How AI Overviews Choose Sources: What We Know
Photo: Tim Witzdam / Pexels
Google says AI Overviews run on the same index and ranking systems as regular Search, with no special markup required. Independent analysis of millions of citations tells a more complicated story: pages that never crack the top 10 organic results get cited constantly, and the overlap between "ranks well" and "gets cited" has been shrinking for over a year. Both things are true at once, and understanding why is the actual answer to how source selection works.
What Google says versus what gets measured
Google's own documentation is short and specific. AI features "run on the same index and ranking systems as regular search," there are "no additional requirements to appear in AI Overviews or AI Mode," and "no special structured data" is needed beyond standard helpful-content practices (Google Search Central, AI features and your website). Google product leadership has repeated this in public talks and interviews: AI Overviews are described as a synthesis layer sitting on top of the existing ranking pipeline, not a parallel system with its own signals.
That is a narrow, defensible claim. It says AI Overviews draw from the ranked web index. It does not say every citation comes from a page that ranked in the visible top 10 for that specific query, and it doesn't say the passage selected for the summary is chosen the same way a snippet is chosen. Those are two different steps: retrieval (which pages are candidates) and synthesis (which passages get pulled into the generated answer and cited). Google is precise about the first step and largely silent on the mechanics of the second.
That drop matters because it undercuts the simplest version of "just rank #1 and you'll get cited." Something in the synthesis step is pulling in pages that traditional ranking signals alone wouldn't surface as the single best answer, at least not in a way a human editor would recognize from a SERP.
How the pipeline actually works, step by step
Think of AI Overviews less like a single ranking algorithm and more like a research assistant with a strict word count. The assistant doesn't write from scratch. It is handed a stack of already-ranked documents, told to read fast, and told to produce a short answer with receipts. Every step in that process introduces its own filter, and a page can fail at any one of them even if it's excellent at the others.
Step one: retrieval. Google's query fan-out expands a single search into several related sub-queries — sometimes reformulating "best budget laptops" into implicit questions like "what makes a laptop budget," "what specs matter," "current price ranges." Each sub-query pulls its own set of ranked candidate pages from the normal index. A page you'd never associate with the visible query can still get pulled in because it ranks well for one of the invisible sub-queries.
Step two: passage extraction. From the candidate pool, the system doesn't grab whole pages. It extracts specific passages, paragraphs, or list items — the exact sentence or two that answers a sub-query cleanly. This is why a page ranking #7 with one dense, self-contained paragraph can out-cite a page ranking #1 that buries the same fact in a rambling introduction. The unit being scored is the passage, not the URL.
Step three: synthesis and citation. The extracted passages get stitched into a generated summary, and each cited chip links back to its source page. This is where redundancy gets trimmed — if three pages say the same thing in similar words, only one or two typically get the citation slot, and it tends to go to whichever version is clearest or most recently crawled.
Why does a page rank #4 in organic but win the AI Overview citation while the #1 page doesn't?
Because organic ranking scores the whole page's relevance and authority for the query, while citation selection scores individual passages for how cleanly they answer a specific sub-question. A page can rank lower overall but contain the single best-written paragraph for the exact fact being synthesized, which is enough to win the citation slot even against a higher-ranked competitor.
The signals that correlate with getting cited
No public technical paper documents Google's exact citation-scoring formula. What exists is a mix of Google's stated best practices and pattern analysis from large third-party citation datasets. Treat the second category as observed correlation, not confirmed causation — Google has not verified these mechanisms.
