EVIDENCE

Google AI Overviews and AI Mode do not run a single query. They use a query fan-out technique, issuing multiple related searches across subtopics and data sources to assemble a candidate pool, and they continue to identify supporting pages while the answer is being written, which means grounding is an active process running in parallel with generation ^[1]. ChatGPT’s pipeline starts with roughly 40 to 50 search results, then narrows to 10 to 20 using titles, descriptions, and domain signals before any page is loaded, and finally to 3 to 5 pages that the model actually uses ^[3]. Perplexity pulls 60 or more candidate sources per query through hybrid keyword and semantic retrieval, then applies a multi-layer reranker with an approximate 0.7 quality threshold that discards entire weak result sets rather than demoting individual entries ^[4].

The decisive insight is that retrieval and citation are two gates, not one. In Ahrefs’ study of 1.4 million prompts, ChatGPT cited approximately 50 percent of the URLs it retrieved and left the other 50 percent uncited ^[2]. Reddit alone accounted for 67.8 percent of the non-cited URLs, meaning the model leaned on it heavily for consensus and context while almost never crediting it ^[2]. What separated cited from non-cited pages was measurable semantic relevance: cited page titles scored 0.602 cosine similarity to the prompt versus 0.484 for non-cited titles ^[2]. Two structural facts compound this. Only about 38 percent of Perplexity’s citations come from pages ranking in Google’s top ten ^[7], and pages that answer the core question within their first 100 words appear in 90 percent of top Perplexity citation slots ^[7]. The same shape, retrieve a pool, score candidates at the passage level, then attribute only some of them, recurs across independent reverse-engineering of these pipelines ^{[5] [6]}. The concentration this produces is visible at scale: across a dataset of 36 million AI Overviews, YouTube received about 23.3 percent of citations and Wikipedia about 18.4 percent, because corroboration and multi-format authority dominate the pool far more than any single page can ^[8].

EVIDENCE

The overlap between AI Overview citations and Google’s organic top ten collapsed from 76 percent in July 2025 to 38 percent by February 2026, so 62 percent of citations now come from outside the top ten ^[9]. Citation probability still tilts toward higher positions (33.07 percent at position one versus 13.04 percent at position ten), but pages below position five account for 47 percent of all AI Overview citations ^[9]. AI Overviews also concentrates heavily on a few platforms: Google-controlled domains represent 22.81 percent of all mentions, and Reddit’s citation rate surged roughly 450 percent in three months ^[17].

Google AI Mode is a different engine. It and AI Overviews cite identical URLs only 13.7 percent of the time across 730,000 response pairs, despite reaching semantically similar conclusions 86 percent of the time ^[12]. AI Mode averages 310 citations per query against 51 for AI Overviews, draws from 3,621 unique domains versus 615, and triggered on 100 percent of test queries versus 49 percent for AI Overviews ^[13]. It also weights domain types differently, citing Wikipedia in 28.9 percent of responses versus 18.1 percent and Quora about 3.5 times more often ^[13].

ChatGPT search runs on Bing’s index: 87 percent of its citations match Bing’s top organic results for the same question ^[10]. But Bing rank is a threshold, not a lever. Being in Bing’s top three predicts an actual ChatGPT citation only 6.8 to 7.8 percent of the time ^[10]. After the index threshold, citation is governed by domain trust (pages with a trust score of 97 to 100 average 8.4 citations versus 1.6 for low-trust domains) and freshness (content updated within 30 days receives 3.2 times more citations) ^[16].

Perplexity is the recency engine. It retrieves 5 to 10 candidates and cites only 3 to 4 survivors, and temporal freshness accounts for 44.2 percent of its selection algorithm, the strongest recency bias of any major engine ^[14].

The engines barely agree on sources. ChatGPT’s top-ten most-cited sources are 47.9 percent Wikipedia; Perplexity’s are 46.7 percent Reddit ^[11]. Only 11 percent of domains are cited by both for the same query, and 71 percent of all cited sources appear on a single platform ^[11]. Domain preferences also move under your feet: Semrush’s study of 100 million citations found ChatGPT’s Reddit share collapsing from roughly 60 percent to about 10 percent in six weeks, while Perplexity stayed stable ^[15].

EVIDENCE

In the Princeton GEO study, adding quotations from relevant sources produced a 41 percent relative visibility improvement, the largest single lift of any tactic tested ^{[18] [19]}. Adding verifiable statistics produced a 33 percent lift on the benchmark and 37 percent on Perplexity, making it the most platform-consistent tactic ^[18]. Citing credible external sources inside your own content produced a 28 percent average lift, rising to a 115 percent gain for pages that started at position five ^[18]. Keyword stuffing did the opposite: it performed about 10 percent worse than the untouched baseline, confirming that retrieval runs on semantic matching, not keyword frequency ^[18].

Large-scale correlational work points the same way. Semrush’s study of more than 300,000 AI-cited URLs found clarity and summarization correlated with citation at plus 32.83 percent and clear section structure at plus 22.91 percent, while promotional language correlated at minus 26.19 percent ^[20]. Front-loading is decisive: 44.2 percent of all LLM citations originate in the first 30 percent of page text ^[21]. Cited passages use definitive language 36.2 percent of the time versus 20.3 percent for non-cited passages, nearly double ^[22]. Norg.ai’s analysis found semantic completeness at the passage level correlated with AI Overview selection at r equals 0.87, that pages with 19 or more data points averaged 5.4 citations versus 2.8 for low-data pages, that pages with expert quotes averaged 4.1 versus 2.4, and that sections of 120 to 180 words between headings performed best at 4.6 citations ^[23]. Finally, original data is structurally irreplaceable: an engine cannot paraphrase what exists nowhere else, so derivative or AI-rephrased content that adds no information gain is systematically deprioritized ^[24]. This is why practitioner syntheses of the research keep landing on the same short list: original information, clear structure, and genuine authority, not formatting tricks ^[25].

EVIDENCE

Ahrefs’ study of 75,000 brands found that branded web mentions correlate with AI Overview citations at 0.664, while backlinks correlate at only 0.218, roughly a three to one advantage for mentions over links ^[27]. The distribution is brutally skewed: brands in the top quartile for web mentions averaged 169 AI Overview mentions versus 14 for the next quartile, and 26 percent of brands had zero ^[27]. Citation is also decoupled from rank: about 80 percent of AI-cited sources do not appear in Google’s top ten, rising to 88 percent for AI Mode ^[28]. Most of what gets cited is earned media, which accounts for an estimated 82 percent of AI citations ^[32].

Corroboration is measurable. Brands with positive mentions across four or more unaffiliated platforms were 2.8 times more likely to appear in ChatGPT responses ^[31]. Within E-E-A-T, Google’s raters treat Trust as the gating component, and Experience is the differentiator that AI-generated content cannot credibly fake ^[29]. Entity recognition has an infrastructure layer: Google’s Knowledge Graph draws on Wikidata, and without a structured entity record Google must infer your identity from unstructured text, a less reliable process that can cause disambiguation errors or absence from AI answers ^[30]. Concentration is a risk in itself. When ChatGPT’s Reddit citation share fell from roughly 60 percent to 10 percent in six weeks, every brand that depended on Reddit alone lost ground at once ^[15].

EVIDENCE

OpenAI runs three separate crawlers with distinct robots.txt identifiers: GPTBot for model training, OAI-SearchBot for ChatGPT search indexing, and ChatGPT-User for live user-triggered fetches, so blocking one does not block the others ^[34]. User-triggered fetchers generally ignore robots.txt because they act for a live user, which OpenAI and Perplexity both document ^{[34] [36]}. Perplexity states that PerplexityBot is not used for foundation-model training, only for surfacing and linking sites in results ^[36]. Anthropic offers the most granular separation, letting publishers block ClaudeBot for training while allowing Claude-SearchBot for citation indexing ^[35]. Google-Extended is a control signal layered on Googlebot that governs Gemini training data and does not affect search ranking or AI Overview eligibility ^[38]. Microsoft offers no equivalent split: blocking Bingbot removes a site from Bing and from Copilot answers at once ^[42].

The traffic economics are lopsided. Cloudflare’s crawl-to-refer ratio, which measures content consumed against referrals returned, reached extreme levels for some platforms, while training crawlers ran up to 8 times the volume of search crawling in 2025 ^[37]. Structured data helps but does not guarantee: Google frames it as eligibility for rich results, not a citation lever, while practitioner benchmarks associate FAQ and HowTo markup with a roughly 3.2 times higher likelihood of AI Overview appearance ^[39]. Freshness is now a ranking-adjacent signal for citation. Research by Lily Ray and the Amsive team found that 50 percent of content cited in AI responses is less than 13 weeks old, because every cited URL is pulled from a live index, so a page that drops out of the index drops out of AI answers immediately ^[40]. And rendering is a real constraint: a November 2025 analysis found 69 percent of AI crawlers cannot execute JavaScript, so content that only appears after client-side rendering is often invisible to them ^[41].

EVIDENCE

On llms.txt, Google could not be clearer. Its official AI optimization guide states that you do not need to create new machine-readable files, AI text files, or Markdown to appear in Google Search, and that doing so will not harm or help your visibility ^[26]. Gary Illyes confirmed Google does not support llms.txt and is not planning to, and that to appear in AI Overviews you should simply use normal SEO practices ^[43]. John Mueller called converting pages to Markdown for bots a stupid idea and compared llms.txt to the long-dead keywords meta tag ^[44]. The apparent contradiction, that Chrome’s Lighthouse audits for llms.txt, is a difference in product-team priorities: the Search team’s position governs Search citation, and a browser-readiness audit is unrelated to it ^[45].

Three more myths fall to Google’s own words. Content chunking is not required, because Google’s systems understand multiple topics on a page and surface the relevant piece ^[26]. No special schema is required for generative search, and there is no AIPage type or LLMOptimized property ^[26]. And the FAQ-schema-everywhere tactic is now a zero-return investment, because Google completed the deprecation of FAQ rich results in 2026 ^[46]. Keyword stuffing actively backfires, performing 10 percent below baseline in the controlled Princeton study ^[18]. Finally, the belief that AI-assisted content is penalized on sight is wrong: Google’s policy targets scaled content abuse, the mass generation of low-value pages, not the use of AI itself to research or structure genuinely valuable content ^[47]. Google’s own 2026 AI search guidance reinforces the point by framing AEO and GEO as still SEO, a continuation of established quality practice rather than a separate machine-readable discipline ^[48].

EVIDENCE

The core metric is AI share of voice, your citation frequency against competitors for a fixed prompt set, and it is the most actionable GEO metric available given that engines cite only two to seven domains per response ^[49]. The measurement gap is structural. Google Search Console can filter by AI Overview and AI Mode appearances, but a citation that earns no click is invisible in the data ^[49]. GA4 can capture referral traffic from standalone platforms like ChatGPT and Perplexity through a custom channel group, but AI Overview clicks inside Google Search are recorded as ordinary organic and cannot be separated ^[52]. A Conductor study found that 89 percent of brands cannot properly attribute AI referral traffic, which makes this an industry-wide operational gap, not a tooling quirk ^[52].

A workable method exists. A prompt basket of 25 to 50 queries across category, problem-aware, and comparison intents is the minimum meaningful sample, and fewer than 15 prompts is too variable to act on ^[53]. Because citation patterns shift after model refreshes, the right cadence is weekly pulse checks on a 10 to 15 prompt subset, monthly competitive benchmarking, and a full quarterly audit ^[53]. Tooling helps but must be read with care: independent testing of one snapshot-based tracker found it reported 3 mentions where 123 existed, because the same prompt produces different answers hours apart ^[54]. The prize justifies the effort. AI referral traffic is about 1 percent of total volume but grew 357 percent year over year to 1.13 billion visits in June 2025 ^[55], and it converts far better than organic, with Perplexity referrals converting at about 10.5 percent versus 1.76 percent for Google organic ^[49].