Query Fan-Out: The Hidden Sub-Questions That Decide Whether AI Recommends You

When a stranger asks an AI assistant “who’s the best commercial roofer in Tampa,” the model almost never runs that one query and reads back the top result. It quietly expands the question into a fan of narrower ones — commercial roofing companies Tampa, licensed roofers Hillsborough County, roofing contractor reviews Tampa, flat roof repair specialists Florida, roofing warranty comparison — runs several of them, and assembles a single confident answer out of whatever came back. You didn’t lose because you ranked eleventh for the headline phrase. You lost because you were absent from six of the eight sub-questions the model actually asked.

This is query fan-out, and once you see it, you can’t un-see it. It’s the single biggest reason smart operators keep optimizing for the wrong thing. They’re trying to win a query the AI never really runs, while the recommendation is being decided in a dozen quieter ones they’ve never looked at.

What query fan-out actually is

Traditional search was a matching problem. You typed words, the engine returned pages whose words matched, ranked by a mix of relevance and authority. One query in, one ranked list out. The whole discipline of SEO grew up around that one-to-one relationship: find the keyword, win the keyword.

Answer engines don’t work that way. A system like ChatGPT search, Perplexity, or Google’s AI Mode treats your question as an intent to be satisfied, not a string to be matched. Before it writes a word, it plans. It decomposes your question into the sub-questions a knowledgeable human would need to answer to give a good recommendation, retrieves sources for each, and then synthesizes. The industry name for the decomposition step is query fan-out or query expansion. Whatever you call it, the mechanic is the same: one prompt becomes many retrievals.

Think about how you’d answer “best commercial roofer in Tampa” if a friend asked and you knew the trade. You wouldn’t reach for a single fact. You’d think about who’s actually licensed and bonded, who does flat commercial roofs versus residential shingle work, who has a reputation that holds up, who won’t vanish when a warranty claim comes in. Each of those is a separate line of inquiry. The model does the same thing — just explicitly, and at machine speed.

Why this breaks the old keyword mental model

The keyword model tells you to find the phrase with volume and beat everyone on it. Fan-out tells you that phrase is a doorway, not a destination. Behind it sit the sub-questions that do the real filtering, and they’re often phrases with almost no traditional search volume at all — which is exactly why nobody optimized for them.

Here’s the uncomfortable part. Most of the sub-questions in a fan are qualifying questions, not discovery questions. Discovery is “roofers in Tampa.” Qualifying is “which of these are licensed for commercial work,” “which handle TPO and modified bitumen,” “which have credible reviews for warranty follow-through.” A business can dominate the discovery query and still get filtered out in qualifying, because the qualifying facts simply aren’t present anywhere the model can find them. The model can’t recommend a strength you never made legible.

In the “State of AI Search 2026” work we did at AIrecommend.ai, the pattern that kept surfacing was this gap between being findable and being qualifiable. Plenty of businesses show up when the model searches broadly. Far fewer survive the specific sub-questions, because the specific answers live only in the owner’s head or on a page no crawler ever indexed. That’s my read of the mechanism, not a claim about any one engine’s internals — but it’s held up across enough categories that I plan around it.

There’s a second-order effect worth naming. Because the model synthesizes across the whole fan before it answers, a single strong sub-question rarely carries you, and a single weak one rarely sinks you — but a pattern of absence is fatal. If you’re missing from five of eight sub-questions, the model quietly concludes it doesn’t know enough about you to recommend you with confidence, and confidence is exactly what a recommender is protecting. It would rather name the competitor it can describe completely than gamble on the one it only half-understands. That’s why coverage compounds: each sub-question you can answer doesn’t just win that slot, it raises the model’s overall willingness to put your name in the answer at all.

How to see the fan for your own category

You don’t need special tooling to map your fan. You need to think like the model and then check your work.

1. Write down the headline prompt a real customer would type. Not a keyword — a full, human question. “What’s a good accountant for a small e-commerce business in Austin?”
2. List the sub-questions a careful human would ask before answering it. Licensed and in good standing? Actually works with e-commerce clients? Handles sales-tax nexus? Responsive? Reasonably priced for a small business? Aim for eight to twelve. These are your fan.
3. Run the headline prompt in three assistants — ChatGPT, Perplexity, and Google’s AI experience — and read what the model asked itself. Perplexity will often show you the sub-searches directly. The others reveal the fan through what they emphasize in the answer and which sources they cite. The sub-questions the model raises are the ones you have to be present for.
4. For each sub-question, ask: where would the model find my answer? If the honest answer is “nowhere public,” you’ve found a hole in your coverage. That’s the whole game.