| Signal | What Google says | What citation studies observe |
|---|---|---|
| Ranking in top 10 organic | Not explicitly required | Correlated with citation odds, but the correlation weakened from 76% to 38% co-occurrence between mid-2025 and 2026 (Ahrefs) |
| Structured data / schema markup | "No special structured data is needed" | Mixed evidence: one controlled Search Engine Land test found the schema-marked page got cited and a non-schema version of similar content didn't; a separate Ahrefs test across multiple AI platforms found schema made no measurable difference to citation rates (Search Engine Land) |
| Answer-first structure (direct answer near the top) | Recommended as general helpful-content practice | Strongly associated with passage extraction success in the Princeton GEO study, which found specific statistics, direct citations, and quotable phrasing improved a page's odds of being pulled into generated answers |
| Freshness / recent crawl date | Part of normal ranking systems, no AI-specific boost stated | Frequently cited informally by SEOs tracking citation churn, though no controlled public study isolates freshness as an independent variable |
| Fresh, well-known domains (YouTube, Reddit, Quora) | No stated preference | Semrush's 2026 AI Visibility Index, built from 126 million analyzed prompts, found YouTube was the single most-cited domain across AI Overviews, with Reddit and Quora also in the top tier — likely reflecting how often those platforms directly answer conversational queries, not a domain-level trust boost |
The pattern across the honest data: nothing here is a magic switch. Structure and directness show up repeatedly as helpful; schema markup shows up as unreliable, sometimes helping and sometimes doing nothing measurable; raw domain authority shows up as a weaker predictor than it was for classic organic ranking.
From featured snippets to a synthesized answer
AI Overviews didn't appear out of nowhere. They're the latest step in a decade-long Google project to understand queries well enough to answer them directly instead of just pointing to a ranked list. RankBrain (2015) applied machine learning to query interpretation for the first time. BERT (2019) improved Google's grasp of context and nuance in phrasing, and Google has said BERT specifically helps identify which passage is the most appropriate one to surface as a featured snippet (Google, BERT announcement context via Search Central history). MUM, introduced in 2021, extended that further, handling text, image, and video understanding together across roughly a thousand times the model capacity of BERT.
Featured snippets were the direct ancestor of today's citation chips: a single passage lifted from a single ranking page, displayed above the rest of the results. AI Overviews generalize that idea. Instead of pulling one passage from one page, the system pulls multiple passages from multiple pages, across multiple ranking positions, and stitches them into a synthesized paragraph with several citations attached. The shift from "one snippet, one source" to "one summary, many sources" is exactly why source selection got harder to reverse-engineer. A featured snippet had one clear winner. An AI Overview can have five or six partial winners, each contributing one sentence.
How this differs across ChatGPT, Perplexity, and Google
Treating "AI search" as one thing is a common mistake, because the major platforms pull from meaningfully different source pools. Independent citation-pattern analysis comparing the three largest assistants found real structural differences, not just stylistic ones.
| Platform | Typical citation count per answer | Dominant source type | Retrieval behavior |
|---|---|---|---|
| Google AI Overviews | Roughly 12 sources per answer | YouTube and multi-modal content lead; SeoClarity found 97% of AI Overviews include at least one source from the top 20 organic results | Draws from Google's existing crawled index; query fan-out expands one visible query into several sub-queries |
| ChatGPT (Search) | Roughly 7 to 8 sources per answer | Encyclopedic and editorial sources; Wikipedia alone accounted for close to half of top-10 citations in one large sample, supplemented by wire services like Reuters, AP, and the BBC | Hybrid of training-data knowledge and selective live web retrieval, triggered when the model judges a query needs current information |
| Perplexity | Over 21 sources per answer | Community and forum content, with Reddit as the single most-cited domain in some samples | Performs a live web search on essentially every query, with a strong recency bias toward content published in the last 30 days |
Table: Citation behavior comparison across three major AI search platforms, based on independent third-party citation-pattern analysis (Profound, AI Platform Citation Patterns).
The practical implication is that "getting cited by AI" isn't a single target. A page optimized to win a Perplexity citation, dense, recent, conversational, forum-adjacent, isn't the same page that wins a ChatGPT citation, which leans toward encyclopedic framing and institutional sourcing. Independent analysis has also found that domain overlap between platforms is low: only around 11% of domains get cited by both ChatGPT and Perplexity for the same query, and roughly 71% of all cited sources show up on just one platform. Building for "AI visibility" in the abstract, without picking which assistant's retrieval behavior you're actually targeting, is a bit like optimizing a resume for "getting a job" without knowing whether the reader is a hiring manager or an applicant-tracking-system keyword filter. The mechanics reward different things.
