How AI Search Decides What to Cite, and What It Ignores

Every AI search answer passes through a multi-stage reduction pipeline. The stages explain why an AI system never cites most content, whatever that content ranks in organic search.

Dan Petrovic at DEJAN AI reverse-engineered Google's Vertex AI Search pipeline and found a five-step process. The user query decomposes into fan-out sub-queries. Each sub-query retrieves candidate pages. The pipeline trims those pages to condensed versions. The condensed snippets become the LLM context. The model then writes an answer with citations.

Each stage removes content. Across five test domains, only 32% of total page characters from cited pages survived into the final answer, the content survival rate. The variation was wide: one domain kept 65% of its content, another only 21%. The survivors were service descriptions, pricing structures, and process instructions. The filter removed navigation, promotional claims, unrelated product categories, and customer review quotations.

Separately, AirOps analyzed 548,534 pages retrieved by ChatGPT and found that only 82,108 (15%) earned any citation at all. Retrieval does not equal citation. The reranking layer between the two filters aggressively.

The two filters compound. An AI system cites roughly 15% of the pages it retrieves, and about 32% of the text on a cited page survives into the answer. About 5% of the retrieved content reaches the reader.

Google uses extractive summarization , exact sentences from source pages, not paraphrases. DEJAN confirmed this by fine-tuning a DeBERTa model to replicate the behavior. The system applies query-focused selection with a heavy lead bias, and it extracts an opening paragraph near-wholesale. Every sentence has to work as a standalone extractable claim, a principle explored in depth in content engineering for AI extraction. Pronouns and anaphora ("it," "they") create extraction failures because the model cannot resolve them outside the original context.

The filter decides what survives. A second question decides what enters the filter, and the answer used to be the organic top 10. In mid-2025, Ahrefs found that 76% of AI Overview citations came from pages ranking in Google's organic top 10. By February 2026, that number was 38%. The remaining 62% came from positions 11-100 (31.2%) and beyond position 100 (31%). Google's Gemini 3 upgrade in January 2026 caused the shift, because it widened the source pool that the query fan-out system draws from.

The decoupling is even more pronounced across platforms.

ChatGPT has only 10% URL overlap with Google's top 10. An arXiv study found 4% domain overlap between GPT-4o and Google. The two products run functionally different retrieval systems, and the divergence compounds the gap between Google's public statements and its internal behavior.

Shashko's 42,971-citation study across six platforms gives the most granular view of the divergence. The organic SERP overlap varies widely by platform:

74.7% of all cited URLs across all platforms do not appear in the organic top 10. Perplexity is most aligned with organic rankings (43.5% URL overlap). ChatGPT is most independent (6.5%), which confirms the arXiv 4% domain overlap from a different angle. Copilot shows the strongest position sensitivity, with a mean organic rank of 2.99 when it cites. The Bing SEO fundamentals matter most for Copilot visibility.

One nuance matters. AirOps measured the relationship in the other direction. Of the pages that ChatGPT does cite, 55.8% rank somewhere in Google's top 20 for at least one query, and that count includes the fan-out sub-queries. A page at Google position #1 earns a citation 43.2% of the time, against 12.4% for a page beyond position 20.

The two findings agree. Google ranks far more pages than ChatGPT cites, so most Google-ranked pages never earn a citation. A page with strong fundamentals surfaces in both systems, because both systems reward similar quality signals. Shared quality creates the correlation. Rank does not cause the citation.

The divergence from Google is one gap. The divergence between the AI products is a second one, and it is wider. 86% of top-mentioned sources never appear across ChatGPT, Perplexity, and AI Overviews at the same time. Only 7 of the top 50 domains appear in all three platforms' top 50. Each platform has distinct retrieval architecture, source preferences, and citation behavior.

Citation rates vary 615x across platforms. Grok cites sources in 27% of responses. ChatGPT cites in 0.59%. A brand visible on one platform may be invisible on another.

Even Google's own AI products diverge. AI Mode, AI Overviews, and Gemini cite very differently despite sharing an owner. Ahrefs compared 730K response pairs and found 86% semantic similarity with only 13.7% citation overlap. The products agree on what to say, and they cite different sources for it. AI Mode cited 143% more unique domains than AI Overviews by January 2026, includes 2.5x more brand mentions, and behaves more like ChatGPT than like AI Overviews. The divergence extends even to Gemini: Shashko found only 3.5% shared domains between AI Mode and Gemini (Jaccard 0.035), or 247 shared domains out of 7,057. Two sibling Google products draw from nearly separate retrieval pools. A self-citation bias sits on top: AI Mode cites Google's own properties at 17.42%, up from 5.7%.

