How LLMs Select Citation Sources: The Query Fanout Pipeline Explained

When you ask ChatGPT a question, the answer cites a handful of sources. Behind that answer, the model evaluated hundreds of candidates. The process that determines which sources survive is called query fanout.

Query fanout is the mechanism AI search engines use to decompose your prompt into multiple parallel sub-queries, retrieve results for each, and filter down to the few pages worthy of citation. Understanding this pipeline gives you a direct advantage in Answer Engine Optimization (AEO).

AirOps tracks the sub-queries AI search engines generate behind each prompt, giving you visibility into exactly what LLMs search for before selecting citations. This article walks through the four-stage pipeline that connects a user's prompt to the sources in an AI-generated answer.

What Is Query Fanout and Why Does It Matter?

AI search engines do not run your prompt as a single search query. They decompose it into multiple parallel sub-queries. Each sub-query retrieves its own results. The system then synthesizes an answer from the combined set. This process is query fanout.

The scale of this decomposition is significant. Traditional search queries average 3-4 words. AI search prompts average 70-80 words, a 17-26x increase in query length. Each of those long-form prompts then breaks down into multiple sub-queries that hit search indexes independently.

Google AI Mode uses a custom Gemini 2.5 model specifically designed for fan-out query decomposition. Its Deep Search mode generates hundreds of sub-queries for a single prompt. ChatGPT takes a different approach. AirOps research found that 88.6% of ChatGPT queries generate exactly two fan-out sub-queries per prompt.

Here is what makes this critical for your SEO strategy: 95% of fan-out queries have zero traditional search volume. These are queries that no keyword research tool will surface. They exist only inside the AI search pipeline. You cannot optimize for them unless you know what they are.

This is the fundamental shift query fanout creates. Your content is evaluated against queries you have never seen in GSC. These queries carry no volume in any traditional keyword tool. They determine whether your page gets cited or filtered out.

‍

The Four-Stage Citation Pipeline

Every AI-generated answer with citations passes through four stages. Each stage narrows the candidate pool. Understanding where pages get filtered out tells you exactly where to focus your optimization effort.

Stage 1: Query Decomposition

The LLM takes your prompt and expands it into multiple sub-queries. The decomposition strategy depends on query type. AirOps research shows that informational queries stay near-verbatim 39.9% of the time, meaning the LLM searches for something close to what you typed. Commercial queries behave differently: they split into component parts 38.4% of the time, breaking a product comparison into separate feature-specific searches.

This means a single prompt like "best project management tool for remote teams" generates sub-queries about collaboration features, pricing models, and integration capabilities as separate searches.

Stage 2: Retrieval

Each sub-query hits a search index and retrieves a set of candidate pages. Retrieval position is the strongest predictor of citation. Pages at position 1 earn a 58% citation rate. At position 10, that drops to 14%. If your page does not rank in the top results for the fan-out sub-queries, it never enters the candidate pool.

This is where traditional SEO and AEO converge. Ranking well in organic search directly increases your chances of being retrieved by AI search engines.

Stage 3: Down-Selection

This is the stage where most pages die. 85% of retrieved pages never get cited. The LLM evaluates each candidate against multiple content signals to decide which sources earn inclusion in the final answer.

Title-query overlap is a strong signal. Pages with 50%+ title-query overlap earn a 20.1% citation rate versus 9.3% for pages with lower overlap, a 2.2x advantage. Heading-query match is even more decisive: pages with heading-query alignment earn a 41% citation rate versus 29% for those without it.

Stage 4: Citation

Surviving passages get woven into the final answer. Citation placement follows a consistent pattern: 41% of citations land in the first third of the answer. Each page is typically cited exactly once. The LLM extracts the most relevant passage and attributes it to your page.

‍

What Content Signals Survive the Down-Select Stage?

Once your page clears retrieval, content signals determine whether it gets cited or filtered. AirOps research identifies the specific signals that predict citation success.

Heading-query match is the top content signal. Pages where headings align with fan-out queries earn a 41% citation rate versus 29% for pages without alignment. This makes heading structure one of the highest-ROI optimizations you can make.

Focused pages outperform guides that try to cover everything. Pages covering 26-50% of a topic earn higher citation rates than pages covering 100%. LLMs prefer extractable, specific answers over sprawling overviews. The "write the definitive guide" strategy works against you in AI search.

Word count has a clear sweet spot. Pages between 500 and 2,000 words perform best. Pages over 5,000 words underperform. Longer content dilutes the signal the LLM is trying to extract.

Schema markup creates a measurable advantage. JSON-LD structured data adds a 6.5 percentage point lift to citation rates. FAQPage schema earns a 45.6% citation rate, one of the highest-performing schema types.

Domain Authority (DA) does not predict citation success the way you would expect. Pages from domains with DA 20-40 earned more citations than pages from DA 80-100 domains (26.0% vs 25.4%). Authority helps with retrieval. It does not help with down-selection.

Sequential heading structure delivers a 2.8x citation rate boost. Pages with logically ordered H2 and H3 headings that follow a clear progression make it easier for LLMs to extract and attribute specific passages.

