Query Fan-Out Means AI Visibility Is Still Just SEO

Someone in Galway asks ChatGPT whether they should move their shop onto Shopify. It does not go looking for a page about moving a shop onto Shopify. It rewrites the question into several of its own, runs those, reads what comes back, then asks a few more. That rewriting step is called query fan-out.

Google’s AI Mode does the same thing a different way, firing its questions off all at once. Perplexity does it too, and will show you the breakdown if you ask.

None of them search for what your customer typed. They search for what they decided the question was really about, and your website is either inside one of those answers or it is not.

Query fan-out is also the reason you can ignore about half of what you are currently being sold. Follow those sub-questions to where they actually get answered and they land back in ordinary search results. Which makes AI visibility a change in what you rank for, not a replacement for ranking.

What query fan-out is

Query fan-out is the step where an AI system takes one question and turns it into several of its own before it goes looking for anything. It runs those sub-questions, gathers evidence for each, and assembles a single answer out of the pieces.

The name is Google’s. In its announcement of AI Mode, Google said the feature “uses our query fan-out technique, breaking down your question into subtopics and issuing a multitude of queries simultaneously on your behalf.” Dounia Berrada, a Search Senior Engineering Director at Google, put it more plainly in a company interview: “AI Mode is basically doing a dozen searches for you in the time it takes to do one.”

Google’s Deep Search runs the same method further, issuing hundreds of searches before it writes anything. AI Mode itself runs on a custom version of Gemini, and it went live for Irish users on 8 October 2025, as a tab on the results page and in the Google app. