How Google AI Overviews Select & Cite Sources: Query Fan-Out, Conversational Prompts & Passage Extraction (400-Query Commercial SERP Dataset & Industry Synthesis)

Summary Open dataset on Google AI Overviews and AI Mode: of 1,116 top-three organic results across 372 US searches, 80.5% were not among the first three AI citations. The audit covered 400 US searches in four commercial sectors, and 372 of them showed an AI answer. In 200 of those 372 (53.8%), none of the top three organic pages was cited. The deposit includes a 12-slide research briefing (PDF) and the two CSV files behind it. The full web study, a 15-slide practitioner checkmap, and a short video walkthrough are linked at the end. How Google’s AI answers find their sources People now type longer, more specific questions and can ask follow-ups inside Google’s AI answers. Google Search Central’s documentation on AI features says AI Overviews and AI Mode can use a technique called query fan-out, which runs several related searches at once and draws supporting pages from across them. In practice, a page can rank first for a short keyword and still be passed over when someone asks about a price, a symptom, or a standard that the page does not state clearly. Our working model This is a simplified model we used to organize the audit. It is not Google’s official pipeline. The searcher’s question is split into sub-questions: definition, price, specifications, symptoms, standards, comparisons. Related searches run for each sub-question. Passages that answer a sub-question cleanly, such as a short answer, a table, or a numbered list, are easier to quote than long narrative sections. Off-site signals such as YouTube videos and third-party mentions may help confirm who is behind the content. What other studies report Semrush’s 2026 study of 600,000+ SERPs reports AI Overviews on 88% of commercial evaluation searches, up 71% across commercial queries. Pew Research Center found people clicked a traditional result in 8% of visits when an AI summary appeared, compared with 15% when it did not. Ahrefs found that the average click-through rate of the top-ranking page was 58% lower on keywords with an AI Overview (December 2025 data, 300,000 keywords). This is a correlation, not a controlled test. Seer Interactive reports that brands cited in an AI Overview earned 35% more organic clicks and 91% more paid clicks than uncited brands on the same results. This is an observed difference, not a controlled test. Aggarwal et al. (ACM SIGKDD 2024) found in their generative-engine benchmark that adding statistics, authoritative citations, and quotations improved source visibility, while keyword stuffing did not. They did not test Google AI Overviews directly. Ahrefs’ 75,000-brand study found that YouTube mentions (0.737), digital PR mentions (0.664), and branded anchors (0.527) correlated more strongly with AI visibility than backlink volume did (0.218). These are correlations. What we found in the 400-query audit Scope We ran 400 US queries, 100 each in Local Home Services, E-Commerce and Retail, B2B Commercial and Industrial, and SaaS and Enterprise Software. For each query we recorded the top three organic results and the first three AI citation cards. 372 queries (93.0%) showed an AI answer and 28 (7.0%) did not. Of the 372 answers, 215 were standard AI Overviews and 157 were AI Mode responses. Organic rank versus AI citations Across 1,116 top-three organic slots, 182 (16.3%) were cited with the exact same page, 36 (3.2%) were cited through a companion guide on the same domain, and 898 (80.5%) were not cited among the first three cards. Organic position Exact page cited Companion guide cited Not cited #1 (372 slots) 70 (18.8%) 16 (4.3%) 286 (76.9%) #2 (372 slots) 63 (16.9%) 8 (2.2%) 301 (80.9%) #3 (372 slots) 49 (13.2%) 12 (3.2%) 311 (83.6%) All (1,116 slots) 182 (16.3%) 36 (3.2%) 898 (80.5%) At the query level, 200 queries (53.8%) cited none of the top three organic pages, 129 (34.7%) cited one, 40 (10.8%) cited two, and 3 (0.8%) cited all three. Results by sector Sector Queries with an AI answer Top-3 slots Exact page Companion guide Combined Local Home Services 86 of 100 258 45 6 51 (19.8%) E-Commerce