A worked example: how one query might get answered
Consider the query "does creatine cause hair loss." A classic organic SERP would return a ranked list of pages, mostly health sites and forums, and the user would click through and read one. Under the AI Overview pipeline, query fan-out likely expands this into implicit sub-questions: what creatine does physiologically, whether any clinical study links it to DHT (a hormone tied to hair loss), and what the current scientific consensus says about causation versus correlation.
Each sub-question pulls its own candidate pool. The DHT mechanism question might pull a passage from a sports-medicine journal summary page. The "current consensus" question might pull a passage from a health-information site that explicitly states uncertainty and cites a specific trial. A general wellness blog that ranks well overall but only vaguely says "some people worry creatine causes hair loss" is unlikely to win any of these citation slots, because it doesn't contain an extractable, specific answer to any single sub-question. This is the mechanism behind the Ahrefs finding that top-10 organic ranking and AI Overview citation have diverged: the blog might still rank on page one for the broad query while losing every citation slot to more specific, deeper-ranked pages that answer the fan-out sub-questions directly.
E-E-A-T and why it's the closest thing to a stated factor
Google has never published a numeric E-E-A-T score, and E-E-A-T (Experience, Expertise, Authoritativeness, Trustworthiness) isn't a ranking factor in the traditional sense; it's a framework Google's own Search Quality Rater Guidelines use to describe what quality content looks like. But it's the closest thing to an officially sanctioned lens for AI feature eligibility, because Google's AI-features guidance explicitly points back to the same helpful-content and quality-rater principles that govern normal Search.
In practice, for a passage to be both retrieval-eligible and extraction-friendly, it typically needs to demonstrate: a clear author or organizational identity, verifiable claims with sources of their own, evidence the page isn't just repackaging someone else's reporting, and content that matches genuine user need rather than being built primarily to rank. None of that is unique to AI Overviews. It's the same bar Google has been raising for organic Search since the 2018 "Medic" update and the 2022 Helpful Content update. AI Overviews didn't invent a new quality bar; they raised the cost of failing the existing one, because a vague or unsourced page now has two ways to lose visibility (organic ranking and citation extraction) instead of one.
What the Princeton GEO research actually found
The most rigorous public study on this topic isn't from an SEO vendor. It's an academic paper from researchers at Princeton, Georgia Tech, IIT Delhi, and the Allen Institute for AI, published as "GEO: Generative Engine Optimization" and presented at the ACM SIGKDD conference (arXiv 2311.09735). The researchers built a 10,000-query benchmark across eight domains, paired with the actual source documents a generative engine draws from, and then systematically tested which content changes moved the needle on citation frequency.
Their controlled result: adding specific, verifiable statistics and clear source citations to a passage increased its visibility in generated answers by up to 40% relative to an unmodified baseline. Keyword density and content padding, the two levers that dominated 2010s-era SEO, showed no comparable effect and sometimes hurt. Authoritative, quotable phrasing — sentences that could be lifted whole into a summary without needing rewriting — also correlated with higher inclusion rates.
This is useful because it's the rare study with an actual control group, rather than observational correlation pulled from a citation crawl. It doesn't prove Google's specific model works identically to the generic generative-engine benchmark the researchers tested, but it's the closest thing to causal evidence in this space, and it lines up with what Google's own "helpful content" guidance already recommends: be specific, be sourced, be direct.
Why citation patterns keep shifting under your feet
If this feels less stable than classic SEO, that's because it is. Three forces are compounding.