Social media influence splits the same way. Tinuiti/Profound's Q1 2026 report found Reddit drove 24% of Perplexity's citations in January 2026 but effectively 0% of Gemini's. YouTube matters for Gemini but barely registers on ChatGPT. Reddit's citation share grew 73-100%+ across all verticals between October 2025 and January 2026. However, Yext data (via Surfer SEO) adds a caveat. When the study controls for intent and location, brand-controlled sources take 86% of AI citations, and Reddit drops to 2%. Indig/Johnson corroborate this from a different angle. Across 98,217 citations and 7 verticals, corporate content accounts for 94.7% of all AI citations, and UGC accounts for 5.3%.

The UGC share moves with the vertical. Finance is the most corporate-locked at 0.5% UGC. Healthcare sits at 1.8%. Crypto shows the highest UGC penetration at 9.2%, where community posts fill documentation gaps. Reddit's 24% share on Perplexity is a platform-specific and intent-specific amplification, and not a universal pattern.

The source preference differences are structural:

• AI Overviews: Over-indexes on UGC, YouTube (9.5%), Reddit (7.4%), Quora (3.6%)
• ChatGPT: Wikipedia dominant (16.3%); leans toward publishers and news sources
• Perplexity: YouTube (16.1%) + Wikipedia (12.5%); broadest international corpus
• AI Mode: Favors commercial and authoritative sources. Highest entity density

The divergence extends beyond source domains to content types. Wix AI Search Lab analyzed 1,056,727 citations across ChatGPT, Google AI Mode, and Perplexity. All three models agree on the listicle as the most-cited content type, and they disagree on what comes second. ChatGPT shows the highest article representation (+4.38% above average) and the lowest discussion preference (-4.32%), which fits its Bing-based, publisher-heavy retrieval. Perplexity cites discussions at 17.35%, more than double the 7.52% cross-model average, which fits its paragraph-level Vespa.ai retrieval architecture. Google AI Mode is the most balanced, and it spreads citations across all 11 content types with little bias.

The platform decides which pool it draws from. The format of the page decides what the platform can take out of it. Structured content, which means a table, a comparison, or a guide with a clear section hierarchy, outperforms narrative prose in every study in this set.

The aggregate numbers hide one variable: query intent is more predictive of content type citation than either industry or model choice. Wix AI Search Lab analyzed 1,056,727 citations (75,000 AI answers across ChatGPT, Google AI Mode, and Perplexity) and found the content format hierarchy inverts depending on intent:

An article dominates the informational queries at 45.5%, and it drops below 7% for every other intent. A listicle takes 40.9% of the commercial citations. Product and category pages together take 40% of both the transactional and the navigational citations. Third-party listicles produce 80.9% of those listicle citations, which matters, because Google now demotes the self-serving kind.

The intent-structure relationship extends beyond content type to on-page formatting. Previsible analyzed 5,000 prompts across four intent types and 12 structural attributes. Branded factual queries favor entity naming (82%), feature lists (64%), and short paragraphs (71%). Branded competitive queries trigger comparison tables (52%) and evaluation criteria headers (67%). Category buying queries surface ranked lists (74%) and ICP segmentation (46%). Informational queries follow a Definition-Context-Example pattern (58%), with taxonomy structures (42%) and almost no brand naming (11%).

Some structures hold across all four intents. Headers appear every 100-200 words. Lists appear in 63% of the cited content, tables in 39%, FAQ structures in 47%, and interrogative headers in 58%. An LLM converts the source content into a list 76% of the time, whatever format the source used.

For branded queries specifically, the source type hierarchy inverts. Omniscient Digital ran 240 branded prompts through ChatGPT, Perplexity, Gemini, AI Mode, and AI Overviews. The set covers 23,387 unique citation sources, 4 industries, and 6 intent types. Third-party validation dominates it. Reviews and social proof account for 57% of the branded citations. Directories and reference sites take 17%, product pages 12%, education and thought leadership 5.4%, and brand foundation content 4.5% of them. Brand-owned content is a minority of the citations about the brand itself.

