AI Query Fan-Out Infographic: How Searches Become Cited Answers

View Original Article: GEO AI Fanout's Explained for SEO Experts

What this page is, and what it is not

This is a working reference for SEO, GEO, content, digital PR, analytics and technical web teams. Every mechanism described here is either documented by the platform that operates it, an inference supported by published information-retrieval research and labeled as such, or flagged explicitly as unknown. The single most common failure in GEO writing is presenting a plausible mechanism as a confirmed one, so this page keeps those three categories separate throughout.

1. The short answer

Some marketers informally call this behavior "AI fanning." Google's documented term is query fan-out.

Google defines fan-out as a set of concurrent, related, model-generated queries used to gather additional results, and confirms that its generative Search features use its core Search ranking and quality systems to retrieve pages from the Search index. [1]

Other companies document comparable behavior without always using Google's term. ChatGPT Search can rewrite a prompt into one or more targeted queries and issue more specific searches after reviewing initial results. [5] Claude can search repeatedly inside a single request. [11][12] Perplexity Pro conducts multiple searches. [13] Microsoft's Bing Orchestrator historically generated internal queries iteratively, and current Copilot Search says it uses Bing results for the user's query plus additional queries issued on the user's behalf. [8][32]

2. Ten terms that are not synonyms

Most bad GEO analysis is a vocabulary problem. These ten words describe different events in the pipeline, and collapsing any two of them produces a conclusion the data cannot support.

3. A vendor-neutral model of how AI pulls from rankings

Eight stages, combining disclosed platform behavior with standard information-retrieval and RAG research. This is a conceptual map, not a claim that every commercial system implements every box identically.

1 The system decides whether to search

Some surfaces always search; others can answer from model memory. The decision may depend on freshness, complexity, confidence, the selected mode, system instructions, or which tools are available. OpenAI distinguishes fast non-reasoning search, agentic search that may continue after analyzing results, and deep research that may consult hundreds of sources. [6] Anthropic similarly describes Claude searching for current or changing facts while answering stable questions without searching. [12]

This is the first visibility gate, and it is the one most GEO dashboards ignore. If the platform does not activate search, live organic rankings cannot directly supply the answer at all. Model priors, licensed datasets, connected files, product feeds, maps, knowledge graphs, or cached indexes may dominate instead. A measurement program that does not record whether retrieval happened will silently mix two different populations of answers and then average them.

2 The prompt becomes one or more retrieval queries

Google's own example starts with a broad lawn-care prompt and expands it into searches about herbicides, chemical-free removal, and prevention, and it says AI Mode can issue a multitude of queries simultaneously. In the May 2025 AI Mode announcement Google said Deep Search could issue hundreds of searches. [1][4]

ChatGPT Search documents the same broad pattern: a prompt about a current cancer-drug target might first become a general development query, then a narrower conference query after the initial results are reviewed. Approximate location and enabled memory may also affect the rewrite. [5]

Gemini's developer interface is unusually transparent. With Google Search grounding enabled, the model may generate one or multiple queries, and depending on the API interface the executed queries can appear as google_search_call.queries or as webSearchQueries in GenerateContent grounding metadata, alongside retrieved sources and claim-to-source support mappings. [7][36]

The consequence for diagnosis: the visible prompt is an incomplete diagnostic. The platform might never rank pages against that exact sentence. It may rank them against a set of narrower information needs instead.

3 Each query creates its own ranked candidate set

This is where conventional search systems matter most directly, and it is the strongest argument against "SEO is dead." Google states that AI Overviews and AI Mode rely on core Search ranking systems to retrieve relevant, current pages from the Search index, and it documents systems for semantic matching, passage relevance, links and PageRank, originality, freshness, reliability, deduplication, site diversity, and spam detection. [1][3]

Microsoft describes the same inheritance in different language: search and grounding share crawling, understanding, quality signals, and ranking foundations. Traditional search asks which documents a person should visit; grounding asks which specific information an AI can responsibly use to support an answer. [9]

Perplexity has disclosed more of its own Search API architecture than most: it retrieves candidates through lexical and semantic methods, merges them into a hybrid candidate set, removes clearly stale or non-responsive material, then applies progressively stronger ranking stages whose later passes use cross-encoder rerankers scoring both documents and subdocument spans. [14] That is strong evidence about Perplexity's search infrastructure, not a complete formula for citations in every consumer answer.

4 Candidate lists are combined

Once several queries return results, the system needs some way to reconcile them. Commercial platforms do not disclose a universal merge formula. Reciprocal Rank Fusion, usually shortened to RRF, is frequently asserted in SEO articles, but there is no public evidence that Google or OpenAI uses one fixed RRF implementation for ordinary answers.