The underlying model changes. AI Overviews moved onto more capable Gemini models over the course of 2025 and 2026, and each model version can weigh passage quality, sub-query decomposition, and redundancy differently. A citation pattern documented in one quarter isn't guaranteed to hold in the next, which is exactly what the Ahrefs top-10-overlap number (76% to 38%) shows happening in real time.
Query fan-out multiplies the effective competition. A single visible search query might expand into five or six sub-queries behind the scenes. A page doesn't just compete against everyone ranking for the literal search term; it competes against everyone ranking for every implicit sub-question the system generates. That's a much larger and less visible competitive set than a traditional SERP.
AI Overviews suppress the very signal that used to validate strategy. Ahrefs' earlier 300,000-keyword CTR study found that pages ranking #1 organically saw meaningfully lower clickthrough when an AI Overview appeared above them (Ahrefs, AI Overviews reduce clicks). That means the traditional feedback loop — publish, watch clicks, iterate — gets muddier, because clicks are being redirected or eliminated by the summary itself before a reader ever reaches the page.
What people get wrong about this topic
"Add FAQPage schema and you'll get cited." The evidence doesn't support a reliable causal link. One controlled test found schema correlated with a citation; a separate multi-platform Ahrefs analysis found no measurable lift from schema across Google AI Overviews, ChatGPT, and other generative engines. Schema still helps machines parse structure and can support classic rich results, but treating it as an AI-citation lever oversells what's been demonstrated.
"Ranking #1 guarantees the citation." It used to correlate strongly; now it doesn't, based on the shift from 76% to 38% top-10 overlap. A page can rank #1 and lose the citation to a lower-ranked page with a cleaner, more extractable passage on the exact sub-question the system needed answered.
"AI Overviews are a separate ranking system you have to game differently." Google's position, that AI features run on the same index and ranking systems, is consistent with what retrieval-stage evidence shows: candidate pages still have to earn organic relevance first. The divergence happens downstream, at passage extraction and synthesis, not at the initial retrieval gate.
"This is all guesswork, so there's no point optimizing." The Princeton controlled study is real evidence, not speculation: specificity, sourcing, and quotable phrasing produced a measurable, repeatable lift in a benchmark environment. The uncertainty is about Google's exact production weights, not about whether structure and specificity matter at all.
Is AI Overview citation the same thing as ranking well in Google Search?
No. Ranking well in the organic top 10 remains the entry ticket for being considered as a retrieval candidate, but the actual citation decision happens at the passage level during synthesis. A page can rank respectably and never get cited, or rank modestly and get cited repeatedly, depending on how extractable its specific passages are.
Does structured data (schema markup) help a page get cited in AI Overviews?
The evidence is mixed. Google's official guidance states no special structured data is required. One controlled Search Engine Land test found a schema-marked page got cited over a non-schema version, but a broader Ahrefs analysis across multiple AI platforms found no measurable citation benefit from adding schema. Treat schema as good technical hygiene, not a proven citation lever.
Why do YouTube, Reddit, and Quora show up so often as cited sources?
Semrush's 2026 AI Visibility Index, based on 126 million analyzed AI search prompts, found YouTube, Reddit, and Quora among the most-cited domains overall. The likely explanation is format fit: those platforms are dense with direct, conversational answers to specific questions, which matches what passage extraction is optimized to pull, rather than any domain-level trust boost from Google.
Do AI Overviews reduce clicks to the pages they cite?
Yes, on average. Ahrefs' analysis of 300,000 keywords found that the presence of an AI Overview correlated with meaningfully lower clickthrough for the top organic result compared to similar queries without one. A citation credit doesn't guarantee traffic, since many users read the generated summary and never click through.
Does ChatGPT choose sources the same way Google AI Overviews does?
No. ChatGPT Search leans heavily on encyclopedic and editorial sources, with Wikipedia alone representing close to half of top citations in independent analysis, and it typically cites far fewer sources per answer than Perplexity or Google. Google AI Overviews draw from its own crawled organic index through query fan-out and favor multi-modal content, including video.
