This text will be a translation, analysis, and semantic commentary on the phenomenal collective work organized by Andrea Volpini, " Query Fan-Out: A Data-Driven Approach to AI Search Visibility ".

The work that generated this article and an incredible tool was built through the analysis of many professionals and brought us a concept that I believe will be very important for the future of SEO.

To help you understand what is my commentary and what is translated text from the original content, I will always comment in a pre-formatted block, like this one here.

According to the original post:

“Google’s AI Mode doesn’t just process your query — it breaks it down into a network of sub-queries, spreading across its Knowledge Graphs and web index to generate rich, synthesized answers.”

They launched a tool that helps predict the next likely follow-up query for your published content — so you can optimize for how AI actually thinks.

I've been testing this tool in my daily work and I highly recommend you do the same. It runs on Google Colab, and with a little knowledge and attention, you can use it in your projects.

How does Google search work in AI mode?

Andrea explains to us that:

“AI-powered search doesn’t just process your query—it expands it into multiple subqueries, spreading across Google’s knowledge graphs and its web index to synthesize comprehensive answers. This process, called “query fan-out ,” represents the most significant shift in search behavior since mobile-first indexing, but it feels like we’re all flying blind, unable to gauge how well our content performs in this new reality.”

Here I want to invite you, the reader of this text, to pay close attention. The shift from a search mediated by an algorithm that assembles a list of 10 blue links to what we called at the time "position 0" was already radical, but nothing compares to what we are seeing with AI Overview. 

According to what Andrea tells us, I interpret Gemini as simulating our search behavior, and being much more selective when preparing to answer our question. And that's completely understandable. It's one thing to give us a list of 10 posts and pass the responsibility on to us, quite another to answer the question, being the sole source of the answer.

The article continues:

“After reviewing fundamental patents such as US 2024/0289407 A1 (Stateful Chat Search) and new insights from WO2024064249A1 (Systems and methods for generating prompt-based queries for diversified retrieval), the latter highlighted by Michael King’s analysis, and studying AI search patterns for several months, I’ve updated our practical framework and the accompanying Colab tool. The breakthrough isn’t just understanding query fan-out , it’s being able to more accurately simulate its initial stages and score your content against it.”

From here I'll leave you with the translated text and will add my comments at the end of the article. Happy reading!

The Reality of Query Fan-Out: Beyond Deterministic Classifications

When someone asks Google's AI, "What is the best sustainable marketing strategy for small e-commerce businesses?", the AI ​​doesn't search for that exact phrase. Instead, it breaks down the query into multiple subqueries that vary based on the user's context:

• What makes a marketing strategy sustainable?
• Which marketing channels work best for e-commerce?
• What are the budget constraints for small businesses?
• How do sustainable practices impact customer acquisition?
• What are some successful case studies in sustainable e-commerce marketing?

But here’s the critical insight from my work since the SGE (Search Generative Experience) days: these follow-up questions aren’t the same for everyone. They’re deeply contextual, stochastic, and impossible to predict deterministically. The problem for marketers isn’t just optimizing for subqueries—it’s accepting that there’s no way to “rank” or even track visibility in traditional terms. The game has fundamentally changed.

A New Framework: Predicting Questions

Traditional SEO tools struggle with the dynamic nature of query fan-out . What we need is a way to probe how our content might perform when AI deconstructs user intent. Our updated Colab simulator now takes a more AI-native approach to this challenge:

• (Future Vision) Validation Function and Reward Policies: While the current tool simulates and scores, the long-term vision remains the development of robust validation functions—reward policies within DSPy to substantiate and improve predictions of real contextual follow-ups based on user interactions and knowledge graphs.
• Understanding Entities and Context with AI: Instead of traditional scraping, the tool leverages Google's Gemini model and the url_context tool, which simulate how AI Mode might interpret URLs. Gemini identifies the main ontological entity and extracts the key snippets of well-reasoned content that it considers relevant on the page. This more closely reflects how AI search systems digest content: not by analyzing the HTML from top to bottom, but by focusing on semantically rich and quotable snippets.
• Synthetic Query Generation with Chain of Thought: Informed by the entity identified by Gemini and guided by patent principles such as WO2024064249A1, the simulator uses a Chain of Thought (CoT) module from DSPy with Gemini. This CoT module first reasons about the different types of information facets and query types (Related, Implicit, Comparative, etc.) relevant to your entity and then generates a diverse set of fan-out . This is a step beyond simple keyword expansion, aiming for a more grounded decomposition.
• Semantic Coverage Assessment: Using content snippets extracted by Gemini (or a summary generated by Gemini if ​​direct snippets are not available), we assess how well your page's key information covers these fan-out using semantic similarity (embeddings). The goal is not to predict all possible queries—it's to build a semantic infrastructure that increases our accuracy in predicting the next likely question based on the user's context.

