Query Fan-Out: How AI Search Breaks One Prompt Into Many Searches, and What It Means for Your Content

Query fan-out is the step where an AI search engine turns one prompt into several searches it writes itself. The engine may pick your page for a sub-query it wrote, not for the exact prompt the buyer typed. That is how a brand gets cited for prompts it never targeted, and misses citations for the ones it did.

Fan-out sits at the retrieval end of answer engine optimization: when an engine searches the web, it can cite a page only after its retrieval process finds it. Google attached the name to AI Mode in March 2025.

Other providers describe the same pattern without the label. Each lets you see a different amount of it.

Key Takeaways

• One prompt, several searches: Google says AI Mode breaks a question into subtopics and issues "a multitude of queries simultaneously." OpenAI, Anthropic, Perplexity, Grok's API docs, and Google's Gemini API each describe multi-search behavior in their own words.
• Not every recorded search shows fan-out: across the 15 highest-volume payment-gateway and payment-processor questions in Ahrefs’ Perplexity captures, the recorded query matched the question every time.
• Sub-queries name brands the buyer never typed: ChatGPT turned "What is the best payment processor to use?" into four searches. Three of them asked for one vendor's official fees.
• APIs show what most apps hide: the Gemini, OpenAI, Claude, and Grok APIs can return the queries they ran. Only Perplexity's Pro Search documents a visible breakdown in a consumer app. Treat API output as evidence about the mechanism.
• Google's sub-queries stay private: Google publishes no list for AI Mode or AI Overviews. Tracking those surfaces means recording what gets cited per prompt.

What Is Query Fan-Out?

Query fan-out is a retrieval pattern. An AI search system writes several search queries from one user prompt. It runs them at once or in sequence, then merges the results into one answer.

AI query fan-out is Google's term, but the pattern shows up across engines. Google's own version on its AI Mode tab is covered in how AI Mode picks its sources.

Where the Term Comes From

Google attached the term to AI Mode when it introduced the experiment on 5 March 2025. Google's AI Mode launch post said: "It uses a 'query fan-out' technique, issuing multiple related searches concurrently across subtopics and multiple data sources."

At I/O on 20 May 2025, Elizabeth Reid, VP, Head of Search, restated it in Google's I/O 2025 Search announcement: "Under the hood, AI Mode uses our query fan-out technique, breaking down your question into subtopics and issuing a multitude of queries simultaneously on your behalf."

The idea is older than the product name. Ma et al. made the case in a 2023 paper on query rewriting for retrieval, presented at EMNLP: "there is inevitably a gap between the input text and the needed knowledge in retrieval." Their fix: "We first prompt an LLM to generate the query, then use a web search engine to retrieve contexts."

Self-ask goes a step further. In Press et al.'s self-ask method, the model "explicitly asks itself (and answers) follow-up questions before answering the initial question." A search engine can then answer each follow-up.

Fan-Out, Rewriting, and Decomposition

Providers use these words loosely, so this guide holds them to three working definitions:

• Rewriting: one prompt becomes one different query, phrased for a search engine.
• Fan-out: one prompt becomes several queries, run in parallel or one after another.
• Decomposition: the research term for splitting a multi-part question into sub-questions, each answered before the final answer.

How AI Search Breaks Down Queries: A Worked Example

To see how AI search breaks down queries, read the searches an engine ran for a real prompt. Payments makes a clean test case. The prompt names nobody, and the engine's searches name vendors you can check.

One Prompt, Four Searches

On 26 September 2026, ChatGPT answered "What is the best payment processor to use?" in an Ahrefs capture. It ran four searches, and only one of them resembles the prompt:

1. "best payment processors Stripe Square PayPal Adyen fees online payments 2026"
2. "Stripe pricing payment processing official"
3. "Square processing fees official"
4. "PayPal merchant fees official"

The answer drew on 48 links. Twelve sat on stripe.com, twelve on squareup.com, and twelve on paypal.com, subdomains included. The other twelve were third-party comparison, review, and fee-calculator pages, with NerdWallet appearing three times.

Same Prompt, No Fan-Out

In Ahrefs’ capture, Perplexity’s recorded query matched the question exactly. That was the consistent pattern in this sample. In the 15 highest-volume payment-gateway and payment-processor questions Ahrefs captured for Perplexity, the recorded search query matched the question every time.

