Findings from 2,594 ChatGPT Fanout Queries

Fanout queries are the subqueries that an AI assistant searches (when it performs web search) based on the user’s prompt. Most major AI assistants use fanout queries, like ChatGPT, Gemini, AI Overviews, AI Mode, Claude, etc.

Being relevant for fanout queries can help make your content an eligible source candidate for AI generated answers.

I track 50 or so prompts in my prompt-tracking tool. Over the past 30 days, those prompts have generated 2,608 query rows representing 2,594 distinct fanout queries from ChatGPT. The prompts cover a variety of topics, including GEO consulting, RAG, SEO news, etc.

What I’m interested in are the details of the fanouts: like how many contain multiple entities or use quotation marks.

I’ve asked ChatGPT to analyze my fanout queries for such insights. Here’s what I found.

Limitations of this research are that it’s based on fanout queries from my tracked prompts that skew toward SEO/GEO topics, not real users, and thus the findings aren’t necessarily representative of fanouts at large but rather my unique dataset.

The most interesting story is how the fanout queries are constructed to connect entities, disambiguate concepts, select source types, and satisfy different evidence needs.

TL;DR:

Fanout queries encode retrieval strategy:

• Entities: what things should be connected
• Quotes: how precisely terminology should be matched
• site:: where information should come from
• Years: when information should come from
• Evidence modifiers: what kind of proof is needed
• Platform combinations: which entities should be compared
• Acronym + expansion: how ambiguity should be resolved

Quick Findings

Based on the 2,608 query rows:

Pattern	What I found	Why it’s interesting
Multiple entities	~1,741 queries have 2+ recognizable entities*	Fanout seems heavily relationship-driven
3+ entities	~700	Many searches are really miniature entity graphs
4+ entities	~201	Some fanouts combine platforms, people, concepts, and products
Quotation marks	224 queries, 8.6%	Exact phrases appear to be deliberately preserved
2+ quoted phrases	114	ChatGPT sometimes constrains several entities simultaneously
site: searches	194, 7.4%	Fanout includes explicit source selection
A year included	560, 21.5%	Freshness is frequently encoded directly into retrieval
Queries beginning like questions	Only 19	Fanouts overwhelmingly look like search queries, not conversational questions
Median query length	10 terms	They are relatively dense keyword/entity bundles

*Entity counts are approximate and based on a broad entity classification that includes named people, companies, platforms, locations, products, and recognizable concepts.

1. Some Fanout Queries are Basically Entity Graphs

Look at this fanout query:

generative engine optimization AI search content citations GEO Google AI Overviews Bing Copilot OpenAI ChatGPT search guidance

That isn’t one search concept. Rather, it connects:

• GEO
• AI search
• Google AI Overviews
• Bing
• Copilot
• OpenAI
• ChatGPT
• citations and content guidance

Here’s another interesting fanout:

AI search optimization experts AEO GEO Lily Ray Mike King Wil Reynolds Kevin Indig

This names SEO/GEO practitioners by name, connecting them to the concepts of AI Search, AEO, and GEO.

Another interesting fanout is:

Generative Engine Optimization AI search Google AI Overviews Bing Copilot ChatGPT Perplexity methodology citations content retrieval 2026

That’s exploring a concept, multiple platforms, a methodology, citations, retrieval, and recency simultaneously.

And so, fanout queries aren’t merely a form of query expansion. They can provide entity-relationship expansion, as well.

2. Quotation Marks are Surprisingly Common

There are 224 queries containing quotation marks, including 114 with two or more quoted phrases and 35 with three or more.

Some searches get extremely constrained, for example:

“GEO” “generative engine optimization” agency methodology “ChatGPT” “Perplexity” “Gemini” consultants

Or this one:

“New York” “GEO” “SEO consultant” “generative engine optimization” -linkedin

What’s interesting is the quotation marks don’t appear random. Rather, they’re often wrapped around specific entities, coined terminology, names, or phrases whose exact wording matters.

This pattern suggests the fanout mechanism sometimes shifts from broad semantic retrieval toward more lexically constrained retrieval, especially when exact terminology or named entities are involved. Thus, exact-string retrieval is one tool in the fanout process.