Gianluca Fiorelli has been working on this for years, recommending that we retrieve all the query refinements (Topic filter, PAA, People also search, image search tag queries, etc.) that Google presents for a query.

How Google Chunking Works — and Why It Matters

Google doesn't analyze content as a whole. Instead, it segments documents into smaller, meaningful chunks . Based on our research, Google appears to use a hybrid strategy that includes:

• Fixed-size chunking: used to adjust the content to the template limits, such as the 2048 token limit for gemini-embedding-001.
• Recursive chunking: to divide unstructured content based on paragraphs → sentences → words.
• Semantic chunking: where related phrases are grouped together.
• Layout-aware chunking: This segments content based on HTML structure (headings, lists, tables) and is the default in Vertex AI Search via LayoutBasedChunkingConfig.

Among these, layout-aware chunking is the most crucial for ingestion. It aligns with how documents are visually structured, how humans process them, and how Google's AI determines relevance at the pass-through level.

Now, before we share how we plan to look for follow-up questions, let's review the core principles behind the framework.

The Strategic Framework for AI Search Optimization: 10 Core Principles

Based on successful AI-visible content testing and analysis, these are the strategic principles that truly make a difference:

1. Focus on the Ontological Core, Not the Volume of Content: The shift isn't about producing more content—it's about building a robust semantic foundation. Your ontological core (as introduced by Tony Seal) is the structured representation of essential domain knowledge. It's what allows AI systems to generate dynamic, context-aware responses.
2. Build a Dynamic and Conversational Content Architecture: Because follow-up questions are contextual and stochastic, their content should be conversational and adaptable—not merely static or exhaustive.
3. Prioritize EEAT and Structured Data Rigorously: Google has made it clear: EEAT (Expertise, Authority, and Trust) is fundamental to AI-generated answers. But, in addition to traditional trust signals, you also need to provide explicit semantic context through schema markup.
4. Adapt to Conversational and Intent-Driven Queries: AI excels at understanding natural language — your content strategy should reflect this, focusing on underlying intent, not just keyword phrases.
5. Develop Predictive Models for Contextual Follow-ups: Universal coverage is not the goal. Instead of optimizing for individual queries, focus on building data—and eventually models—that can predict likely follow-up questions based on the user's context.
6. Stand Out in “Zero-Click” and “Citation-Based” Visibility: Success metrics are changing. Being cited in AI-generated responses can be more impactful than traditional clicks—especially for brand authority and consideration.
7. For e-commerce, focus on rich product data and comparative content: AI-driven search doesn't just answer straightforward queries—it supports users through complex decision-making journeys.
8. Monitor Paid Search Performance and Adapt Strategies: AI Overviews are reshaping the search landscape — including paid visibility.
9. Stay True to Your Values ​​and Listen Strategically: As AI reshapes how we create, optimize, and discover content, staying true to your core values ​​is more important than ever.
10. Experiment and Iterate Continuously: The AI ​​search landscape is rapidly evolving. Use tools like our query fan-out to regularly assess the visibility of your content by AI and adapt strategies based on real-world reach data.

The Simulator: Understanding Your Content's Readiness for AI Mode

Understanding query fan-out is one thing — simulating how your content might be perceived and deconstructed by AI is the first practical step. Our updated Colab notebook aims to provide that initial insight.

This new release enhances the simulation by:

• Leveraging Gemini for Initial Understanding: Uses Google's Gemini (through its url_context-like reasoning or direct analysis capabilities) to identify the main entity in your URL and extract important content snippets from the page.
• Sophisticated Query Fan-Out Simulation with DSPy CoT: Employs a DSPy Thought Chain (CoT) module powered by Gemini to generate fan-out .
• Semantic Coverage Score: Evaluates how well the content snippets identified by Gemini in your URL cover these various synthetic queries using embedding-based similarity.