Academic RAG systems show why a merge stage is technically plausible. A 2025 ACL paper decomposed complex questions, retrieved passages separately for each subquestion, merged the candidate pool, and reranked it before answer generation; that pipeline improved MRR@10 by 36.7% and answer F1 by 11.6% on its evaluated datasets. [16] The finding demonstrates that decomposition plus reranking is useful. It does not reveal any vendor's proprietary implementation.

5 Relevance is re-evaluated for answer utility

The best page for a person to click is not always the best passage for a model to place in a limited context window. A general guide might rank first because it best satisfies broad intent. A lower-ranking official table might contain the exact current number needed for one sentence. A specialist source may cover a rare exception that no top result explains.

RAG research formalizes this gap. RichRAG builds a multi-faceted candidate pool and then uses a listwise ranker to choose a collectively valuable set of documents for the generator. [17] Uplift-RAG notes that documents judged relevant by a retriever are not necessarily helpful for generation, and that misleading material can actively reduce answer quality; its reranker estimates the marginal benefit a document provides beyond the model's internal knowledge. [18] Google Research has likewise described a preference gap between human-friendly retrieval and LLM-friendly context selection. [19]

Together these support a defensible inference: commercial answer engines have strong reasons to optimize the final evidence set for usefulness, coverage and supportability rather than merely preserving the original blue-link order. They do not prove the exact scoring features used by any named product.

6 Passages compete for a limited context budget

Google's passage ranking system can identify relevant sections within a page. [3] Perplexity says it parses pages into self-contained spans that can be individually retrieved and ranked. [14] Claude's current web-search tool can filter search results before they enter the model context, keeping only relevant material. [12]

These mechanisms are consistent with page-level rank being an imperfect predictor. A page can be broadly authoritative and still contain no compact, current support for the needed claim; another page can be weaker overall yet hold the single most useful passage for a subquestion. Note the limit of the evidence: Google has not said that its documented passage-ranking system directly selects AI-answer evidence.

It does not follow that every paragraph should be chopped to a rigid word count. Google explicitly rejects a required tiny-chunk strategy and says there is no ideal page length. [1]

7 The model synthesizes an answer and maps claims to sources

The selected context is combined with the model's internal knowledge and instructions. The model may summarize several sources into one statement, use one source for a number and another for a qualification, or decide a retrieved candidate is redundant. The final answer is therefore not a ranked list translated into prose. It is a generated composition conditioned on a selected evidence set.

Citation mapping is a later layer again. A system can consult more sources than it displays, show a source drawer containing both used and merely related links, or attach more than one citation to a single claim. Citation order should not be read as retrieval rank unless the platform explicitly exposes that relationship.

8 Answer, citations and referrals

What the user sees is the narrow end of the funnel: some prose, a subset of sources, and possibly a link they never click. Each of the four outcomes a business cares about, being cited, being mentioned, being recommended, and being visited, is produced by a different combination of the stages above, which is why they need to be measured separately rather than collapsed into one "AI rank."

4. What each platform actually documents

The pattern is shared. The disclosure is not. Below is what can be cited to a first-party source, and what is simply unavailable no matter how confidently it gets repeated in conference decks.

5. What the overlap studies really show

The published numbers appear to contradict each other. They mostly do not. They answer different questions, about different products, with different denominators.

The best empirical work supports two statements at the same time. First, higher organic rankings are associated with a greater chance of citation, especially in Google AI Overviews. Second, the exact-prompt SERP explains only part of the final citation set. Everything below is an elaboration of those two sentences.

The denominator problem

The word "overlap" can describe at least ten different calculations. Until a report says which one it used, its number cannot be compared with anyone else's.

Why the studies vary so much

Differences can arise from any of the following, and most published comparisons differ on several at once.

No responsible report turns 11.9%, 17% and 37.1% into a trend line without matched methodology.

6. The strongest current academic evidence

Peer-reviewed work is doing something commercial studies mostly do not: repeating collection, reporting variance, and defining its units before it counts anything.

Generative source sets are broader, and far less stable

The 2026 ACL study by Kirsten and colleagues began with a 4,706-query corpus. About 81% of its US queries generated an AI Overview, and cross-engine comparisons were conditioned on queries for which an AIO actually returned a response. [15]

Small query edits move generative sources more than organic ones

A SIGIR 2026 paper introduced an 11,500-query benchmark and analyzed 14,212 queries in total after adding two time-sensitive sets. Its principal source-overlap analysis used the 7,439 queries for which conventional Google results, AI Overviews and Gemini 2.5 Flash all returned sources. Aggregate URL-set Jaccard similarity ranged from 0.11 to 0.18, although individual query-subset comparisons reached 0.24. AI Overviews were less consistent across repeated queries and less robust to minor query edits than traditional results. [25]

Decomposition and reranking measurably help retrieval systems

This is the research that makes the merge-and-rerank stages of the pipeline plausible rather than speculative. It is evidence about RAG architectures in general, not a disclosure about any commercial product.