The structural advantage is measurable at the element level, according to Onely's compiled research and the AirOps 2026 State of AI Search:

• Semantic HTML tables increase citation rates approximately 2.5x versus paragraph text
• ChatGPT citations include tables 2.3x more frequently than traditional search (30% vs 13%)
• FAQ-structured content shows 28-40% higher citation probability
• Sequential headings correlate with 2.8x higher citation likelihood. But heading count is vertical-specific: Indig/Johnson found 3-4 headings is worse than zero in every vertical (a dead zone between no structure and committed structure). CRM/SaaS peaks at 20-49 headings (18.2% high-cited at 50+). Healthcare inverts: citation drops from 15.1% at zero headings to 2.5% at 20-49. Finance peaks at 10-19, crypto at 5-9. Education is flat across all heading counts. The "20+ headings" advice is a CRM/SaaS-specific finding, not universal
• Pages with title-query alignment of 50%+ see 2.2x citation lift

One writing signal holds across every vertical. Indig/Johnson analyzed 1.2M ChatGPT responses across 7 verticals. Only one writing-level signal predicts a higher citation rate in every vertical they tested: declarative phrasing in the opening paragraph, with a +14% aggregate lift. The form is "[X] is [Y]" or "[X] does [Z]". A hedge in the intro, such as "This may help teams understand", suppresses the citation. "Teams that do X see Y" outperforms it, and no other writing change holds this consistently.

A specific structural pattern keeps surfacing. An analysis of ChatGPT-cited blog posts covered 15 domains, about 2M monthly organic sessions, and 7,500 ChatGPT referrals. 72.4% of the cited posts carried an "answer capsule". The capsule is a declarative statement of 120-150 characters (about 20-25 words) placed right after a question-based H2 heading.

52.2% of the same posts carried original data or a branded insight. Both traits appeared together in 34.3% of the cited posts, and only 13.2% carried neither. About 91% of the capsules held no outbound link, so clean self-contained text extracts more reliably.

DEJAN's reverse-engineering of AI Mode found that the AI snippet selection caps at about 160 characters. The algorithm ranks semantic relevance, structural importance (the HTML hierarchy), content density, and value proposition detection. It also selects customer-centric language ("you", "your team") more often.

A page also holds what Petrovic calls a "semantic topography". Different regions of the same page sit at different semantic coordinates. When AI systems decompose a query into sub-queries, each sub-query surfaces different passages from the same page. A query about "risks of X" surfaces the avoidance language. A query about "benefits of X" surfaces the benefit language. A well-structured page can serve multiple fan-out queries simultaneously if each section is independently extractable, the same architectural principle behind programmatic SEO.

The inverse also holds. SE Ranking tested 2,000 pure AI-generated articles across 20 brand-new domains. 71% of the pages entered the index within 36 days, and 28% briefly reached the Google top 100 results. The rankings then collapsed to 3% by month three. By month 16 the 2,000 pages had earned 1,381 clicks in total, or about 0.7 clicks per page.

Ahrefs shows that AI content fails at DR 91. SE Ranking shows that it fails at zero authority. Information originality is the variable that separates the two outcomes, and domain strength is not.

Originality decides the citation. Domain authority looks like it decides the citation, and the two studies below explain why the appearance holds. Ahrefs found that ChatGPT's most-cited pages carry a median DR of 90. AirOps measured something different: the rate at which a retrieved page converts to a citation.

The citation rate holds at 21.5-23.6% across DA 0-80, and it drops to 15% for a DA 80-100 site. The retriever pulls a high-authority site more often, so the site collects more citations in total. It converts retrieval into citation at a lower rate, because it covers a topic broadly instead of answering a specific query.

Both findings hold at once. A high-DR site collects more citations in total, and a mid-authority site competes on the citation rate. A base rate separates the two. A mid-authority site (DA 20-80) competes on citation when it holds topical precision, as the Banksparency case study demonstrates with 10K+ monthly visits on a low-authority domain. The barrier is relevance, not domain authority.

The paradox extends to entities. Indig/Johnson ran Google's Natural Language API on the first 1,000 characters of 5,000 pages across 7 verticals and found that Knowledge Graph-verified entities are a negative citation signal (0.81x lift). High-cited pages average 1.42 KG-verified entities versus 1.75 for low-cited pages. Pages built around well-known, KG-verified entities (a major brand, an institution, a famous person) tend toward generic coverage, and the AI systems deprioritize it. A high-cited page is dense with specific, niche entities: a named methodology, a precise statistic, a named comparison. Many of those entities hold no KG entry at all. A Wikipedia entry, a brand panel, and KG verification are the wrong levers.

The entity type data adds a second dimension: DATE and NUMBER are the most universal positive citation signals. Include a publish date and at least one specific number. PRICE is the strongest universal negative, suppressing citation in 5 of 6 verticals (it signals commercial intent). Finance is the exception, where a price means a fee percentage or a rate comparison, the reference data that a financial query asks for. A phone number is a positive signal in Healthcare (1.41x) and Education (1.40x). The number stands in for institutional presence, so do not read it as an instruction to add phone numbers.

