What Query Fan-Out Actually Is
When Elizabeth Reid, Google’s Head of Search, introduced AI Mode at Google I/O 2025, she described the mechanism directly: “Search recognises when a question needs advanced reasoning. It calls on Gemini to break the question into different subtopics and issues a multitude of queries simultaneously on your behalf.” That mechanism is query fan-out. It is not a future feature. It is running right now, on every AI Mode search, and on every ChatGPT and Perplexity query that triggers web retrieval.
The shift it represents is structural. Traditional search was one-to-one: one query, one set of results, one ranking to compete for. Then search evolved to many-to-one: semantically equivalent queries returning the same results. AI search is one-to-many: one query expands into multiple parallel searches, with results scored and fused before the AI composes its answer. Most businesses are still operating with a one-to-one strategy in a one-to-many world.
The Building Your Buyers Are Already Exploring
Think of search visibility as a building under construction. The ground floor has been standing for twenty-five years: traditional Google, established rules, a known inspection regime. Every competent practitioner understands the code. Most businesses that have invested in SEO have some presence here.
The first floor opened recently. Google AI Overviews and AI Mode are built directly on the ground floor’s index — same foundations, different logic applied above them. Many businesses have found their way here without realising it: their well-indexed pages appear in AI Overviews as a byproduct of existing SEO work. But appearance here is not automatic, and as AI Mode intercepts a growing proportion of searches, the cost of being absent rises.
The second floor exists, and your buyers are already on it. Standalone LLMs — ChatGPT, Perplexity, Claude — access this floor through what is currently a temporary ladder: a real-time web retrieval layer that is still being formalised into something permanent. The rules on this floor are not yet codified. But buyers are asking questions here, and AI systems are retrieving answers and making recommendations.
Query fan-out is the mechanism that determines which rooms get opened when a buyer asks a question. A business with well-furnished, clearly labelled content across all three floors — extractable, attributed, entity-declared — gets found. A business with one impressive room on the ground floor and nothing accessible above it gets bypassed. Floor by floor. Sub-query by sub-query.
What Your Buyers Are Actually Searching When They Search for You
This is where most businesses and SEOs miss the point entirely. Your buyers do not search using the language on your service pages. They type their problem, in their language, at the moment they have it. Fan-out then generates the sub-questions they would ask next — not the questions your marketing team would answer in a brochure.
Consider a prospective client who needs to speak to the right employment lawyer — not a firm in general, but a solicitor with specific experience in constructive dismissal cases where the employer is a technology company. They do not type “employment law firm London.” They type: been constructively dismissed by tech company what are my options.
Fan-out generates: constructive dismissal success rates UK, employment tribunal claims against tech employers, how to find the right employment solicitor, what evidence is needed for constructive dismissal, employment solicitor fee structures, Law Society accredited employment specialists, constructive dismissal time limits.
If your law firm has one “constructive dismissal” service page — no named solicitors, no case outcome data, nothing on evidence gathering, no Law Society directory presence, no FAQ content covering tech employer scenarios — you are invisible to six of those eight sub-queries. The AI does not recommend you. Not because you are not capable. Because it cannot verify you are relevant across the dimensions it is checking.
Fan-out does not just retrieve your content. It cross-examines your topic space against the full range of questions a buyer in that moment actually has. That is a fundamentally different competitive challenge than ranking for a keyword.
What SEOs and Businesses Are Getting Wrong
Seven gaps account for the majority of fan-out invisibility. Most are not new problems — they are existing weaknesses that fan-out makes immediately consequential.
