Reputation (how AI answers characterize the brand), Perception (the sentiment and framing AI answers apply), and Influence (how much the brand shapes category narratives) on Google AI Overviews, ChatGPT, Claude, and Perplexity. CheckThat powers that measurement and covers 5,800+ brands, nearly 200 categories, and 2.6M+ AI responses, which provides a category baseline for the analysis.
Google can return different sources for identical searches, and pages can lose citations as its retrieval set shifts and a cluster stops covering the sub-queries Google uses.
Ship the cluster, revise on a cadence, keep the tracker's prompt list fixed, and check GSC impressions monthly. When that loop is spread across a rank tracker, a crawler, GSC, and a separate AI-visibility tool, the GrowthX platform consolidates it onto one context layer, and production runs through governed workflows where nothing ships without human approval.
If that maintenance is your bottleneck, book a demo and we'll show you the consolidated vision.