Concentration at the top is still real. Goodie AI analyzed 5.7M citations and found that the top 50 domains capture roughly 53% of all citations, while 40,000+ sites split the remainder. 74% of the most-cited domains are "susceptible to marketing influence", so a deliberate presence on those properties moves part of the citation share.

A brand in the top quartile for web mentions receives 10x more AIO citations than the next quartile. The relationship is non-linear, so a mention below the top quartile returns little. Directional data from ConvertMate ties an active profile on a review platform (Trustpilot, G2, Capterra) to a 3x higher ChatGPT citation probability. ConvertMate does not publish the methodology behind that figure.

The concentration is also mostly out of reach. 67% of ChatGPT's top 1,000 citations sit on properties that no optimization touches: Wikipedia (29.7%), homepages (23.8%), app stores, and reference sites. Only 32.3% leave room for content work or outreach. Most brands then meet what RankScience calls the ghost citation problem. The model cites them as an evidence source, and it never recommends them as a brand.

Authority sets a weak prior. Publication date sets a strong one. AI assistants cite content that is 25.7% fresher than the content in the organic results. Ahrefs analyzed 16.975M cited URLs and found that ChatGPT's citations are 458 days newer on average. Google's AI Overviews are the exception, and they prefer slightly older content.

The owners had updated 76.4% of ChatGPT's top-cited pages within 30 days, and 89.7% at some point during 2025. Seer Interactive found that 65% of AI bot crawl hits target content published within the past year, and 50% of Perplexity citations are from 2025 content alone.

Perplexity is the most freshness-sensitive platform. ConvertMate data (via Surfer SEO) estimates that freshness accounts for roughly 40% of the Perplexity ranking factors. Content updated two hours ago earns 38% more citations than month-old content.

Freshness is an active ranking factor. Reverse-engineering of ChatGPT's configuration revealed a use_freshness_scoring_profile: true flag that nobody can turn off. Freshness scoring is an active layer that can override content quality. Adding fake publication dates boosted AI visibility by up to 95 rank positions.

The platform freshness split is even starker than the averages suggest. Shashko's sentence-level study found that for Google's AI Mode specifically, the median cited page is 2.2 years old. Over half (52.8%) of AI Mode's cited content is 2+ years old, and 26.3% is 5+ years old. The ChatGPT and Perplexity freshness bias runs the other way. AI Mode cites evergreen content that the other platforms deprioritize.

The freshness strategy therefore splits by platform. Run an aggressive refresh cadence for ChatGPT and Perplexity visibility, and choose depth over recency for AI Mode. A page that misses the quarterly refresh is 3x more likely to lose its citations on a freshness-sensitive platform.

Freshness and format both describe a page. Fan-out describes the query that finds it. When a user asks an AI search engine a question, the system does not search for that question. The system decomposes the query into sub-queries, between 2.9 (AirOps) and 10.7 (Gemini 3, Seer Interactive) on average, and retrieves results for each one independently. Google Patent US11663201B2 defines 8 variant types: Equivalent, Follow-up, Generalization, Specification, Canonicalization, Translation, Entailment, and Clarification.

95% of these fan-out queries have zero traditional search volume. No keyword research tool on the market reports them, which is why a diagnostic method has to account for the AI retrieval paths. They still carry the primary retrieval pathway. 89.6% of ChatGPT searches generate 2+ fan-out queries. For 32.9% of cited pages, the citation comes from a fan-out result alone, and not from the top 20 for the original query.

That gap is the "invisible keyword" problem. Nearly a third of AI citations come from queries that the user never typed and that conventional measurement cannot detect. The count also grows fast. Gemini 2.5 averaged 6.01 sub-queries, and Gemini 3 (January 2026) raised that by 78% to 10.7. Each model generation widens the retrieval surface, so a page that no AI system could reach becomes reachable with no change to the page itself.

Fan-out behavior is not random. It varies by intent:

• Definition queries stay close to the original phrasing (51.6% near-verbatim)
• Research queries add temporal modifiers, 21.3% of fan-out queries contain a year
• Comparison queries decompose most aggressively (38.4% sub-question splitting)

The recency injection is notable. An AI-generated sub-query adds a temporal bias when the user asks for none. The term "2026" appeared 184x more often than "2025" in Gemini 3's sub-queries.

