Query Fan-Out: Why AI Mode Changes the Question You're Optimizing For

The mental model most teams carry into AI search is inherited from SEO: there is a query, there is a set of candidate pages, the best one wins. AI is just a different ranking function.

That model is wrong in one specific way, and the error compounds. In a conversational search surface, the question the user typed is frequently not the question the system searches for. It is decomposed into several, run in parallel, and the answer is synthesised from the union of what comes back.

This is query fan-out, and once you see it, a lot of otherwise confusing behaviour becomes obvious.

What actually happens

Someone asks: “I run a 12-person agency, what project management tool should we use and is it worth paying for the premium tier?”

A traditional search engine matches that against an index and returns documents. It is one query.

A reasoning-driven answer surface does something else. It recognises several separable sub-questions and issues its own searches for them — something along the lines of:

• project management tools for small agencies
• best project management software for teams under 20
• [tool] pricing tiers comparison
• [tool] premium features worth it
• project management software reviews agencies

Then it reads across all of those results and composes one answer. Google’s documentation on its AI features describes AI Mode as using this kind of decomposition — issuing multiple related searches to assemble a response — and its guidance for succeeding in AI features is built around that behaviour rather than around single-query matching.

The user issued one question. The system issued many. You were never competing for the question you saw.

Four consequences that change what you should do

Your best page can lose to a broader competitor. Suppose you have the single finest “project management for agencies” page on the web, and nothing else. Your competitor has five mediocre pages covering pricing, comparisons, reviews, small-team fit and feature depth.

The fan-out surfaces your competitor in four of five sub-searches and you in one. The synthesised answer draws on the weight of evidence, and the weight is theirs. You won the query and lost the answer.

This is a genuine reversal of the SEO instinct, which says depth on one asset beats thin coverage across five. Under fan-out, coverage of the decision beats perfection on any single node of it.

You are competing for questions nobody types. “Is the premium tier worth it” may have negligible search volume. It has a great deal of fan-out volume, because it is generated as a sub-query whenever anyone asks about pricing conversationally. Your keyword tool will never show it, because nobody searches it directly. This is one more reason volume-based prioritisation misfires here — see there is no keyword volume for AI prompts.

Long, specific, personal questions are now viable. The reason people ask assistants questions they would never type into a search box is that the assistant can handle them. “I run a 12-person agency” is context a search engine could not use and a decomposer can — it becomes a constraint on several sub-searches at once. The consequence is that highly qualified questions, previously unreachable, now route to whoever covers their component parts.

One weak sub-answer contaminates the whole thing. If four sub-searches surface you favourably and the fifth returns a two-year-old thread describing a problem you fixed, that thread is in the context the model is reading. Answers frequently carry a “though some users report…” clause traceable to exactly one stale source. Your average is good; your worst source is what gets quoted.

What to do differently

Map the decision, not the keyword. For each buying decision you care about, write down the questions a buyer works through: what are my options, how do these two compare, does it fit my situation, what does it cost, is it worth it, what goes wrong. That list approximates a fan-out. Coverage across it is the goal.

Build clusters, not flagships. A hub with five satellites covering the sub-questions outperforms one exhaustive page, because it can be retrieved five times. This is close to the classic topic-cluster model, which is convenient — the structure was already good practice and the fan-out mechanism gives it a sharper rationale. Building topical authority covers the construction.

Answer each sub-question in its own self-contained passage. Retrieval operates on chunks, not documents. A page where the pricing discussion is a clearly delimited, independently readable section can be retrieved for the pricing sub-query. One where it is woven through a narrative cannot be extracted cleanly.

Audit your worst source, not your best. Search your brand plus “problems,” “issues,” “vs,” “alternatives,” “review.” Whatever is stale, negative and prominent there is being retrieved by some sub-query. Addressing it — updating it, responding to it, or publishing something that answers it better — is often higher-value than another new page.

Track the sub-questions, not just the headline prompt. If your tracked prompt set is ten broad category questions, you are monitoring the tip of the fan-out. Adding the decomposed questions — pricing, comparison, fit, objection — gives you a much better picture of where you actually drop out. Choosing tracked prompts covers the selection, and multi-engine monitoring is where the pattern across engines becomes visible.

What we cannot see, and should not pretend to

Being precise about the limits, since this is an area where confident-sounding claims outrun the evidence.

The sub-queries are not exposed. Google does not publish the fan-out for a given question. Anything presenting itself as “the exact sub-queries generated for your topic” is inferring, not reading. The decomposition is real; a specific enumeration of it is a reconstruction.

Decomposition behaviour varies by engine and by question. Not every query fans out, and different surfaces do it differently. Treating fan-out as a universal constant of AI search overstates it.

Attribution within a synthesised answer is partial. When an answer draws on eight retrieved documents and names three sources, the influence of the unnamed five is invisible. Citation counts undercount influence, structurally, and no tool resolves this.

What survives all three caveats is the strategic implication, which does not depend on seeing the sub-queries: coverage of a decision beats optimisation of a page. That conclusion holds whether the fan-out is three queries or fifteen.

The relationship to AI Overviews

Worth separating, since they get conflated. AI Overviews is a summary surface layered on top of a results page — you still see the links. AI Mode is a conversational surface where the answer is the product and follow-ups are expected.

The fan-out mechanism appears in both, more aggressively in the conversational one, because a follow-up question inherits the whole prior context and decomposes against it. Google AI Overviews vs AI Mode covers the surfaces properly; the optimisation implication is the same in each, just stronger in the second.

The summary

The shift is from “which page ranks for this query” to “how much of this decision does our material cover.”

That is a harder question, and it rewards a different shape of investment — breadth across a decision rather than depth on a term. The teams that struggle with AI visibility despite good SEO are frequently the ones with one excellent page per topic, competing against sets. The fix is not better pages. It is more of the decision covered adequately.