7. Why rank one gets dropped, and why a non-ranker gets cited

Both directions have specific, identifiable mechanisms. Neither requires believing that rankings stopped mattering.

8. Known, plausible, and still unknown

If you take one table from this page into a client deck, take this one. It is the difference between an argument that survives scrutiny and one that does not.

9. Practical GEO strategy for fan-out

Everything here is compatible with Google's own guidance. Nothing here requires guessing a hidden query and publishing a page against it.

Optimize for a task graph, not a page per phrase

For each high-value customer task, map the likely information needs across these eight dimensions. Treat it as a research taxonomy, not a claim that you have reverse-engineered the platform's hidden queries.

Preserve foundational SEO

For Google, a potential supporting page must be indexed and snippet-eligible. Confirm successful HTTP responses, canonicals, crawl access, internal links, textual availability, renderability and appropriate preview directives. [2]

For other products, crawler access can affect direct discovery or eligibility on the surfaces each provider documents, but it is never sufficient and it does not create a rank boost. OAI-SearchBot [27], PerplexityBot [30] and Claude-SearchBot [28] serve search-related functions distinct from those providers' training bots.

Crawler activity proves only that a bot could fetch a page. It does not prove retrieval, citation, mention, recommendation, traffic or conversion.

10. A rigorous testing protocol

Seven steps. The point of every one of them is to prevent the two standard failures: measuring a moving system once, and collapsing five different outcomes into one number.

Step 1. Define one primary outcome

Choose a single question before collecting any data: did citation prevalence increase, did the entity become part of the answer, did recommendation sentiment improve, did generative impressions or referral visits grow, or did those visits convert? Do not collapse these into one opaque "AI rank."

Step 2. Create a balanced prompt panel

• Include definition, comparison, how-to, evidence, constraint, local, current and troubleshooting prompts.
• Include natural paraphrases.
• Separate locale, language, device and account state.
• Test clean sessions separately from conversational follow-ups.
• Keep a holdout prompt set that did not influence the content edit.
• Save the exact prompt text and the collection conditions.

Step 3. Record both AI and organic data on every run

A rank for a hypothesized subquery is evidence consistent with fan-out, not proof the hidden subquery was executed. Use permitted data sources and respect platform terms when collecting search results. [26]

Step 4. Repeat, then repeat again

Do not use one screenshot per prompt. The ACL 2026 study found large source and decision variability, and in SE Ranking's 2025 experiment the mean three-run exact-URL Jaccard similarity was 9.2%. [15][24]

There is no universal minimum run count. Continue balanced repeated collection until a predeclared precision or stability threshold is reached. Bootstrap at the canonical prompt-family level, with paraphrases and repeated runs nested inside families, so repeated variants are not treated as independent observations. Predeclare either a minimum detectable effect or an acceptable confidence-interval width.

Step 5. Make one attributable change

Replace a vague section with a direct sourced explanation; add original data and a visible methodology; consolidate duplicate pages; repair crawl, index, canonical or WAF problems; clarify an entity and correct supported structured data; or add one genuinely missing subtopic or decision constraint.

Use simultaneous matched control pages or staggered rollouts on identical collection schedules. Confirm recrawling and reindexing before the post-period begins, and remember that confirmation alone does not control for model, index or product changes happening at the same time.

Step 6. Calculate separate metrics

Keep every scheduled run in the dataset, including runs where no AI feature, web search, citation or brand mention appeared. Removing null outcomes inflates visibility.

Step 7. Use native reporting carefully

Covered in the next section. No single report observes the whole pipeline.

11. What the native reports can and cannot tell you

Two first-party reports now exist. They are useful, they are not equivalent, and neither one observes the whole pipeline.

12. Black-hat fan-out myths and failure modes

Understanding fan-out inspires manipulative ideas. These range from unsupported to policy-violating, and the durability of none of them is established.

13. Safe wording for articles, decks and client reports

If you publish about fan-out, these are the sentences that survive a technical reviewer.

AI search may turn one prompt into several related searches, retrieve ranked candidates for each, select the most useful supporting passages, and cite a smaller evidence set. Traditional SEO helps sources enter that process, but the exact-prompt SERP is not the final citation list.

The core statement

14. The model to keep

The right mental model is not "SERP position becomes citation."

Traditional SEO remains essential because access, indexing, relevance, quality, authority, originality and freshness shape the candidate universe. Query fan-out changes where a page can enter that universe. Passage and evidence selection change whether it survives. Synthesis changes whether its claims are used. Citation design changes whether the source is shown. User behavior determines whether visibility creates value.

