Here is something almost no one who talks about "SEO for AI" knows: when someone asks ChatGPT or Google's AI Mode a real question, the system does not search for their question. Not once. It rewrites it into five, ten, sometimes twenty separate hidden searches, runs them all at the same time, and stitches the answer from whatever it finds. Your one page is not competing for one query anymore. It is competing to be the best passage for a fan of sub-queries you will never see.
That mechanic has a name. Query fan-out. And if you are optimising your site the way you did in 2023, you are optimising for a game that is not being played.
Here is how badly the rules have shifted. A December 2025 analysis of about 174,000 URLs found that roughly 68 percent of the pages AI cited in its answers were not even in Google's top ten organic results for the same query. Read that again. The pages the AI trusted enough to name were, in the majority, pages that never ranked. A separate 2026 study put the overlap between top rankings and AI citations at about a quarter to a third. So a brand that only chases rank misses most of the AI citation opportunities. Not some. Most.
The reason is fan-out. Ranking answers "which page is best for this exact query." Fan-out asks "which passage best answers each of these hidden sub-queries," and the winning passage for a narrow sub-query often sits on a smaller page that never ranked first for the head term. Different question, different winner.
The mechanic itself takes one sentence to describe. Your reader types one question. The engine's language model rewrites it into a set of sub-queries, each targeting a facet of what the reader probably actually wants. All the sub-queries run at once against the live web. The engine grabs specific passages, not whole pages, and picks the best ones. It writes one answer and cites a few sources. Google confirmed this publicly when it launched AI Mode, and its own patent calls it "query variant generation."
How wide the fan opens depends on the engine. Google's AI Mode fans wide: most prompts fire five to eleven parallel sub-queries, and its Deep Search can run hundreds. ChatGPT is surgical: it only searches when it needs fresh data, and when it does, it fires a handful of precise sub-queries. Perplexity typically runs a small fan and reviews ten or so sources before writing. Simple factual questions may not fan at all. But the moment a question is layered or commercial, which is exactly the kind of question worth winning, the fan opens.
Here is the surgical-strike detail that stops most makers cold.
When ChatGPT does fan out, it does not search the open web from scratch. It leans heavily on Bing's index. Analyses in 2026 put the overlap between ChatGPT's citations and Bing's top-ten results at around 87 percent. So if you rank number one on Google for your money keyword, and you are nowhere in Bing's top ten for the sub-queries under it, ChatGPT does not cite you. Full stop. Setting up Bing Webmaster Tools takes about ten minutes and is free. Your Bing rank for each sub-query is a leading indicator of whether ChatGPT will cite you. Most people who care about being cited by ChatGPT have never opened Bing Webmaster Tools. That is the free win nobody takes.
Now the mechanical heart of it, and the piece that changes how you write. AI systems do not grade whole pages. They chunk content into semantic passages and score each passage on its own. A single well-written paragraph from a small site can beat a five-thousand-word guide from a major domain, if that paragraph answers the sub-query more directly. Research on AI Overview extraction in late 2025 put the sweet spot for an extractable passage at roughly 130 to 170 words. Long enough to answer completely. Short enough to be lifted cleanly into a synthesis.
That means your unit of optimisation is no longer the page. It is the self-contained passage: a section that answers one question fully, on its own, without needing the paragraphs around it for context. A page can rank #1 and still lose because none of its sections are extractable on their own.
The ten-minute test you can run right now
Here is how to see the fan-out for your own topic, using nothing but a browser and a fresh chat window.
Pick one high-intent question a buyer in your category would ask. Not your brand name. A problem, phrased naturally: "what is a good tool for X" or "how do I solve Y without Z."
Open ChatGPT or Claude in a new chat with no history. Paste the question with one small change at the end: "If someone asked you this, what sub-questions would you search the web for before answering? List them." The list you get back is not the real fan, but it is a strong first draft of it. You will see six to ten sub-questions you never wrote content for.
Now open your site and honestly ask yourself: do I have a self-contained 130-to-170-word passage that answers each of those sub-questions? For most sites, the answer is no. And that gap, question by question, is where AI is citing your competitors instead of you.
If you want the real fan-out that a specific engine actually fires, the free Backlinko Chrome extension exposes ChatGPT's sub-queries, cited sources and entities for any prompt you run. Perplexity shows its citations natively in every answer. Both are worth ten minutes of your afternoon.
What to do with what you find
The strategic shift is this: you are no longer writing pages for keywords. You are writing self-contained passages that answer specific hidden sub-questions, organised into topic clusters that cover the whole question space around your subject. The pillar-plus-cluster structure, atomic answers of ~150 words each, and a real refresh date matter more than any single ranking signal you optimised in 2023.
The specifics of how to predict the fan reliably, build the entity-attribute matrix that turns messy sub-queries into a content plan, use the peer-reviewed tactics that increase citation rates up to forty percent (from Aggarwal et al., ACM KDD 2024), and set up the free monthly loop that tells you if any of it is working, live in the play book I wrote on this: Query Fan-Out: SEO for AI Search https://nataliiap.gumroad.com/l/seo-fan-out, part of a trilogy on AI search visibility for web apps.
But the ten-minute test above needs no book. Run it today. If your site cannot answer at least four of the sub-questions the model lists back, you now know exactly why AI keeps naming somebody else.