| The gap | What most businesses do | What fan-out requires |
|---|---|---|
| Competing for keywords | Targeting the primary search phrase | Occupying the full topic space the fan-out generates — themes and intent types, not strings |
| Optimising pages, not passages | Writing for the page as a whole | Every H2 section must be extractable in isolation — fan-out retrieves at passage level, not page level |
| Ignoring implicit questions | Answering what buyers explicitly ask | Mapping and covering the questions buyers will be asked next — the sub-queries they never type |
| Weak third-party presence | Strong site, minimal external footprint | High-stakes fan-outs always generate trust signal sub-queries — reviews, directories, editorial mentions |
| One angle per topic | Covering a topic from one perspective | Fan-out checks comparative, recency, next-step, and trust angles simultaneously |
| Generic content, one voice | One page covering all audiences | Enterprise, B2B services, SaaS, and B2C fan-outs generate entirely different sub-query patterns. Research across 35 LLMs confirms that when similar entities use identical language, the model’s ability to distinguish between them degrades — commodity positioning is a retrieval failure mode, not just a brand problem (Wang and Sun, NYU/UVA, July 2025) |
| Google-only thinking | Optimising for AI Overviews only | ChatGPT, Perplexity, and Copilot all use fan-out — each with different depth and different trust signal requirements |
Gap three — ignoring implicit questions — causes the most commercial damage and gets the least attention. The 73% uniqueness finding makes it structural: because the majority of fan-out sub-queries are unique each time the same prompt is entered, you cannot build a list of strings to target. A second layer of evidence reinforces this: RAG systems lose approximately 40% of retrieval accuracy when users rephrase the same question (Gloaguen et al., arXiv:2602.11988) — meaning the AI itself gets different results on the same topic when phrasing shifts. String-targeting is not just ineffective; it is structurally impossible. You have to understand the type of question your buyers have around your topic, then ensure your content addresses each type comprehensively regardless of exact phrasing. This is not a keyword strategy. It is a topic coverage strategy.
How Fan-Out Plays Differently Across Business Types
Fan-out depth and character is not uniform. It scales with query complexity, decision stakes, and the information gap the AI needs to resolve before it can give a confident answer. Understanding which fan-out pattern applies to your buyers is the first strategic decision.
| Audience | Typical trigger query | Fan-out generates | What you need to be visible |
|---|---|---|---|
| Enterprise / Regulated | “compliance training software UK financial services” | Vendor credibility, FCA/ICO regulatory alignment, security certifications, named client case studies, G2/Capterra reviews, integration capability, procurement policy fit | Named clients, certification pages, G2 profile, compliance-specific content, integration documentation — all indexed and attributed |
| B2B Services (law, accountancy, consulting) | “employment solicitor constructive dismissal tech company” | Named practitioner credentials, case outcomes, fee transparency, Law Society / ICAEW standing, client reviews, scenario-specific FAQ content | Named solicitor / partner profiles, outcome-oriented case studies, professional directory presence, scenario-specific content (not just service category pages) |
| B2B SaaS / Tech | “alternatives to [competitor]” or “best [category] for [use case]” | Feature comparison, pricing transparency, integration depth, onboarding complexity, customer reviews, migration effort, support quality | Comparison pages, G2/Capterra profile, integration library, pricing clarity, migration guide — absent from any of these means absent from the shortlist |
| B2C Considered Purchase | “best private school Hampshire pastoral care” or “private healthcare knee surgeon Southampton” | Ofsted / CQC data, outcome statistics, fee structure, specialist credentials, parent or patient reviews, comparison with alternatives | Factual, attributed content across each dimension; third-party editorial mentions; review platform presence; named specialists with verifiable credentials |
Notice what is consistent across all four rows: named entities and third-party presence. The source types each platform retrieves from differ significantly even for the same query — only 12% of cited sources match across ChatGPT, Perplexity, and Google AI (Passionfruit + Ahrefs, 15,000 queries). ChatGPT draws from first-party sites, editorial listicles, G2/Clutch profiles, and Reddit threads. Gemini uses editorial sources primarily, largely ignores first-party sites, and relies heavily on training data without URL citations. Perplexity has the widest source range — Reddit, comparison sites, review platforms, and real-time web — making it the most responsive to direct citation work. Optimising for one platform’s source pool while ignoring the others means you are visible in one and absent from two. These are not vanity signals. They are what trust sub-queries — which high-stakes fan-outs generate almost universally — are checking. An AI recommending a vendor, a solicitor, a school, or a surgeon is performing a verification step against external sources it cannot control. If those sources do not mention you, or mention you inconsistently, the verification fails and the recommendation does not happen.
