Query Fan-Out: How One Buyer Question Becomes 12 Hidden Searches

Query fan-out is the step where an AI search system takes a single prompt, decomposes it into multiple related sub-queries, retrieves sources for each one independently, and then assembles a single answer from the results. Google documents the behaviour explicitly for its AI features. The practical consequence is that you can hold the top result for the question a buyer typed and appear in none of the retrievals that actually produced the answer they read.

TL;DR

• One prompt becomes many retrievals. You compete separately in each, and you only see the one you targeted.
• A realistic B2B security prompt fans out into eight to fifteen sub-queries covering compliance, pricing, integration, deployment, and alternatives.
• Semrush found in July 2025 that pages cited by ChatGPT ranked at position 21 or worse in conventional Google results almost 90% of the time, which is what partial fan-out coverage looks like from the outside.
• The failure mode is invisible in every tool you own. Rank tracking says you are winning. The answer does not mention you.
• The fix is coverage architecture, not a longer page.

What fan-out does mechanically

A conventional search engine matches a query against an index and returns ranked documents. A generative engine has a harder job, because the prompt is often a compound question that no single document answers.

Take "best CNAPP for a 200-person fintech". No page in existence answers exactly that. There are pages about CNAPP vendors, pages about security tooling at company size, pages about fintech compliance requirements, and pages comparing specific products. The system's route to a usable answer is to break the prompt apart, retrieve for each piece, and synthesise.

The decomposition is not visible to you. It happens between the prompt and the answer, it varies by engine, and it varies by phrasing of the same question. What you see is the output, which reads like one answer to one question and is nothing of the kind.

A worked example

Here is a plausible fan-out for that prompt, with the kind of content that wins each sub-query.

Sub-query	What wins it	Who typically holds it
What is a CNAPP	A clean one-sentence definition	Analyst sites, glossaries
CNAPP vendor list 2026	A maintained comparison table	Review sites
CNAPP pricing models	Published pricing with units	Almost nobody in security
Security tooling for a 200-person company	Sizing guidance by headcount	Blogs, community posts
Fintech compliance requirements	PCI DSS and SOC 2 specifics	Compliance vendors
CNAPP vs CSPM vs CWPP	A category distinctions table	Vendor blogs
CNAPP deployment effort	Time-to-value with real numbers	Almost nobody
Cloud coverage by vendor	An AWS, Azure, GCP support matrix	Documentation
Alternatives to the market leader	An honest comparison page	Competitors
CNAPP false positive rates	Operational reality, rarely published	Practitioner forums

Now count. A CNAPP vendor with a strong product page and a strong "what is CNAPP" post is competing in two of ten. The other eight are answered by review sites, compliance vendors, competitors, and forums. The synthesised answer draws from all ten, so the vendor appears in a fifth of the underlying material and is described mostly in other people's words.

Two rows in that table are worth staring at. Pricing and deployment effort are sub-queries almost no security vendor answers publicly, and both are decision-stage questions. Whoever does answer them gets retrieved for them, and in this category that is usually a review site working from secondhand information.

Why your dashboards will not show you this

Rank tracking reports the query you configured. Fan-out generates queries nobody configured, including phrasings that never appear in keyword tools because no human types them. They are machine-generated intermediate steps.

Search Console shows queries that produced impressions in Google Search. Sub-queries inside an AI Mode session are not that.

Even a citation-tracking tool running a prompt panel usually records whether you appeared in the final answer, not which of the retrievals you won. So the observable signal is binary and the underlying cause is a coverage map you cannot see directly.

This is why teams report the same confusing pattern: strong rankings, healthy traffic, and a persistent absence from AI answers in their own category. The rankings are real. They are just concentrated in one or two of the sub-queries that matter.

How to reverse-engineer the fan-out

You cannot see the decomposition directly, but you can approximate it well enough to act.

1. Ask the engine to enumerate. Prompt with "What sub-questions would you need to answer before recommending a CNAPP for a 200-person fintech?" The list you get back is not the actual decomposition, and it is close enough in practice.
2. Read the answer's own structure. A generated answer that covers pricing, compliance, and deployment in separate paragraphs is showing you its retrieval buckets.
3. Follow the citations, per section. Note which domain is cited for which part of the answer. That mapping is your competitive picture, and it is usually a different set of competitors per sub-query.
4. Vary the phrasing. Run the same intent five ways. Sub-queries that recur across all five are the stable core worth building for.
5. Check the negative space. Sub-queries where the engine cites nothing good, or hedges, are open. In security that is reliably pricing, real deployment timelines, and honest limitations.