How this was pulled: Ahrefs Brand Radar, AI responses report, US database, using Ahrefs-generated prompts. The first pull phrase-matched “best payment processor to use” across ChatGPT and Perplexity. The second took the 15 highest-volume questions per engine matching “payment gateway” or “payment processor.” Both pulls ran on 5 October 2026. Ahrefs builds its prompts from Google People Also Ask questions and its own semantic fan-out. These are not real buyer sessions.

What the Example Does Not Prove

The vendor-pricing searches line up with the vendor-domain links. Ahrefs does not map each link to the search that found it. "Line up with" is as far as the data goes.

The captures are seven weeks apart, so they are not a same-day comparison. Treat them as one prompt shown in detail, with no rate attached.

Sub-Queries in AI Search: What Gets Searched Behind the Scenes

Each provider documents a different slice of the sub-queries AI search engines write. Some describe the behavior and show nothing. Some return every query in their API.

Others say nothing at all. The table reflects each provider's own pages as read on 5 October 2026.

What Each Engine Documents

Where the Documentation Stops

Three surfaces had no documentation on how they form queries: the Gemini app, the Grok app, and Perplexity's default search. No provider says an API's queries match its app either. Google's AI Overviews get one hedged "may use."

The rest of that surface's behavior belongs with Google's AI Overviews as a topic of its own.

The Perplexity rows look contradictory and are not. Pro Search documents multiple searches. The default answers Ahrefs captured ran the question verbatim, matching the two captures in how Perplexity picks its sources.

API Behavior Is Not App Behavior

API documentation describes a developer tool. It tells you what the model can do when a developer calls it. What ChatGPT, Claude, or the Gemini app does for a logged-in buyer is a separate question.

OpenAI's help pages describe a separate set of app behaviors, covered in ChatGPT's surfaces as OpenAI documents them. The same split applies to what grounding in Google Search means for Gemini. Read API-documented behavior as evidence about the mechanism; whether an app behaves the same way is undocumented.

Query Fan-Out SEO: Why Ranking for One Keyword Isn't Enough

Query fan-out SEO starts from one uncomfortable fact: the engine matches your page to a query it wrote, and you never see that query. A head-term ranking may cover only one branch of the fan-out. The other branches go to whichever pages answer them.

The Prompt You Track Is Not the Query That Runs

The payments prompt named no vendor, yet three of ChatGPT's four searches named one each. A page ranking first for "best payment processor" could compete in one branch: the broad comparison search. The other three asked for one company's official fees, which only that company's pricing page answers well.

Why You Get Cited for Prompts You Never Targeted

The reverse also holds. Stripe's pricing page appeared in a "best processor" answer. The engine had written a Stripe pricing search on the buyer's behalf.

Research systems show the same effect. Rackauckas's RAG-Fusion study credits its better answers to "the generated queries contextualizing the original query from various perspectives." It is a single-author case study at Infineon, so use it for the mechanism and nothing more.

Where Coverage Stops Paying

The same study found that "some answers strayed off topic when the generated queries' relevance to the original query is insufficient." Not every sub-query deserves a page, and chasing every branch spreads a site thin. Build topical coverage around the branches that recur across prompts.

How to Map the Fan-Out for Your Category

You cannot see most sub-queries. You can still collect enough of them to see the pattern.

The procedure below is this guide's own method, built from the surfaces in the table above. The example category is payments for online marketplaces, and every sample prompt and sub-query is hypothetical.

Start With Five Buyer Prompts

1. Write 5 to 10 decision-stage prompts in the words a buyer would use. For example: "Which payment processor handles split payouts for a marketplace?" or "Is Stripe Connect or Adyen better for a two-sided marketplace?"
2. Keep them unbranded where buyers would. An unbranded prompt shows you which vendors the engine adds on its own.

Collect the Queries You Can See

1. Run each prompt through the APIs that return queries: the Gemini Interactions API (google_search_call), OpenAI's web search tool (web_search_call), Claude's web search tool (server_tool_use), and the Grok API's web_search tool calls. An API run exposes the mechanism. It does not reproduce the consumer app.
2. Each search also carries a price for API users. Google bills Gemini 3 API projects “for each search query that the model decides to execute,” according to its Gemini API grounding documentation. That makes fan-out depth a cost variable in API implementations. It does not tell us how many branches consumer apps choose or why.
3. Read Ahrefs' recorded queries if you have Brand Radar. Ahrefs' API documentation states that the search-query field stays empty for every engine except ChatGPT and Perplexity.
4. Run the prompts in Perplexity Pro Search and copy the steps it shows.

