A practical guide to what fan-out actually does inside a RAG pipeline, the misconceptions leading SEOs astray, and how to build strategy that works with the process instead of against it.
This post covers how query fan-out actually functions within RAG processes, the three most consequential misconceptions circulating right now, and what a more well-rounded, practical approach looks like. It focuses on Google’s RAG process, since AI Overviews is perhaps the most disruptive surface in terms of information and answer-seeking traffic and its impact on businesses, AI Mode, and Gemini’s chat interface. ChatGPT works slightly differently and will need to be its own post.
What Query Fan-Out Actually Does
Before diagnosing where the thinking goes wrong, you need a precise model of what fan-out is doing — not just “it breaks queries into sub-queries,” but why, when, and what those outputs are used for. This section looks at key points in the QFO process to help you understand what the LLM is trying to achieve at each step and how that can influence content strategy and writing style if your goal is to be more visible in AI/LLM answers.
Fan-out isn’t designed to find the best page on a topic. It’s designed to collect the best passage for each discrete information need the original query implies. The unit of retrieval is the passage, not the URL.
The model generating sub-queries is essentially asking: “If I were building a complete, well-grounded answer to this question, what specific pieces of information would I need, and where might I find them?”
Sub-queries are the information architecture of the final answer, written in advance.
When Fan-Out Triggers
Fan-out is not universal. It fires based on conditions:
- Intent complexity.
- Simple, high-confidence factual lookups may not trigger fan-out. The model may answer from training data without touching the web.
- Recency requirements.
- Queries containing signals like “latest,” “current,” or time-sensitive topics trigger web grounding.
- Confidence thresholds.
- Some implementations have configurable thresholds — if the model’s internal confidence exceeds the threshold, no retrieval occurs.
- Query type.
- Comparative, evaluative, and exploratory queries almost always trigger fan-out. Narrow factual queries often don’t.
Understanding when fan-out triggers is as important as understanding what it does. It tells you which of your content is even in play.
Before the Fan-Out Happens
When a user submits a prompt to an AI search system such as ChatGPT with web search, Perplexity, Google AI Mode, or Gemini, the system runs a RAG pipeline. Most SEO writing summarizes this as “it searches the web and synthesizes an answer.” That glosses over the steps where your content either survives or gets dropped.
Prompt classification and chunk size allocation
Before any retrieval occurs, the model classifies the incoming prompt. This step is rarely discussed in SEO coverage, and it matters.
Classification has two major outputs.
First, it determines whether retrieval is necessary at all. For high-confidence factual queries where training data is sufficient, no web retrieval occurs. The fan-out never fires and your content is never in play, unless it’s already in the training data (though that likely means it won’t be cited).
Second, it sets the size and number of retrieval chunks allocated for the retrieval and generation process. The number varies based on the prompt, but the retrieval sizes are either 512 or 1024 tokens (tokens are roughly ¾ of a word each). The system decides how much to budget before fetching a single document, and that budget is set by prompt classification.
A simple factual lookup gets a small retrieval token size: retrieve one or two brief, confirming passages, then generate a brief answer. A comparative or exploratory query gets a substantially larger window, because the model anticipates holding and cross-referencing evidence across multiple dimensions.
Why this matters for content strategy:
Query classification directly influences how many retrieved passages the model evaluates simultaneously, how many sub-queries the fan-out generates, and how many citation slots are available in the final answer. The retrieval chunk size limit (512 vs 1024) determines the size and type of passage the LLM is looking for on your site. Narrow window = fierce competition for very few slots. Wide window = more citation slots, but more candidates competing for them. That’s just the first takeaway.
It’s critical to know what prompt or prompts your content is attempting to optimize for because retrieval size matters. Not just from a competition standpoint, either. Smaller context windows are expecting short, clear, factual information nuggets. They’re more likely to be seeking simple validating statements rather than trying to understand and synthesize information about a complex or nuanced topic.
Meanwhile, larger retrieval sizes are evaluating multiple sources for relevance, factuality, and consensus. They’re disambiguating entities and evaluating trust. They’re providing more opportunity for discovery, context and entity-attribute identification.
Notably, queries classified with a shopping intent (looking for specific products, browsing categories, or asking for recommendations) are classified into searchable, conversational, or browse-based intents. Retrieval window size for shopping intent queries can shrink to as little as 256 tokens (about 190 words). This happens because shopping intent queries typically need to identify very specific products, often using a brand and/or product name or a specific product attribute. Since the LLM doesn’t need to understand and synthesize a concept to the extent it does for more complex queries, it saves memory and compute by reducing the retrieved chunk size and relying on these keys to identify relevant passages. Shopping queries also tap into the Google Product Graph, offering another dimension of context and information, which simplify the retrieval process.
Key Takeaways for Content Creators and Marketers
- Shape your content to the expected retrieval chunk size.
- Simple questions should have simple answers
- Retrieval uses a rolling window based on chunk token size, with a 256 (190ish word) overlap. If the answer isn’t in one chunk, ensure it’s nearby. This is where brevity and clarity are key.