and Retail 99 of 100 297 40 20 60 (20.2%) B2B Commercial and Industrial 87 of 100 261 54 8 62 (23.8%) SaaS and Enterprise Software 100 of 100 300 43 2 45 (15.0%) In e-commerce, 257 of 297 category or product slots (86.5%) were not cited directly, and another 20 slots were cited through a companion buying guide on the same site. B2B had the highest combined rate and SaaS the lowest. Examples from the files: for “net revenue retention formula”, Stripe was cited while the first organic result, Klipfolio, was not. For “mattress firmness guide”, NapLab was cited while Mattress Mill ranked first. For “what is ltl freight”, Old Dominion Freight Line was cited while GoComet ranked first. For “why is my ac blowing warm”, Airflow Designs was cited while Valiant Energy ranked first. Who got cited The first three AI cards produced 1,100 citation URLs across 850 unique domains, about 1.29 citations per domain. Sixteen citation slots had no URL because some answers showed fewer than three cards. YouTube accounted for 82 of the citations. We labeled the main structure of each of the 1,100 cited pages by hand, one primary label per page: Primary format of the cited page Citations Share Direct short answer under a question heading (about 40 to 60 words) 353 32.1% HTML pricing or specification table 299 27.2% Ordered symptom-to-action checklist 185 16.8% Timestamped video (including 82 YouTube citations) 132 12.0% Named official standards or benchmarks 131 11.9% These labels are our own classification. They are not a column in the CSV files. How we read these patterns Cited pages tended to state a number, a rule, or a step plainly in one self-contained block. We treat that as a practical guide for writing pages, not as proof that any one format causes a citation. The full playbook is in the web study linked below. Putting these patterns into practice usually means working on three layers together: page structure, entity and trust signals, and off-site mentions. Our overview of that work is on the AI search optimization page. Scope and limitations This is an observational snapshot of US English searches from one collection period. AI answers change often. We recorded only the first three citation cards. Some AI answers show more sources. The audit describes what cited pages looked like. It does not test whether adding those formats to a page causes a citation. The “Why Ranked” and “Why Cited” columns in the CSV files are analyst notes and are best read as hypotheses. Dataset guide File 1: SERP Analysis 2026 - SERP.csv 1,200 rows and 10 columns: the top three organic results (Pos #1 to #3) for each of the 400 queries. Target Keyword, Industry, and AI Overview Triggered? appear on the first row of each three-row block. Column I, Cited in AI Overview?, uses three kinds of values: YES (Citation #1, #2, or #3) for the exact page (182 slots), YES (Citation #1, #2, or #3 — Domain Cluster) for a companion page on the same domain (36 slots), and NO (Organic Only) (898 slots). The 28 queries without an AI answer account for the remaining 84 rows. File 2: SERP Analysis 2026 - Google AI Overview.csv 1,200 rows and 10 columns: Citation #1, #2, and #3 for each query, with the AI answer text (column D), cited brand (F), cited URL (G), in-text anchor (H), page structure (I), and analyst rationale (J). It holds 1,100 cited URLs across 850 unique domains. Sixteen slots in answers with fewer than three cards, and the 84 slots for queries without an AI answer, have no URL. How to reproduce the counts The counts of citations, skipped slots, sectors, and domains can be reproduced by filtering these two files. The format shares in the table above come from hand labeling. Cite this dataset Rahman, M. M., & Invrse Marketing Research Team. (2026). How Google AI Overviews Select & Cite Sources: Query Fan-Out, Conversational Prompts & Passage Extraction (400-Query Commercial SERP Dataset & Industry Synthesis) [Data set]. Zenodo. Companion materials Strategy to Get Cited in Google AI Overview (2026), the full web study AI Overview Citation Checkmap, 15-slide PDF How to Get Cited in Google AI Overviews, video walkthrough (5:13) Invrse Marketing LLC AI search optimization at Invrse Marketing