‍

How to Optimize Your Content for Query Fanout

Knowing how the pipeline works gives you a clear optimization framework. Here are the highest-impact actions you can take, ordered by the stage they target.

Start With Retrievability

Retrieval position is the strongest single predictor of citation. A page that ranks position 1 for a fan-out query has a 58% citation rate. No amount of content optimization overcomes poor retrieval. Focus on ranking well for the specific sub-queries LLMs generate, not just your target keywords.

Match Headings to Queries

Heading-query match delivers a +12 percentage point advantage at equal retrieval rank. Structure your H2 and H3 headings to mirror the exact phrasing of questions your audience asks. Use AirOps Insights to see the actual fan-out sub-queries AI engines generate for your target prompts, then align your headings to match them.

Stay Focused, Not Exhaustive

Target 26-50% topic coverage per page. LLMs prefer specific, extractable answers. If you have a 5,000-word guide, consider splitting it into three focused articles of 1,200-1,800 words each. Each page earns its own retrieval opportunities and gives the LLM a cleaner extraction target.

Track Fan-Out Queries

32.9% of citations come from fan-out Search Engine Results Pages (SERPs) only, meaning the original prompt SERP did not surface these pages. If you are optimizing only for the primary query, you are missing a third of your citation opportunities. AirOps tracks these invisible sub-queries so you can build content that targets them directly.

Structure for Extraction

Sequential headings deliver a 2.8x citation boost. Lists, tables, and schema markup all increase extractability. Add JSON-LD structured data for a 6.5pp citation rate lift. Use FAQPage schema on pages that answer specific questions.

‍

How Query Fanout Differs Across AI Platforms

Each AI search platform implements query fanout differently. These differences affect which content gets retrieved, how it gets evaluated, and how many opportunities you have for citation.

• Google AI Mode uses a custom Gemini 2.5 model built specifically for query decomposition. Standard queries generate a moderate number of sub-queries. Deep Search mode generates hundreds, casting the widest retrieval net of any platform. Google's retrieval pulls from its own search index, giving pages with strong organic rankings a direct advantage.
• ChatGPT routes fan-out queries through Bing's search index. Most prompts generate exactly two sub-queries. This narrower decomposition means fewer retrieval opportunities per prompt, making each ranking position more valuable. Content that ranks well in Bing has a structural advantage in ChatGPT citations.
• Perplexity runs parallel searches across multiple indexes simultaneously. Its retrieval model is aggressive, pulling candidates from diverse sources. Perplexity tends to cite more sources per answer than ChatGPT, giving your content more chances to appear if it ranks across multiple indexes.

The optimization implication is clear: a single content strategy does not work across all AI platforms. You need to track how each platform decomposes queries for your target topics and optimize accordingly. AirOps Insights shows you fan-out behavior broken down by platform so you can tailor your approach.

‍

Key Takeaways

• AI search engines decompose every prompt into multiple sub-queries through query fanout. 88.6% of ChatGPT prompts generate exactly two fan-out sub-queries.
• 95% of fan-out queries have zero traditional search volume. You cannot find them in keyword research tools.
• Retrieval position is the strongest citation predictor: 58% citation rate at position 1 versus 14% at position 10.
• 85% of retrieved pages never get cited. Down-selection is where most content fails.
• Heading-query match (+12pp), sequential heading structure (2.8x), and focused coverage (26-50%) are the top content signals for surviving down-selection.
• Domain Authority does not predict citation. DA 20-40 sites earned more citations than DA 80-100 sites.
• 32.9% of citations come from fan-out SERPs only. Track sub-queries with AirOps to capture these hidden opportunities.

AirOps for Query Fanout Visibility

AirOps Insights shows you the exact fan-out sub-queries AI search engines generate for your target prompts. You can see which sub-queries surface your content, which ones surface competitors, and where your content gets filtered during down-selection. This turns the query fanout pipeline from a black box into an optimization framework you can act on.

See how AirOps tracks query fanout for your brand. Book a call.

Frequently Asked Questions

What Is Query Fanout in AI Search?

Query fanout is the process AI search engines use to decompose a single user prompt into multiple parallel sub-queries. Each sub-query retrieves its own set of candidate pages. The LLM then evaluates all candidates and selects which sources to cite in the final answer.

How Many Sub-Queries Does ChatGPT Generate per Prompt?

AirOps research found that 88.6% of ChatGPT queries generate exactly two fan-out sub-queries. Google AI Mode generates more, with Deep Search producing hundreds of sub-queries per prompt.

Does Domain Authority Affect AI Citation Rates?

DA helps with retrieval (getting into the candidate pool) but does not predict citation success during down-selection. Pages from DA 20-40 domains earned a 26.0% citation rate versus 25.4% for DA 80-100 domains.

What Is the Best Word Count for AI Citations?

Pages between 500 and 2,000 words perform best for AI citations. Content over 5,000 words underperforms because it dilutes the extractable signal LLMs are looking for. Focused, specific pages outperform sprawling guides.

Can You Track Which Fan-Out Queries Surface Your Content?

Yes. AirOps tracks the exact sub-queries AI search engines generate behind each prompt. This gives you visibility into the invisible queries that determine whether your content gets cited, queries that carry zero traditional search volume and do not appear in GSC or keyword research tools.