A page that answers several facets of a topic holds several retrieval entry points. The count of entry points is the structural case for comprehensive, well-sectioned content. The grounding still plateaus at 540 words, so the gain comes from the semantic coverage that lets more fan-out queries match. Writesonic's GPT-5.4 analysis found that 67% of cited domains don't appear in traditional Google or Bing results. Surfer SEO corroborated this from the other direction: 67.82% of AI Overview-cited sources do not rank in Google's top 10 organic results. Two independent studies, one on ChatGPT and one on AI Overviews, report the same number. Fan-out retrieval operates independently of the conventional SERP.

Kevin Indig analyzed 21,482 ChatGPT citations and found that 67% of cited URLs appear in only one prompt. Most citations are one-hit appearances. The top 4.8% of URLs, the ones cited in 10+ distinct prompts, share three structural patterns. They use a category-level guide format, they cover a broad topic inside a single page, and they anchor an explicit year. One comprehensive page that covers 10+ query intents is worth more than 10 single-intent pages. No thin, single-topic page reached the 11+ prompt tier in any vertical.

Fan-out changes what the system retrieves. The answer format changes what the reader does next. AI Overviews cut organic clicks by about 34.5% on the queries where they appear. In B2B SaaS, the figure is steeper: Kevin Indig measured a 56.6% click decline across 10 sites and ~450M impressions since the March 2025 AIO rollout intensification. Zero-click rates in AI Mode reach 92-94%, with users clicking only once per 20 prompts, a dynamic explored further in the zero-click paradox.

But the clicks that do happen are dramatically more valuable.

An AI visitor arrives pre-briefed by the answer. That visitor engages more deeply, converts faster, and bounces less. The pattern is the "educated click". Fewer clicks arrive in total, and each one carries 5x the conversion value of a traditional organic visit.

The measurement compounds the problem. AI referrer attribution breaks on every platform. Google AI Mode strips the referrer, which John Mueller confirmed as a bug. ChatGPT strips it for a paid account, and the Perplexity and Copilot desktop apps drop it too. GA4 then files the visit as "Direct".

Seer Interactive's 2026 analysis measured a 70.6% misclassification rate. A standard GA4 configuration files most AI search traffic under the wrong channel. Without custom channel groupings and UTM parameters, a traffic impact analysis of AI search runs on inaccurate data.

An AI system decides what to cite in the reranking layer and in the extraction step, and not in the ranking that SEO tools measure. Eight working rules follow from the studies above.

Density over length, with vertical nuance. The grounding budget is ~1,929 words. Content past the first 540 words of grounding has diminishing returns. A page under 1,000 words retains 61% of its content in an AI answer. A page over 3,000 words retains 13%. But the length-citation relationship is vertical-specific. In finance, shorter content wins, and the relationship inverts. In education and crypto, length pays off linearly. SaaS shows the weakest length effect. A page under 1,000 words underperforms in every vertical studied. The near-zero aggregate correlation that Ahrefs reports hides these vertical effects.

Structure for extraction. Every sentence must work as a standalone, citable claim. Use semantic HTML. A table increases the citation rate 2.5x. Use a clear heading hierarchy. Do not write a pronoun chain, because it breaks when the system lifts one sentence out of its context.

Optimize for fan-out and the primary query. A single user query generates 3-28 sub-queries. Content that covers multiple facets of a topic, with each section independently extractable, has more entry points into the AI retrieval pipeline.

Freshness is a lever. On the non-Google AI platforms, content updated within 3 months earns nearly 2x the citation rate of older content. A quarterly refresh cadence is the baseline.

Platform-specific strategies are not optional. The 615x citation rate variance across platforms means a single "AI optimization" strategy will fail. Each platform has different source preferences, retrieval architectures, and content biases.

Mid-authority sites can compete. The citation rate is flat from DA 0-80, and it drops for DA 80-100. The barrier is topical precision, not domain authority.

Traditional SEO metrics are near-irrelevant for AI citation. Traffic and backlinks explain less than 5% of citation behavior. The fundamentals that drive organic quality still transfer: entity richness, semantic relevance, and clear structure. Both systems reward the same content characteristics. The clinical diagnostic framework maps these shared quality signals systematically.

Measurement fails. Analytics undercounts AI traffic. Citation visibility moves from run to run: a brand holds 30% retention per answer, and 20% across five consecutive runs. An AI visibility program needs continuous monitoring.

The eight rules answer the question the page opened with. An AI search engine cites the passage that its extractor can lift cleanly and that its retriever can reach through a fan-out query. The freshness profile of the platform then decides whether the passage is recent enough. The system ignores the page that ranks well and reads as continuous prose. Write for the extractor first, and the ranking follows as a second-order effect.