Technical Pipeline: URL → Entity Extraction → Query Fan-Out → Embedding Coverage → AI Visibility Score

While predicting every exact contextual match is impossible, this tool helps you understand the readiness of your content for an AI-driven search environment that relies on this decomposition.

From Coverage to Context: The Real Challenge

The notebook reveals something deeper than coverage gaps — it shows the impossibility of deterministic optimization in an AI-driven search world. Most of the content only addresses samples of possible follow-ups, and that's not a bug, it's a characteristic.

Ready to start building your ontological foundation?

The simulator is your first step toward understanding how contextual tracking works — and why traditional SEO thinking no longer applies.

Access the Query Fan-Out Simulator (version 3) — a free tool to measure the AI ​​search visibility of your content [runs on Google Colab].

My comments on this incredible work:

The first thing I want to contribute is the notion of facets. I've talked about Shiyali Ramamrita Ranganathan here on the blog before. He's the person who defined the five laws of Library Science and inspired me to create the five laws of SEO.

If you don't know him, he was a mathematician and librarian in India and is considered the father of library science in the country and one of the most influential in the world. Ranganathan wrote fundamental articles and books on the organization and classification of information and greatly developed the concept of facets.

According to him, facets are components of a composite subject, essential for the analytical-synthetic organization of knowledge. They are manifestations of the five fundamental categories (personality, matter, energy, space, and time) and can be used to form classes, terms, and numbers.

Facets allow a complex subject to be divided into its constituent parts, facilitating the analysis and synthesis of knowledge. 

I found it very curious that the number of facets Andrea mentions in the text is exactly the same number that Gemini used, as defined by Ranganathan.

But what's important here is the idea behind Gemini using the concept of facets when generating an answer for us.

Earlier in the text I mention that AI Overview attempts to simulate our way of searching for information, and it's even more careful because it now only needs to generate one answer.

If you look at how Gemini generates search results, you'll see it creates a response with multiple perspectives, citing multiple sources.

In the search above, Gemini's AI-generated overview uses various types of information sources to generate a fairly complete answer, although it may contain some incorrect information due to errors in one of the sources.

But what matters to us here is the fact that Gemini makes this panoramic movement around the subject of our question, trying to look at the issue from various angles, trying to understand all facets, asking many questions that are semantically related.

This allows me to state: when creating any content on the web, we need to get as close as possible to the completeness of the subject. We must address as many subtopics of the main subject as possible.

This is nothing new for those already working on projects using Semantic SEO. We've already adopted this strategy and are seeing good results in search, both in the algorithm and in AI-generated summaries.

Fan-out logic and ontologies.

I did a search in the knowledge base related to the study of LLMs, Knowledge Graphs and Search Engines and found an interesting relationship, somewhat technical but interesting nonetheless.

The idea of ​​"expansion" or "branching out" of a query to a bank, system, or database, which the term " fan-out " can evoke in the domain of information search and retrieval, is a central and recurring theme in various techniques that combine LLMs, KGs, and Search Engines, but I had no idea about it.

Let's explore this in detail, starting with the direct mention I found in the book Semantic Web Technologies – Trends and Research in Ontology-based Systems (2006) , and then expanding on the broader implications that " fan-out " in queries (or information generation) has for your fields of interest.

The concept of "Fan Out" in ontologies.

In the more than 30 sources (including books, technical and scientific articles) in my knowledge base, I found that the term "fan out" has been used to describe the behavior of a selection function in reasoning systems with inconsistent ontologies .

Now we're going to get to the trickiest part, hold on tight, we're going to do this together.

1. In this context, the selection function s starts with a query formula ( f ) as a starting point.
2. She then “selects the formulas c that are directly relevant to f as a working set”.
3. If a satisfactory answer is not found, the function “increases the relevance level by 1, thus adding more formulas that are relevant to the current working set”.

The formula is then this:

s(Δ, f, k) = {c ∈ Δ | c is directly relevant to s(Δ, f, k – 1)} , for k > 1

I don't know if you noticed, but this approach leads to a "fan-out" behavior (which can also be understood as branching or expansion) of the selection function:

• The first selection is the set of all formulas directly relevant to the query;
• Next, all formulas that are directly relevant to that set are selected, and so on.

