Query Fan-Out Rewards One Page That Answers Many Questions

TL;DR: Ask Google’s AI Mode one question and it searches several related ones. They call this query fan-out. Many agencies and tool vendors answered by scaling content with a page per variation, but Google’s own documentation says that method isn’t effective. Instead, build one page per distinct question and answer its phrasings inside it.

One of your website’s pages holds position three in Google for the search term it was built to target. A buyer opens Google’s AI Mode and asks the same thing in slightly different words. The answer names three other companies. Your page is still indexed and still ranking, but they got the introduction.

For twenty years, holding a top position meant getting found. Now, AI search has created a new way to be found, and its users are increasing rapidly.

How Query Fan-Out Turns One Question Into Many

Google’s AI Mode searches beyond strictly the exact question it’s asked. It generates a set of related questions and assembles the answer out of all of them.

Google calls this query fan-out. The term arrived with the AI Mode announcement in March 2025, and Liz Reid, Google’s Head of Search, explained the mechanism at Google I/O that May. The definition is a set of related queries the model generates to request more information. Reid’s post describes the system breaking a question into subtopics, so the sub-queries cover differences in phrasing as well as the angle of questioning.

Chuck McCarthy, Xponent21’s Service Operations Lead, described the behavior in his own words on Episode 3 of To the Power of X, the agency’s podcast on AI and marketing. The model, he said, “fans out to a bunch of other questions that an expert would ask” (13:22).

The fan-out multiplies your question while still orienting around and searching the original. The technique that replaces one question with another is query rewriting, a separate mechanism.

Why the Page-Per-Phrasing Fix Fails

Many agencies and tool vendors seem to understand this concept, but often treat every sub-query as a separate target. The advice circulates under the banner of Generative Engine Optimization (GEO), the practice of optimizing content to appear in AI-generated answers, and it usually arrives in three parts: a page for every fan-out query, the long page split so each sub-question gets its own URL, and the content cut into passages short enough to be retrieved alone.

Google published guidance on this, and it contradicts the page-per-variation advice. Its guide to optimizing for generative AI features says that creating separate content for every variation of how people might search, fan-out queries included, violates Google’s scaled content abuse spam policy when it is done primarily to manipulate rankings or AI responses. Even if not doing so to intentionally influence AI search results, the guide still calls the approach ineffective in the long run, because a high quantity of pages does not make a site more relevant. It sets no chunking requirement either, and tells publishers not to worry about whether they have captured every variation of how someone might search for their content.

The fan-out is happening. The leap from that mechanism to a page count is where some marketers diverge from best practices.

Splitting one question across its phrasings produces pages that all sit on the same topic and compete with each other for it, and it hands a reader the same thing several times. Covering distinct questions produces pages that each answer something different and point at each other, which is a content cluster.

Best practice is a page per question, not a page per phrasing.

We sell content production, which is exactly why that rule has to govern the AI search work we do. Undifferentiated volume is a violation.

On Episode 3 the crew was talking about the same failure from the reader’s side, output so alike that audiences scroll straight past it. Machine learning operates similarly. LLMs are built out of everything already published and holds every version of the common answer, so it favors original content with actual sources. 

Tracking 78 Buyer Questions Across Four AI Assistants

Checking whether an assistant names you takes an instrument. The first instrument Xponent21 evaluated wasn’t ours. In June 2025, our CEO, Will Melton, wrote about a third-party browser tool that exposes the backend searches ChatGPT fires behind a prompt. His instinct was right, but analyzing single prompts one at a time isn’t sustainable, so we built our own instrument. 

What we run now is our own internally developed tool. Xponent21 asks four AI assistants the same 78 buyer questions every day in CARL Intelligence. The set covers branded, local, category, comparison, and informational questions, written the way a buyer would actually phrase them. AI answers shift from day to day, so the only way to know where you stand is to keep asking.

The majority of questions that return us name Xponent21, or name a person, or name a city. Richmond, Virginia shows up. Our people show up. As the question broadens toward generic informational and how-to phrasing, the set stops returning us and starts returning other people’s domains.

We read that pattern as the assistant naming a specific source it can credit the answer to. Coverage of more phrasings gives it more text. Only a source gives it something to name.

A question that contains our name is asked by someone who already knows it. So part of what we are measuring is awareness the brand already has. Seventy-eight questions is also a small sample. The evidence provides direction, and that is all we are claiming.

Assistants differ in how much of their searching you can see. ChatGPT’s fired searches are capturable, demonstrated in our 2025 article. Google’s fan-out sub-queries stay hidden, and Google says plainly that no third-party tool has access to its internal ranking or AI systems. For Google, the measurable surface is Search Console’s own generative AI performance report. Our instrument works from the outside. We ask a question set and record whether we are named in what comes back.

Which is why the work starts one step earlier than the tooling suggests, with the questions that actually matter to your buyers.

Ranking Still Decides Eligibility for AI Citation

Ranking still counts, according to data. In its July 2025 study, Ahrefs found that 76.1% of AI Overview citations came from pages ranking in the top ten. Its early-2026 study put that figure at 37.9%. The two figures measure differently, by Ahrefs’ own account, because it improved its citation parsing between them. The steadier number is the top-100 figure. 85.6% of AI Overview citations in that July 2025 study came from pages ranking somewhere in Google’s top 100.

Read together, they describe eligibility. Ranking gets a page into the pool the system draws from, and something else decides which page gets named out of it. Optimizing to rank in AI search results sets the floor here.

A page that genuinely answers a question does its work inside the assistant’s response. The buyer gets the answer and you get the attribution, whether or not they ever land on your site, often referred to as “zero-click search.”

So think back to that page at position three. Another company got the introduction, but you can use this information to make your brand more visible. Document the distinct questions a buyer asks on that topic, find which ones the assistants currently answer with someone else’s name, and develop your content on the page already targeting those queries. This is the same approach we use with clients, and a first pass needs only your buyers’ questions and an afternoon. The page that wins is the one that answers more questions.

Frequently Asked Questions About Query Fan-Out

• What is query fan-out in Google’s AI search?
• Should I create a separate page for each query fan-out variation?
• Why does a page that ranks on Google not appear in the AI answer?
• How can I tell which questions AI assistants answer with my brand?
• Does query fan-out mean keyword research no longer matters?