Group Sub-Queries by Intent

1. Sort every sub-query into an intent bucket, then decide which page or source should win each bucket.

Where Fan-Out Queries Pull Sources From: Reviews, Reddit, Comparisons, Docs

The shape of a sub-query predicts the kind of page it retrieves. The payments captures are the only measured case here. Everything else in this section is inference from them, and it is labeled that way.

Vendor Sub-Queries Pull Vendor Docs

ChatGPT's three "official" fees searches line up with 36 of its 48 links landing on the three vendors' own domains. All but one of those links were deep pages: pricing pages, a Square developer docs page, and Stripe's support topic on pricing. Fan-out queries that name a vendor and ask for "official" information reward whoever has published the answer in plain text.

Comparison Sub-Queries Pull Roundups and Reviews

Perplexity's single, unchanged comparison question retrieved ten third-party pages, every one a roundup or buyer's guide. Review platforms enter through the same door: G2's editorial roundup on learn.g2.com was one of the ten. ChatGPT's broad first search lines up with its twelve third-party links, NerdWallet's three among them.

Where Reddit Enters

No reddit.com page appeared in either payments capture. That is one prompt in one category, so it says nothing about Reddit's share elsewhere. A hypothesis worth testing: experience-shaped sub-queries ("is it legit," "complaints," "worth it") are where forum threads compete.

Understanding how a thread becomes citable tells you whether your category's threads could win those branches.

Content and Off-Site Moves That Win More Sub-Queries

Each move below ties to a bucket from the map. On-site moves cover the branches your own pages can answer. Off-site moves cover the branches only other people's pages can.

Answer the Sub-Query on the Page That Should Own It

Put each answer on the page a search would expect to find it. A fees question belongs on the pricing page, and a how-it-works question belongs in a guide whose heading states it. When the answer lives in the wrong place, the branch goes to a third party that put it in the right one.

Make Pricing and Specs Retrievable

The "official" fees searches line up with pricing pages that state their numbers in text. The same answer also cited three fee calculators and rate comparisons: alltools.dev, spark.money, and myPayAdvisor. The inference: gated pricing hands the branch to someone else's estimate.

Get Named in the Pages Comparison Queries Retrieve

Comparison branches retrieve roundups, so the work is off-site. Get named accurately in the buyer's guides, review sites, and community threads your comparison searches return. Start with the domains already in your captures, because those are the sources the engine retrieved and cited last time.

Tracking Which Sub-Queries You Win

No tool sees an engine's consumer-app sub-queries unless the engine shows them. Tracking fan-out means separating what you observe from what you infer. Write down which is which.

What You Can Observe

• API queries for your own runs: the Gemini, OpenAI, Claude, and Grok APIs return the searches they ran for your prompt. OpenAI says its output includes them "usually (but not always)."
• Perplexity Pro Search steps: visible in the interface, per Perplexity's help center.
• Ahrefs' recorded queries: available for its own ChatGPT and Perplexity captures. The field can be empty. All 15 of the highest-volume ChatGPT payment-processor captures held no query. An empty field cannot be read as "no search ran."

What Nobody Outside the Engine Can See

Google publishes no list of AI Mode or AI Overviews sub-queries. The Gemini and Grok apps document nothing about query behavior. No outside monitoring tool sees every search run inside another buyer’s private session.

A Tracking Record That Holds Up

Log six fields per run: the prompt, the engine, the date, the cited domains, any visible queries, and where those queries came from (API, Pro Search, or Ahrefs). That last field keeps an API approximation from being reported as app behavior. Run the same prompt set on a schedule, and you are building a citation baseline that shows which branches you win over time.

What Klarivo Tracks When the Sub-Queries Stay Hidden

Most of a category's sub-queries never surface, so measurement has to start from what does. Klarivo Monitor works from the visible end. It runs targeted queries across five engines, ChatGPT, Claude, Perplexity, Gemini, and Grok, each tracked independently.

Its AEO Query Generation feature exists to "auto-generate high-intent queries that buyers ask AI." Monitor reports query-level mention rates and top citation sources, the domains the engines drew on, updated on a schedule you control. It does not see the engines' sub-queries, and nothing here claims it does.

Cited domains are the visible trace left by retrieval, even when the engine keeps its sub-queries hidden. When a pricing branch keeps landing on a competitor's fee page, or a comparison branch keeps landing on a roundup that omits you, the domain list shows it. Klarivo's managed AEO and Reddit service then works on those sources: the pages and conversations the engines already trust for your category.

To see which domains the engines draw on for your category’s buyer prompts, start with Klarivo Monitor. To talk through which branches matter in your category, book a Klarivo discovery call.