Query Fan-Out and Why AI Search Decomposes Your Customer's Questions

A client called me last quarter convinced his best page was broken. It ranked first in classic Google for his money keyword, traffic was fine, and yet ChatGPT and Google's AI Mode never mentioned his company when people asked the obvious buying question. He wanted to know what penalty he had triggered. There was no penalty. The page was answering one question well in a world where the machine had quietly stopped asking one question. It was asking eight at once, and his page only had an answer for one of them.

That mechanism has a name, and once you understand it, half the confusing behavior in AI search stops being mysterious. It is called query fan-out, and it is the reason your ranking for a single phrase no longer decides whether you show up in the answer.

What query fan-out actually is

Query fan-out is an AI search technique that splits a single user query into multiple sub-queries, issues those sub-queries in parallel, and merges the results into one answer. Google AI Mode, ChatGPT, and Perplexity use query fan-out to turn one query into a fan of related sub-queries.

The single phrase your customer typed becomes a tree of related searches the system runs on their behalf, and your content competes inside each branch separately. Rank first for the phrase they typed and you can still lose, because that phrase is now one twig on the tree.

Google named the technique publicly at I/O 2025. Head of Search Elizabeth Reid described it on stage: AI Mode "calls on our custom version of Gemini to break the question into different subtopics, and it issues a multitude of queries simultaneously on your behalf," a description Google repeated in its AI Mode announcement. So the label is not agency jargon. It is Google's own word for how the system reasons.

Why AI search decomposes your customer's questions

The reason AI search decomposes a question is that the question your customer types is almost never the whole question they mean. Someone who searches "best half-marathon plan for beginners" also wants to know about gear, injury prevention, nutrition, and pacing, and they did not type a word of it. Classic search waited for you to be explicit. If you did not write "for runners," it would not infer it. AI search runs ahead of the user, fills in the unspoken context, and gathers material for the follow-up questions before they are asked.

The reason AI search decomposes a question is that the question your customer types is almost never the whole question they mean.

Marie Haynes documented this early, noting that AI Mode "ranks websites differently than traditional search" precisely because it is summarizing results from multiple queries rather than one. That is the part my client missed. He was not ranked lower. He was simply absent from seven of the eight branches the model explored, so when it assembled the answer, there was nothing of his to pull in for most of it. Aleyda Solis frames the fix as an "answer a facet" mentality: for any topic you target, you map the sub-questions a searcher might explore and answer each one directly.

How query fan-out works under the hood

There is real machinery behind the buzzword, and it is documented. In December 2024 Google filed a patent called Thematic Search, US12158907B1, which describes taking a query, deriving a set of themes from it, generating synthetic queries for those themes, and assembling results into a single themed response. Search Engine Journal's teardown of the filing notes that the patent "closely parallels AI Mode's query fan-out technique." The word "synthetic" matters here, and I will come back to it, because it changes what you can and cannot optimize for.

Strip the patent language away and the pipeline runs in three moves. First the system decomposes your query, rewriting it into sub-queries and projecting the latent intents around it. Mike King's team at iPullRank describes this expansion stage as where the system maps the slots it needs to fill and generates speculative follow-ups, which means "the real competition is now at the subquery level." Second it routes each sub-query to whatever source and format fits best, then retrieves passages from all of them at once. Third it scores and merges those results. Ahrefs points out that the merge often uses reciprocal rank fusion, a method that rewards a document for showing up consistently across several of the sub-query result sets rather than ranking once for one term. A page that surfaces in four branches beats a page that ranks first in one.

That scoring detail is the whole strategy in a sentence. The system is not grading your page against the query your customer typed. It is grading your content against a constellation of questions the model invented, and it favors the sources that keep reappearing.

How many sub-queries one question really becomes

This is where the numbers get uncomfortable for anyone still optimizing one page for one keyword. Ahrefs analyzed the behavior and found that for a simple shopping prompt, AI Mode runs 5–11 searches vs ChatGPT's 420 for the same intent. Independent research from Seer Interactive and Nectiv, which studied tens of thousands of fan-outs, landed in a similar range: an average of 9 to 11 sub-queries per prompt, with 59% of prompts triggering 5 to 11 searches and 24% triggering 12 to 19, reaching as high as 28. One customer question routinely becomes a dozen.

Here is the trap inside those sub-queries, and it is the thing most "optimize for fan-out" advice gets wrong. The sub-queries are synthetic. The model generates them on the fly, they differ between runs of the same prompt, and Seer found that over 95% of them have no recurring search volume at all. You cannot put them in a keyword tool and target them as long-tail terms, because most of them are not terms anyone searches. You target the patterns behind them instead. I went deeper on the broader shift this represents in relevance engineering and the future of search, and query fan-out is the clearest example of why the keyword-at-a-time model broke.

What query fan-out means for your content

If one question becomes a dozen sub-queries, then comprehensive coverage of a topic beats a perfect page on one slice of it. This is not a vague call to "write more." It is specific. For any topic you care about, the work is to identify the facets a searcher and the model will both explore, then own a clear, self-contained answer for each one on a page or cluster that holds together.

Three things move the needle, in my experience auditing this on client sites. The first is intent coverage. A page that answers the core question and the four adjacent ones the model is likely to spin off gives it four more reasons to retrieve you. The second is chunking. The system selects passages, not pages, so each answer needs to live in its own tight, labeled block under a heading that matches the sub-question, dense with the entities and specifics that make it usable when lifted out of context. The third is topic clusters that link the facets together, which is how you signal topical authority across the whole branching tree instead of one leaf. None of this is exotic. It is the difference between a page that mentions a subject and a page built so a retrieval system can pull a clean answer from it.

How I audit a page for fan-out coverage

When I sit down with the client page that started this story, the process is boring and it works. I take the core question and I write out the sub-questions a reasonable buyer would also have: cost, comparison, timeline, risk, what happens next. Tools help here, but a sharp human and ten minutes of thinking get you most of the way. Then I run the page against that list and mark every sub-question it does not answer cleanly. Those gaps are the branches you are losing.

For my client, the page sold the service well and said almost nothing about price ranges, how it compared to the two obvious alternatives, or what the first month of working together looked like. Those were exactly the sub-queries the model was running and finding other companies' answers for. We did not rewrite the page into a wall of text. We added scoped, self-contained sections, each one answering a single facet under a heading that named it, the way you would build a passage you wanted a machine to quote. Within a few weeks his company started turning up in answers it had been invisible for, and the classic ranking did not budge, because comprehensive coverage tends to help both surfaces at once.

The measurement problem nobody warns you about

You will not see this neatly in a dashboard, and you should know that going in. Because fan-out queries are synthetic and largely zero-volume, rank tracking cannot show you which branches you appear in. What you can do is track AI referral traffic from chatgpt.com and perplexity.ai in your analytics, and run your real customer questions through the major models on a schedule to record whether your brand surfaces and where. Treat those numbers as directional. None of the AI platforms publish a Search Console equivalent yet, which means the work is less about a metric and more about coverage you can reason about. I made the broader case for why the old reporting reflexes fail in the piece on how nobody scrolls anymore.

Query fan-out is not a tactic you bolt on. It is the shape of how AI search now thinks, and it rewards the same thing it always should have: genuinely covering what your customer needs to know instead of guarding one keyword. The brands that map the fan and answer every branch with a clean, retrievable passage are the ones the models keep naming. If you want that built as engineering rather than guesswork, it is the core of the work we do as an enterprise SEO consultant, and it sits at the center of our AI search survival manual.

By Michael McDougald