Query fan-out is what happens between a user's question and an AI's answer: the assistant takes one prompt, splits it into several separate web searches, runs them all at once, and builds its response from whatever those searches return. The user asks a single messy question. The machine translates it into a handful of clean ones. Those hidden queries decide which pages get read and which brands get named. The original prompt doesn't.
That's the whole concept. The rest of this article covers what it means for your site, how to see the actual queries, and how to rank for them.
Why does AI search fan out at all?
Because one search can't answer a real question.
Ask an assistant "what's the best CRM for a 10-person real estate team?" and no single page settles it. So the model decomposes the question. It might search for best CRM for small teams, CRM for real estate agents, and a pricing comparison. Three queries, three result sets, one synthesized answer.
Language models also have a knowledge cutoff. Anything that depends on what's true right now, like prices, rankings, best-of lists, or who's still in business, pushes the model to search rather than recall. A useful rule of thumb: if the straight answer to a question could have changed in the last six months, the assistant is probably searching for it.
Two consequences follow, and they're the foundation of GEO:
- AI search mostly runs on ordinary search indexes. The fan-out queries hit Google and Bing. Ranking still matters. It just matters for queries you've never seen.
- Every fan-out query is a separate door. A question you'd count as one keyword in classic SEO becomes three or four retrieval opportunities. Miss all of them and you don't exist in the answer.
How can you see the actual fan-out queries?
You don't have to guess. In your own ChatGPT sessions, you can watch the searches happen.
- Open ChatGPT in a desktop browser.
- Open DevTools (F12 or right-click → Inspect) and switch to the Network tab.
- Ask a question that forces a search: something time-sensitive or comparative in your niche.
- When the interface shows it's searching, look through the network responses for that conversation and search the payload for the query terms.
- You'll find the literal search strings the model fanned out to.
Do this ten times with ten realistic customer questions and you'll have something most of your competitors don't: a list of the real queries AI runs when your category comes up. Patterns show up fast. Fan-out queries tend to be shorter, plainer, and more literal than the prompts that triggered them. Users ramble. The model searches like a librarian.
This is the closest thing to keyword research that exists for AI search right now, so treat the output like a keyword list. That's exactly what it is.
Two practical notes. First, fan-outs vary between sessions. The model rephrases, so run each question a few times and keep the queries that recur; the recurring ones are your targets. Second, record the queries verbatim, punctuation and all. The temptation is to summarize them back into the keywords you already planned to target. Don't. The gap between what you'd have guessed and what the machine actually searched is the entire value of the exercise.
Google gives you a second window from the other direction. In Search Console, filter queries with the regex (\b\w+\b\s){7,} to isolate long, conversational searches, the kind of phrasing AI systems and AI-influenced searchers produce. We cover that workflow, plus the GA4 side, in how to measure AI search traffic.
How do you optimize for fan-out queries?
The same way you'd optimize for any query, just with more discipline, because the searcher is a machine and machines don't squint.
Start with the rule that organizes everything else: one query, one page. For each fan-out query that matters to your business, a specific page should answer it directly. Not a section buried in a 4,000-word guide. A page whose entire job is that query.
Put the query language in the five placements: title, meta description, URL slug, H1, and the beginning of the first sentence. This is old-school on-page SEO, and it works better in AI search than it has worked in classic search for years, because retrieval systems reward literal relevance. When the model searches "CRM for real estate agents," the page titled exactly that, opening with exactly that, is the easy pick.
Answer in the first three sentences. The model reads to extract, not to browse. If your answer starts after 600 words of preamble, you've made yourself expensive to cite. Answer first, elaborate after. Every article on this site opens that way for the same reason.
Match the machine's vocabulary, not your own. If the fan-out queries say "pricing" and your page says "investment options," you have a language mismatch, and the retrieval system won't do the translation for you. Rewrite the page to speak the query's language, high on the page.
Finally, cover the cluster, not just the head. Fan-outs for a commercial question usually include a comparison query, a pricing query, and a "best for [niche]" query. That's an argument for comparison tables, plain pricing pages, and specific "best X for Y" content: the formats that keep showing up in AI answers. We break down what gets a page pulled into responses in how to get cited by ChatGPT.
What happens after the fan-out?
The fan-out is round one. When an assistant finds brands worth considering, it often runs a second wave of narrower searches to vet them. It digs into a specific company's reviews, pricing, and reputation, sometimes searching within a single site. Those follow-ups are called grounding queries, and they're where recommendations are won or lost. Surviving round one gets you considered. Round two decides whether you get named.
Which brands make it into that second round isn't random either. It tracks the evidence models can retrieve about you, and we cover that in how LLMs pick brands.
Does this change what good SEO looks like?
It sharpens it. Nothing about query fan-out rewards tricks. It rewards a site where every important query has a clean, literal, answer-first page, which is what good SEO always was, executed without laziness.
What changes is the unit of planning. Classic SEO planned around keywords people type. GEO plans around queries machines generate: plainer, more numerous, and invisible unless you go looking. The work of finding them, mapping pages to them, and hardening the pages so citations stick is the core of our Reforge Method.
Query fan-out in one checklist
- One user question becomes several machine searches. Optimize for the searches.
- Watch real fan-outs in DevTools during your own ChatGPT sessions. Build a query list from what you see.
- Mine Search Console with
(\b\w+\b\s){7,}for long conversational queries reaching your site. - Give every valuable query its own page, with the query in title, meta, slug, H1, and first sentence.
- Answer in the first three sentences. Machines don't scroll hopefully.
- Use the machine's vocabulary, not your brand's euphemisms.
- Build the comparison, pricing, and "best for" pages the fan-out predictably includes.
- Then prepare for round two: grounding queries, where the recommendation is decided.
A scope note against our own pitch: if your money pages don't answer their own titles yet, fan-out analysis is premature. Mapping sub-queries for pages that can't win any of them is measurement without a use. Fix the pages, then come back to this.
Start with the DevTools session. Ten questions tonight gets you a real query list by tomorrow, and every page you fix after that is aimed at searches you've seen with your own eyes.