An hour on one high-value prompt produces a coverage map you can actually assign work against.

Coverage architecture

The instinct after seeing a fan-out map is to write one enormous page covering all ten sub-queries. That is the wrong response, for a mechanical reason: retrieval operates on passages, so a 6,000-word guide competes as thirty separate chunks, and its pricing section competes against a dedicated pricing page that is entirely about pricing.

The better architecture:

Sub-query type	Format	Why
Definitional	A section on a hub page	Short, stable, rarely the deciding factor
Comparison	A dedicated page with a table	Table structure survives extraction cleanly
Pricing	A dedicated page with real numbers	Almost no competition in security
Deployment and effort	Documentation, not marketing	Docs get retrieved for operational questions
Compliance	A trust page with named frameworks and dates	Specific, verifiable, high intent
Alternatives	An honest comparison page	Otherwise a competitor writes it

Each page targets one sub-query completely rather than ten partially. They interlink, so a system retrieving one has a path to the others, and the hub page ties the cluster to the head term.

This is topic-cluster architecture, which SEO has recommended for a decade. Fan-out changes the reason it works. Under SEO the argument was topical authority. Under retrieval the argument is that you need a distinct, strong passage for each retrieval that a compound prompt generates.

The audit

For your three highest-value buyer prompts:

1. Run the prompt on Perplexity, ChatGPT, and Google AI Mode. Save the answers.
2. Break each answer into its component claims and note which domain is cited for each.
3. Build a grid: sub-queries down the side, engines across the top, and mark where you appear.
4. For every empty cell, decide whether you have a page targeting that sub-query at all. Usually you do not.
5. Rank the gaps by decision-stage weight. Pricing and alternatives outrank definitions every time.
6. Ship one page per gap, in that order.

Most vendors find between four and seven sub-queries where they have no relevant page. That backlog is more valuable than any keyword research output, because every item on it is a question a real buyer's prompt already generated.

Structure each of those pages so its sections survive extraction, which is a discipline of its own, and check that the engines can reach them at all using the crawler reference. For how the four engines differ in how they fan out and cite, see the engine mechanics reference.

Frequently Asked Questions

What is query fan-out?

Query fan-out is when an AI search system decomposes a single prompt into multiple related sub-queries, retrieves sources for each independently, then synthesises one answer from the combined results. Google documents the behaviour for its AI features. It means a compound question is answered from many retrievals rather than one ranked list.

How many sub-queries does a prompt generate?

It varies by engine, prompt complexity, and phrasing, and the exact count is not exposed. A realistic B2B buying prompt tends to produce roughly eight to fifteen distinct retrievals in practice, covering definition, comparison, pricing, deployment, compliance, and alternatives. Treat any precise published number with suspicion, since the decomposition is not observable from outside.

Can I see the actual sub-queries?

Not directly. You can approximate them well by asking the engine what it would need to know before answering, by reading the structural sections of the answer it produces, and by mapping which domain is cited for which part. Running the same intent in five phrasings and keeping the sub-questions that recur gives you a stable working list.

Why do I rank first but never get cited?

Most likely you hold one sub-query and miss the rest. The answer a buyer reads is assembled from many retrievals, and holding the head term contributes one input among several. Semrush found in July 2025 that pages cited by ChatGPT ranked position 21 or worse in conventional results almost 90% of the time, which shows how loosely rank and citation are coupled.

Should I write one long page or many short ones?

Many focused pages. Retrieval competes at passage level, so the pricing section inside a 6,000-word guide competes against a page that is entirely about pricing, and usually loses. Build a cluster of pages that each answer one sub-query completely, interlinked, with a hub page tying them to the head term.

Which sub-queries do security vendors most often miss?

Pricing, deployment effort with real timelines, and honest alternatives. All three are decision-stage questions, all three are routinely answered by review sites and competitors instead of by the vendor, and all three are where an accurate first-party answer displaces a secondhand one immediately.

Does query fan-out apply to Google Search too?

To Google's AI surfaces, yes. Google describes AI Overviews and AI Mode as using a fan-out approach. Conventional blue-link results still work on direct query matching, which is exactly why a page can rank well in one and be absent from the other.