Why does Perplexity cite so many more sources than ChatGPT or Google?
Perplexity performs a live web search for nearly every query rather than relying partly on stored training knowledge, and independent citation studies put its average source count above 21 per answer, roughly three times ChatGPT's average. That behavior also gives Perplexity a strong bias toward very recent content, with pages published in the last 30 days cited disproportionately often.
Can a page with no organic ranking still get cited in an AI Overview?
It's uncommon but not impossible, since AI Overview citations have been documented coming from pages ranking beyond position 100 in Ahrefs' larger citation study. It's more typical for a citation to come from a page that at least ranks somewhere in Google's broader index for a related sub-query generated during fan-out, even if it doesn't rank in the visible top 10 for the exact search term.
Is optimizing for AI Overviews different from traditional SEO?
Not fundamentally, according to Google's own guidance, which states the same helpful-content best practices apply. The practical difference is emphasis: traditional SEO optimizes a whole page to rank for a query, while AI Overview citation rewards individual passages that answer a specific sub-question clearly enough to be lifted into a generated summary without editing.
The takeaway
AI Overviews are not a separate ranking system with a hidden AI-only algorithm, Google is telling the truth about that much. But the visible organic ranking is only the entry gate; the actual citation decision happens one layer downstream, at the level of individual passages, and that layer runs on rules that are less stable and less publicly documented than classic SEO ever was. The most defensible response isn't chasing a specific tactic like schema markup. It's writing passages that state a specific fact, name the source, and stand on their own well enough that a system skimming for the single best sentence would have no reason to skip yours.
Read next:
- GEO Explained — Generative Engine Optimization
- How to Rank in AI Overviews — AEO Guide (2026)
- AI Agents Explained — How Autonomous AI Systems Actually Work
Sources
- Google Search Central: AI features and your website
- Ahrefs: I Analyzed 300K Keywords. Here's What I Learned About AI Overviews
- Ahrefs: AI Overviews Reduce Clicks
- Search Engine Land: Schema and AI Overviews
- Aggarwal et al., "GEO: Generative Engine Optimization," arXiv:2311.09735 (Princeton, Georgia Tech, IIT Delhi, Allen Institute for AI)
- Profound: AI Platform Citation Patterns — How ChatGPT, Google AI Overviews, and Perplexity Source Information
- Wikipedia: AI Overviews (background and timeline)
Follow new articles
Add SciHub101 to your feed reader to see new science and technology stories.
About the Author
SciHub101 is an independent science and technology publication founded and edited by Buğra Sözeri, a software engineer and app developer based in Türkiye. AI tools assist with drafting and research. Source documentation in the existing archive is uneven, and articles are being reviewed and corrected. The Editorial Team byline identifies editorial responsibility; it does not certify specialist or independent human review. Where we compare products, anything built by our own team is labelled as such — see the editorial guidelines and About page for how corrections and disclosures work.
Related Articles
How to Make Content Citable by AI
AI systems cite pages that make claims easy to verify and quote. This guide explains definitions, tables, sources, author signals, and topical clusters.
Generative Engine Optimization Checklist
A practical GEO checklist for making content easier for AI systems to understand, cite, summarize, and connect to related entities.
How to Rank in AI Overviews — AEO Guide (2026)
How to get cited in Google AI Overviews, ChatGPT, and Perplexity in 2026: crawler access, listicle structure, schema, Wikipedia strategy, and IndexNow.
GEO Explained — Generative Engine Optimization
GEO is the practice of optimizing content to be cited by AI engines like ChatGPT and Perplexity. How it works, tactics, and how it differs from SEO.
AI Search vs Traditional Search: How Discovery Is Changing
Traditional search ranks pages; AI search composes answers. This article explains the difference and what it means for publishers, users, and brands.
Answer Engine Optimization Explained
Answer engine optimization focuses on direct answers, structured sections, FAQ blocks, sourceable claims, and entity coverage for Google, Bing, Perplexity, and AI assistants.