In this context, if the initial query is not satisfied on the first try, it generates a chain reaction of new queries related to each other.

This is a way to progressively extend the set of information considered to resolve a query, especially in scenarios where the information may be distributed or ambiguous, aiming to find a consistent subtheory from an inconsistent ontology to find meaningful answers.

In Andrea's work, that's exactly what I learned: the process is triggered in this way because of the nature of the expected response.

Although this is a specific technical application, the process of "fan out" in ontologies—that is, the idea of ​​expanding the search to include related information that was not explicitly requested in the initial query but is contextually relevant—is found in many modern approaches to information generation.

Broader implications of "Fan Out" in current Information Retrieval

In the context of LLMs, Knowledge Graphs, and Search Engines, the "fan out" of a query can be interpreted as the expansion of a system's ability to retrieve and link information to answer complex questions, improve the relevance of results, or handle ambiguity .

Let's see how this works?

Query Expansion and Rewriting

One of the most direct uses of LLMs working in conjunction with Search Engines is the ability to rewrite and expand queries to improve information retrieval.

Models can analyze user intent and generate additional query terms or reformulations that cover a broader spectrum of relevant information, bridging the "vocabulary gap" between what the user types and the terms contained in the documents.

And then we see a number of interesting applications of this, as I describe what I discovered below:

• For SEO: understanding how LLMs expand or rewrite queries is an interesting increment to our understanding. Instead of focusing solely on exact terms, SEO strategies can consider semantic relevance and the coverage of broader topics that a model using this strategy can infer and expand from an initial query. This means optimizing content for a set of semantically related terms and their variations. Semantic SEO, right?
• In Information Systems: query rewriting increases the accuracy and comprehensiveness of results, especially for ambiguous or complex queries. LLMs can generate clarifying questions for the user (e.g., "Did you mean jaguar as an animal or as a car?") based on the initial results, and then use the user's response to refine the query, essentially "branching" the interaction for greater clarity.
• For AI agents: AI agents can use models to reframe their own "internal questions" when searching for information in a database or on the web, allowing for more effective and adaptive exploration.

Multi-Hop Question Answering

Complex questions often require combining information from multiple sources or performing multiple steps of reasoning, which is a classic example of informational "fan out".

This is known as "multi-hop question answering" .

Instead of a single search, the system needs to "branch out" its search strategy by looking for interconnected facts. But to function correctly, some things need to be taken care of.

• Challenges: Simple vector similarity searches can fail in multi-hop queries because the necessary information may be scattered across multiple documents, or the most relevant documents may contain repeated information, ignoring other crucial facts.
• Mixed Solutions: Graphs and LLMs: Knowledge Graphs are particularly effective for this, as they allow for the modeling of relationships between entities. A multi-hop query can be decomposed into sub-questions, and the KG can be traversed to connect the necessary information. LLMs can assist in decomposing the question and formulating sub-queries for the KG. The ability of LLMs to use "chains of thought" allows them to separate questions into multiple steps, define a plan, and utilize external tools (such as KGs or search APIs) to generate a complete answer.

Retrieval-Augmented Generation (RAG)

The RAG paradigm is a basic example of how models intentionally "fan out" external knowledge sources. Instead of relying solely on the parametric knowledge "frozen" in their training, LLMs "retrieve" relevant information from an external corpus (either through a search engine or by scanning a Knowledge Graph) and use it as context to generate responses.

This process offers numerous advantages, including:

• Reducing hallucinations: This approach mitigates the problem of "hallucination" (generating false or unverifiable information) by grounding the response in retrieved facts. The ability to "fan out" on a broader search for relevant documents allows the model to provide more factual and reliable answers.
• Data and diversity: the quality and diversity of retrieved data are crucial factors for good RAG performance. Considering that the system can "branch" the retrieval and include documents with "noise" or similar information, we have a huge challenge for LLMs in RAG. Modeling long documents and accurately understanding concepts are areas of continuous improvement to deal with this "fan out" of information.

Traversal and subgraph extraction in Knowledge Graphs

In a Generating Framework (KG), a query can "branch out" or "fan out" along the connections (edges) to retrieve not only direct entities, but also their relationships and neighbors. This is what makes Graphs such powerful tools for retrieving contextual information and answering more complex queries.