Do this once and you’ll usually find the same thing every operator finds: you’ve poured effort into the one query you already sort of win, and left half the fan completely uncovered.

Fan-out coverage: the metric that matters

I’ve started using a simple frame for this with clients, and it travels well enough that I’ll put a name on it here: fan-out coverage — the share of the sub-questions in your category’s fan that your public footprint can actually answer, in a source a model would trust.

It’s deliberately not a precise score. It’s a discipline. If your category’s fan has ten meaningful sub-questions and your website, your profiles, and the third-party pages about you can credibly answer three of them, your fan-out coverage is thin, and no amount of hammering the headline keyword will fix it. If you can answer eight or nine — with specifics a crawler can read and, ideally, with corroboration from sources that aren’t you — you become the option the model can defend when it names a winner.

The reason coverage beats ranking is that answer engines are risk-averse recommenders. When a model has to pick, it favors the option it can most safely justify across the whole fan, not the one that shouted loudest on a single term. Broad, verifiable coverage is the safety.

Where businesses lose the fan

Three failure modes come up again and again.

The homepage that answers nothing specific. Beautiful hero image, a tagline about “quality and trust,” and not one concrete, retrievable fact. Licenses, service areas, specializations, what you don’t do — those are the sub-question answers, and vague copy erases all of them. A model can’t qualify you on adjectives.

Facts trapped in the wrong format. Your service list is baked into an image. Your credentials live inside a PDF nobody links to. Your specialization is mentioned once, in a testimonial, in the middle of a video. The fact exists; it just isn’t legible. Fan-out rewards machine-readable text, plainly stated, near where the topic is discussed.

One-source coverage. Every answer to every sub-question comes only from you. That works for basic facts, but on the judgment-heavy sub-questions — reliability, quality, whether you deliver — models weight what other people say. If the only voice vouching for your warranty follow-through is your own marketing, you’ll lose that sub-question to a competitor a review site actually described.

What to actually do

Fan-out coverage is buildable, and it’s mostly unglamorous work.

Start by writing the answers to your fan in plain, specific text and putting them where they belong — on service pages, an honest FAQ, an about page that states credentials as facts rather than vibes. One clear sentence that answers a real sub-question is worth more than a paragraph of positioning. Say what you do, for whom, where, with what qualifications, and — underrated — what you don’t do. Negative facts are strong qualifiers; they help the model rule you in for the right query by ruling you out of the wrong one.

Then make the important facts machine-readable. Real text, not images. Structured data where it fits. Consistent naming, address, and category everywhere you appear, so the model isn’t reconciling three versions of you. This is entity hygiene, and it’s the floor of the whole thing.

Finally, get the judgment sub-questions corroborated by someone who isn’t you. Reviews that describe specifics rather than just star counts. A trade directory that lists your specialization. An interview, a case write-up, a mention in a piece about your field. Every independent source that answers one of your qualifying sub-questions is another slot in the fan you now cover.

The honest caveats

A few things I want to be straight about. Fan-out is not a fixed, published algorithm you can reverse-engineer to a checklist — it varies by engine, by category, and it changes as these systems evolve. The sub-questions I’d list for your category are my informed reconstruction, not a leak of anyone’s internals. And covering your fan doesn’t guarantee the recommendation; it earns you eligibility for it. Plenty of other things — freshness, the specific phrasing of the user’s prompt, how much competition sits in your fan — move the final answer.

But here’s what I’m confident about. The businesses winning AI recommendations right now are not, for the most part, the ones with the cleverest single-keyword play. They’re the ones whose public footprint can answer the whole quiet fan of sub-questions a model raises before it’s willing to name a winner. Stop optimizing the question the AI barely runs. Go win the dozen it actually asks.

About the author

Scott Tischler is the Founder & Chairman of AIrecommend.ai and a practitioner-authority on AI search and Answer Engine Optimization. With 20+ years in marketing technology — including American Express, MetLife, and UBS — and executive and professional study at Wharton, Harvard, and Oxford, he helps businesses become the ones AI recommends.