15. Research method and limitations

Research date: September 12, 2026. Every source was accessible at that date. Where a vendor page is undated, the access date is given instead. If you are reading this months later, treat the platform documentation links as the authority and this page as a snapshot.

16. Sources

Thirty-seven references: first-party platform documentation first, then peer-reviewed research, then commercial studies and benchmarks, then policies, crawlers and reporting. All accessed on or before September 12, 2026.

Platform documentation

1. Google Search Central. Optimizing your website for generative AI features on Google Search. Updated July 10, 2026.
2. Google Search Central. AI features and your website. Updated December 10, 2025.
3. Google Search Central. A guide to Google Search ranking systems. Updated December 10, 2025.
4. Google. AI in Search: Going beyond information to intelligence. May 20, 2025.
5. OpenAI Help Center. Searching the web with ChatGPT. Accessed September 12, 2026.
6. OpenAI Developers. Web search. Accessed September 12, 2026.
7. Google AI for Developers. Grounding with Google Search. Accessed September 12, 2026.
8. Microsoft Bing. Building the New Bing. February 21, 2023. Historical architecture context.
9. Microsoft Bing. Evolving role of the index: From ranking pages to supporting answers. May 6, 2026.
10. Bing Webmaster Tools. AI Performance. Accessed September 12, 2026.
11. Anthropic. Introducing web search on the Anthropic API. May 7, 2025.
12. Anthropic Platform Docs. Web search tool. Accessed September 12, 2026.
13. Perplexity Help Center. What is Pro Search? Updated September 3, 2026.
14. Perplexity Research. Architecting and Evaluating an AI-First Search API. Accessed September 12, 2026.

Peer-reviewed and academic research

15. Elisabeth Kirsten et al. Characterizing Web Search in the Age of Generative AI. Findings of ACL 2026.
16. Paul J. L. Ammann, Jonas Golde and Alan Akbik. Question Decomposition for Retrieval-Augmented Generation. ACL Student Research Workshop 2025.
17. Shuting Wang et al. RichRAG: Crafting Rich Responses for Multi-faceted Queries in Retrieval-Augmented Generation. COLING 2025.
18. Changle Qu et al. Uplift-RAG: Uplift-Driven Knowledge Preference Alignment for Retrieval-Augmented Generation. Findings of EMNLP 2025.
19. Zixuan Ke et al. Bridging the Preference Gap between Retrievers and LLMs. ACL 2024.

Commercial studies and empirical benchmarks

20. Ahrefs. Update: 38% of AI Overview Citations Pull From the Top 10. March 2, 2026. 863,000 SERPs, about 4M AIO URLs.
21. Ahrefs. Only 12% of AI Cited URLs Rank in Google's Top 10 for the Original Prompt. August 11, 2025. 15,000 long-tail prompts.
22. BrightEdge. AI Overviews at the One-Year Mark: Presence, Size, and What They're Citing. February 12, 2026.
23. seoClarity. The Overlap Between AI Overviews and Organic Rankings. October 2025. 362,000 US desktop queries, 5.1M citations.
24. SE Ranking. AI Mode research: Volatility, source patterns, and differences from AIO and organic results. August 29, 2025.
25. Riley Grossman et al. How Generative AI Disrupts Search: An Empirical Study of Google Search, Gemini, and AI Overviews. Accepted to SIGIR 2026.

Policies, crawlers and reporting

26. Google Search Central. Spam policies for Google web search. Accessed September 12, 2026.
27. OpenAI Developers. OpenAI bots. Accessed September 12, 2026.
28. Anthropic Support. Does Anthropic crawl data from the web, and how can site owners block the crawler? Updated April 7, 2026.
29. Google Search Central Blog. Introducing Search Generative AI performance reports in Search Console. June 3, 2026, with a worldwide rollout note dated August 31, 2026.
30. Perplexity Docs. Perplexity crawlers. Accessed September 12, 2026.
31. Bing Webmaster Tools. Bing Webmaster Guidelines. Accessed September 12, 2026.
32. Microsoft. Copilot Search. Accessed September 12, 2026.
33. Google Search Console Help. How impressions, position, and clicks are counted. Accessed September 12, 2026.
34. Kai Greshake et al. More than you've asked for: A Comprehensive Analysis of Novel Prompt Injection Threats to Application-Integrated Large Language Models. 2023.
35. Tobias Pfrommer et al. Manipulating Large Language Models to Increase Product Visibility. 2024.
36. Google AI for Developers. GenerateContent API schema. Accessed September 12, 2026.
37. Microsoft Bing Blogs. Introducing AI Performance in Bing Webmaster Tools: Public Preview. February 10, 2026.