The Fan-Out Isn’t the Guide
This advice has spread in various forms: reverse-engineer the sub-queries AI systems generate, then create dedicated content targeting each one.
The logic seems sound on the surface. There are a few significant problems with it.
Problem one: Fan-out queries are dynamic, not fixed
Sub-queries are generated fresh for each prompt, conditioned on user context, session history, recency signals, and the model’s evolving capabilities. The Surfer SEO study found that only 27% of fan-out sub-queries remain consistent across repeated searches for the same prompt, meaning 73% vary each time. Ahrefs noted that AI Overview citations changed every two days. A content strategy built around a specific enumerated list of fan-out queries has a built-in expiration date.
Problem two: Dedicated sub-query pages often perform worse than comprehensive coverage in a single well-structured piece
The passage retrieval mechanism in RAG doesn’t need a dedicated URL for each sub-query. It needs a passage, a coherent chunk of text that cleanly answers the sub-query, wherever that passage lives. A single well-structured piece that addresses a topic comprehensively at the passage level is more retrievable than ten thin pages each targeting one facet. The comprehensive piece accumulates more link authority, more engagement signals, and more topical trust. and passage chunking is agnostic to URL.
I think this one may just be a remnant of the “hub and spoke” content model, with SEOs and content writers confusing two different approaches and purposes.
Problem three: Over-focus on fan-out isn’t for the user
If all you’re doing is capturing the fan out queries and then adding them and their answer to your content, you’re not writing for the user. You’re writing for the LLM.
The query fan-out is predominantly for the LLM. Many of the answers retrieved from fan outs are presented without citation or aren’t presented at all. Therefore, cramming them into your content is adding unnecessary depth and redundant information for the user. It’s like opening a recipe page and having to read the history and definition of every ingredient before you get to the actually recipe steps. The user has context and knowledge. The LLM generally does not.
How Content Creators Should Be Thinking About Implementing Query Fan Out Answers
In my opinion, the biggest change is also not a change at all. We need to clearly delineate what processes are for the machine and which are for the user. The only way you can do that is by having clarity on who your audience is. What do they know vs what do they need to know? And then, what does the LLM need to know in order to get the information to the user.
I’m not saying don’t write for the machines, because I think to a small extent you have to. What I’m saying is, don’t make it the focus. Weed out the irrelevant parts. If your audience is topically aware, but maybe not brand aware, then don’t waste time rehashing what they already know. Get to work informing them about the brand and positioning it relative to competitors. You’ll likely find you need less content and less lengthy content than you expect.
Here are a few other ways to change your thinking about query fan out.
Think in Passages, Not Pages
Every H2 and H3 section is a retrieval candidate. A well-organized page with distinct, well-labeled sections gives the chunking mechanism clean breaks. Each section can be retrieved independently, meaning a single page can be cited for multiple sub-queries.
Atomic facts matter more than thematic coverage. A passage that states a specific claim, supports it with a verifiable figure or mechanism, and attributes it clearly is more citable than one that covers a theme broadly. Specificity and groundedness are what make a passage selectable.
Passage independence is a feature. A passage that can be understood without reading the rest of the document is more usable by a retrieval system than one that depends on context established elsewhere. Each section should stand on its own as a coherent answer to a specific question.
Grounding and Verifiability Directly Affect Selection
Fan-out’s core purpose inside RAG is to reduce hallucination by grounding responses in verifiable external sources. The selection mechanism is sensitive to this.
- Specificity over scope. A passage explaining exactly how server-side caching reduces TTFB with a concrete example is more selectable than one asserting that performance optimization improves user experience.
- Verifiable claims over assertions. Content grounded in evidence, data, or traceable mechanisms is more useful to the retrieval layer than content that makes claims without support.
- Clear entity relationships. Clear subject-predicate-object structure in your prose, not keyword density, is what allows the model to correctly classify what a passage answers and retrieve it for the right sub-queries.
Leverage Entities and Attributes
You have to think in terms of the graph. Analyze the target prompt, the questions and intent your page is designed to serve. What are the relevant entities? What are the most sought-after, most relevant attributes of your related entities? How well and to what extent are you folding in attributes and specifically mentioning relevant entities in your text? Pairing attention to those elements while also addressing the needs of the query fan-out can improve your odds of retrieval, specifically with your target audience.
I built a fun little tool that decomposes Query Fan Out and identifies:
- Queries seeking disambiguation
- Hello, is it you I’m looking for?
- Entity Attributes
- What qualities should I be optimizing for/ addressing?
- Journey stage
- What is the context of the user? What are their needs? How should I shape answers?
- Trust
- Are you a source worth sharing?
- Comparison
- Identify who the LLM thinks your competitors are and speak to the differences between your product and theirs. Don’t shy away from mentioning other brand names, specifically if the LLM is looking for them.
- Recency
- A lot of information changes over time. If that applies to your target prompt, then ensure you’re providing plenty of clues that your information is recent and relevant.
- Entity Salience
- Is your content focused on the primary entity or entities of the user query? If your text is meandering, introduces too many entities, introduces related entities, or fails to adequately position the primary entity as the focus, you could get passed up in retrieval.