Let's detail some applications again:

• Detailed retrieval: Instead of a keyword search that only finds qualified nodes, a query on a KG can extract entire subgraphs (Query Sub-graphs) that include not only the node of interest, but also its triples and the triples of its neighbors to a certain depth. This allows obtaining information associated with multiple nodes, eliminating redundancies and increasing precision.
• Input optimization for LLMs: frameworks like Auto-KGQA select "smaller fragments of the graph" to serve as context for the model, which is a form of controlled "fan out" to reduce the number of input tokens for the LLM while maintaining performance.
• AI Agents and Graphs: The ability of models to convert natural language into graph query languages ​​(such as SPARQL or Cypher) allows AI agents to use "fan out" in their graph queries, leveraging the structure to infer and aggregate data, which is difficult with unstructured text.

Ontology refinement and search scope

The concept of " fan out " can also be applied to how we navigate and refine ontologies to adjust the scope of a search. Ontologies provide a hierarchical structure (e.g., subClassOf ) that can be used to "expand" ( up-posting ) or "reduce" ( down-posting ) the scope of a query.

Let's look at some more examples:

• Up-posting and Down-posting: If a query for “Oracle 8” does not produce results, the system can automatically “expand” ( up-post ) the query to a more general term, such as “database,” to find more information. Conversely, the query can be “reduced” ( down-posted ) to more specific terms, such as subconcepts of “Object-Oriented Programming Language,” for “Java.” This allows systems to “ fan out ” or “ fan in ” the semantic query along the knowledge hierarchy.
• Usability optimization: This improves usability by allowing users to find appropriate terms in the ontology, even if they are not initially familiar with them.

Practical applications

One of the questions I heard the other day about this topic is: can we apply these concepts to content production? And I wonder about that in SEO projects.

So, let's look at some ideas on how to apply fan-out to our projects?

• SEO projects: instead of just optimizing for exact keywords, understand the semantic "networks" that LLMs create when expanding queries. Build your content in a way that covers a broader spectrum of related terms and concepts, anticipating the " fan out " of user queries. In my book Semantic SEO: Semantic Workflow , I already talk a lot about this.
• Information Systems: Design systems that can decompose complex queries and perform multi-hop searches in Knowledge Graphs. Integrating LLMs for query rewriting and curating graph fragments (as in Auto-KGQA ) will optimize the performance and relevance of responses, especially in domain-specific knowledge bases, such as those found in healthcare or finance. The ability to " fan out " for relevant information (whether in text or KG) and summarize results in a navigable way is a key differentiator.
• Creating Artificial Intelligence Agents: Your AI agents can be designed to utilize multiple LLMs to plan and execute step-by-step searches, "branching out" to external tools (such as search engines to update information or graphs for facts augmented by structured data) when internal knowledge is insufficient or could lead to hallucinations. The ability to " fan out " to gather evidence from multiple sources and integrate it coherently is a fundamental skill for autonomous agents.

In short, " Query Fan-Out " encompasses the idea of ​​strategically expanding the search, retrieval, and generation of information . Whether through rewriting queries by LLMs, exploring relationships in Knowledge Graphs, or orchestrating multi-hop searches, the ability to "branch out" to obtain richer and more contextual information is fundamental to building truly intelligent and reliable information systems and AI agents.

Therefore, I invite you to read Andrea's original article and test his tool. Follow the people mentioned here on LinkedIn and let's talk about this subject, which seems complex, but which has actually given us more certainty that we are on the right track.

Let's keep in touch!

References:

For more information on the evolution of Google's AI Mode and its implications for search and SEO, please see:

• Gianluca Fiorelli's “ Why AI Mode Will Replace Traditional Search as Google's Default Interface” on I Love SEO
• Aleyda Solís ’s breakdown of the Query Fan-Out approach
• Dennis Goedegebuure ’s perspective on SEOs as natural-born marketers in the age of AI
• Michael King ’s in-depth analysis of the AI ​​Mode patent
• The official patent document for “Search with stateful chat” (US20240289407A1)
• The WO2024064249A1 patent “Systems and methods for prompt-based query generation for diverse retrieval”
• DAVIES, John; STUDER, Rudi; WARREN, Paul (Org.). Semantic Web Technologies : Trends and Research in Ontology-based Systems. [S. l.]: Wiley, 2006.

Content translated from the original article written by Andrea Volpini on May 26, 2025.