For a breakdown of which AI platforms matter most for each audience type, and the different fan-out patterns each platform applies, see the AI Platform Priority by Audience guide.
The Mechanism: Why RRF Scoring Changes the Rules
When AI systems merge results from multiple fan-out sub-queries, they use Reciprocal Rank Fusion. Every document is scored based on its position across all result lists. A document appearing at position 2 in one list and position 5 in another scores 1/2 + 1/5 = 0.7. A document appearing in only one list at position 1 scores 1.0 for that list alone.
But a document appearing at position 4 across three different sub-query result lists scores 3 × (1/4) = 0.75 — and that is higher than most single-list top rankings. Consistency across the topic space outscores excellence on a single keyword. The Perplexity pipeline research confirms this is binary, not graduated: a document either passes all five citation gates and earns a citation, or it is invisible. There is no position seven. Unlike Google’s graduated scale where position 7 still sends traffic, Perplexity’s model is cited-or-invisible — which makes the passage-level quality tests decisive rather than marginal. This is the rule change that most SEOs have not yet absorbed.
The implication for content is exact. Fan-out sub-queries retrieve passages, not pages. An AI scanning your page for a specific sub-query angle finds the relevant H2 section and evaluates it in isolation: does it open with a direct, self-contained answer? Does it include a specific, attributed data point? Is the named entity declared within the section? Could this passage be cited as a standalone answer without the surrounding context?
A page where every H2 section passes those tests contributes RRF score across multiple sub-query angles simultaneously. The Perplexity pipeline research makes this quantitatively precise: 90% of top-cited sources answer the core question within the first 100 words (LLMClicks, 2026). This is not a formatting preference — it is the location where Perplexity’s retrieval system actively looks for the answer fragment during candidate scoring. Additionally, citations in Perplexity are not retrofitted after the LLM writes its answer. They are pre-embedded during context assembly — assigned to source documents before generation begins. If your passage does not survive the retrieval and ranking stages, no amount of synthesis quality will cite it. A page where two sections pass and four do not contributes nothing to the other four sub-queries — regardless of domain authority, regardless of how strong the page looks as a whole. This is what the CITATE framework addresses at the passage level: fan-out does not make CITATE’s criteria best practice, it makes them mechanically necessary. For the full page-level blueprint, the AI Page Anatomy guide maps exactly how each section should be structured.
What to Do: The Prioritised Sequence
Fan-out optimisation builds on what already exists. Attempting the second floor without the first produces content that looks comprehensive but is invisible because the entity foundations are missing. The sequence matters.
Before any fan-out work: verify your core pages are indexed, your entity is correctly declared in schema.org markup, and your business is consistently represented across the external sources trust sub-queries will check — Wikidata, your industry directory, Google Business Profile, at minimum one credible review platform. This is ground floor work. It is the prerequisite. The AI Visibility Action Plan sequences the full diagnostic by business type.
For topic coverage: use the fan-out mapping approach in the HowTo section below. For each priority topic, the question is not “what keywords should I rank for” but “what implicit questions does a buyer have when they encounter this topic, and which of those does my current content not answer clearly.” Each unanswered implicit question is either a new page or an existing section to fix. The AI Discovery Stack maps where topic coverage sits in the overall visibility architecture — Layer 3, above entity foundations and Bing indexation, below the trust and recommendation layers.
For passage-level extraction: audit your existing H2 sections against the CITATE criteria. A section covering the right topic but opening with “As we discussed above…” or burying its data point three paragraphs in, or failing to name the entity the section is about, will not be retrieved for a relevant sub-query. Fixing these failures is the lowest-effort, highest-return fan-out work available. For the off-site signals that trust sub-queries generate, see entity corroboration and AI Citation Dominance.