3. Fanouts Sometimes Choose the Source

There are 194 site: queries.

Examples include:

site:developers.google.com search central AI features guidance AI Overviews SEO

site:openai.com search crawler OAI-SearchBot ChatGPT search website

site:blogs.bing.com/webmaster generative search AI content SEO Bing 2025 2026

Here are the most common domains in my dataset:

Domain	site: queries
developers.google.com	47
openai.com	12
searchengineland.com	9
perplexity.ai	9
blog.google	8
ipullrank.com	6
blogs.bing.com	6
platform.openai.com	6
arxiv.org	5

In short, some fanouts contain a source-selection decision, which means retrieval isn’t always looking for the most relevant document across the web. In some cases, the fanout itself specifies where the AI assistant wants to look.

4. Freshness Is Often Explicitly Encoded Into Fanout Queries

Notably, 560 queries, or about 21.5%, contain a year.

Some fanouts even specify multiple years:

Google AI Overviews organic traffic click through rate study 2025 2026

In other words, ChatGPT isn’t always relying on the search engine to infer freshness. In many of these fanouts, a year is explicitly encoded into the retrieval request.

5. Fanout Queries Don’t Often Look Like User Prompts

Only 19 queries began with a typical question word, such as what, why, how, who, which, etc.

So instead of “How does generative engine optimization affect AI search visibility?” we may get fanouts that look like:

generative engine optimization GEO AI search visibility citations methodology 2026

Fanout queries seem to translate conversational intent into compressed retrieval language made up of entities, modifiers, source constraints, and intent terms.

In practical terms, don’t assume the literal wording of a prompt is the retrieval query you need to optimize around; rather, infer the prompt’s intent and also optimize toward that.

6. Fanout Queries Can Mix Acronyms with Their Expanded Entities

Interestingly, 431 fanout queries (16.6%) contain both “GEO” and “generative engine optimization“:

SEO GEO generative engine optimization

That redundancy may serve a purpose of increasing recall while disambiguating the acronym at the same time.

Such is the benefit of making entity relationships unambiguous.

7. Some Fanouts Triangulate Across Competing Platforms

There were several queries (~92) like:

Google AI Overviews Bing Copilot ChatGPT Perplexity

So rather than search one AI-search ecosystem at a time, some fanouts bundle multiple systems together. This appears often around questions involving methodology, citations, retrieval, optimization, and visibility.

That’s another type of entity relationship. The intent seems broader than “tell me about ChatGPT.” It’s closer to: “What is true across this category of entities?”

That could help explain why broad comparative content sometimes becomes useful for AI retrieval.

8. Fanout Queries Can Change the Kind of Evidence Searched For

The fanouts contained recurring modifiers such as:

• research
• study
• methodology
• official
• documentation
• guidelines
• citations
• experts
• consultants
• best practices
• traffic
• visibility
• structured data
• technical SEO

So one original topic can branch into different evidence requirements.

Thus, query fanout can be interpreted as more than topics. The modifiers reveal the kinds of evidence the retrieval process is seeking for a given prompt.

So when it comes to planning content from fanouts, don’t just look at relevant entities but also evidence-type relationships.

Takeaways

Taken together, the patterns above suggest fanout queries encode more than topical relevance. They also can encode:

• what to retrieve
• which entities to connect
• where to retrieve it from
• how precisely to match it
• how fresh it should be
• what type of evidence may satisfy the information need

The broader finding from this dataset is that fanout queries can encode retrieval strategy, not merely topical expansion.

Query fanout thus exposes the retrieval relationships surrounding a prompt, which can include entities, exact terminology, authoritative people, preferred domains, freshness requirements, evidence types, platforms, locations, and technical concepts.

For SEO/GEO professionals, analyzing fanout this way produces a map of retrieval relationships rather than simply a list of related keywords.

If you need help with strategies for AI search, feel free to contact me. I’m an independent SEO/GEO consultant helping brands and agencies.

Until next time, enjoy the vibes:

Thanks for reading. Happy optimizing!

Related posts