Practical Priorities For Working SEOs
Traditional technical SEO — crawlability, indexability, page speed, structured data — remains a necessary foundation. What changes is the layer above it.
Audit existing content for passage quality, not just topical coverage. Pull your top-performing pages and read them as a retrieval system would. Are there clear, specific, self-contained sections that answer discrete questions? Or is the content thematically organized but passage-poor — covering the topic broadly without many extractable, citable chunks?
Edit for section-level specificity. When revising content, make individual sections more specific and more self-sufficient. Add the supporting claim. Name the mechanism. Cite the data point. This doesn’t require new content — it requires better editing.
Map content to information needs, not keyword clusters. Instead of “what keywords should this page target?” ask “what discrete questions does a person need answered to fully understand this topic?” That mapping becomes your content structure.
Treat comprehensiveness as a coverage strategy, not a volume strategy. The goal isn’t to publish more pages to target more sub-queries. The goal is to ensure your existing content addresses every major facet of a topic with a passage that would satisfy a model trying to verify that specific facet. Depth in fewer pieces beats breadth across many thin ones.
Monitor which pages are getting cited and what sections are being pulled. Where you’re being cited tells you which passages are clearing the selection threshold. Where you’re ranking but not being cited tells you where retrieval is happening but selection isn’t — which is an editing problem, not a visibility problem.
Think about non-text assets
Google’s RAG is multimodal. It “looks” at the page and its structure. It notes assets like images and video. It looks to structural elements of the page to understand the purpose and value that those assets add. Adding captions, alt text, readable and relevant file names… all the things you’d do for traditional SEO, can help solidify the value of the page and its alignment to the initial prompt.
Approaching Query Fan Out Utilization From A User-Centered Perspective
Last but far from least, consider all of this through the lens of your audience. Most specifically, I think you’ll find that the attributes and entities that align with your target audience are likely to differ from the broadest possible selection of relevant entities and attributes for your vertical (Unless you’re a massive entity yourself like IBM or AWS, Wal-Mart or Target).
Google’s LLMs are now highly personalized. They’re peeking into your docs, your gmail, your chat history (hi there Samsung users who got that system prompt to switch to Google Messenger), etc. and now they’re using that information to decide what sources to cite. They’re looking for co-occurance.
As it relates to query fan-outs, it’s now on you to whittle down your audience and identify the brands they typically engage with, the email lists they subscribe to, the subreddits they engage with, all to help you narrow down who and what you mention in your answers to your target prompts. You’ll need to know the specific attributes that drive their decision-making. THIS is where SEOs, AEO/GEOs are going to get the most traction from optimizing from LLMs. But that’s another post.
Example: Informing Content with Query Fan Out
Let’s use a practical example.
I’m going to search “What are the best running shoes for gravel roads?”. You can see in the image above that this query was most likely confined to a 512 token retrieval size.
This simple query generated three fanout queries. It altered my original query by adding time constraints, focusing on recently published content. It added the terms “reviews” and “top rated”. And right here is where a lot of folks get lost.
It seems like most SEOs and content writers are seeing this and thinking they need “top rated gravel running shoes” and “gravel running shoe review” pages. Sure, you could. But let’s say you’re running shoe brand Hoka. How much sense do those articles make on your site if you only make one running shoe appropriate for gravel? But, you could add content to your PDPs pointing out all the places where your running shoe is ranked in these lists. Which takes us to the next step: What did the LLM retrieve for these queries?
I can see that affiliate pages on running blogs were pretty commonly retrieved. I might run this same prompt a couple dozen times to see which pages were most commonly retrieved, check the traffic of those pages with Ahrefs or similar, then consider paying for prominent placement if the numbers add up (Note: I probably wouldn’t pay to be included on a page that gets little to no traffic on the off-chance it gets cited in an LLM. Remember, citations change roughly every two days).
There’s also a news article there. That’s another great opportunity and a much better place for you to focus your effort. You could create an independent test of running shoes on gravel and produce your own page with original content, then put out a press release hype it up a little. If you’re a manufacturer or reseller focused on trail running, you might turn this into an entire content marketing campaign with follow-up content, lead-in content, and a multi-channel strategy. THAT is the sort of thing that pulls LLM levers simply creating pages based on QFOs can’t touch. That generates prominence and ubiquity, which if carefully managed can lead to increased trust and more organic mentions and citations, which is what the LLM is actually chasing.
Now, let’s go back to that fan out image. Something else that was added was attributes. The fan out added “road-to-trail”. That wasn’t in my prompt. Somewhere the LLM was introduced to the term and it’s pulling it in here. Including that term in my content (or on my PDP!) could strengthen the relevance of my content to the initial query. Parsing the fan-out for terms that weren’t included in your initial prompt is a great way to find additional relevance triggers to include in your content.
More to the Process
This post focused specifically on query fan out, which is part of a much larger retrieval-augmented generation process. I’ll check back in with some more detail and insights about other parts of the process, but my hope and intent here is to give SEOs and content creators a more practical and nuanced understanding of the fan out process, how to use it and how not to use it.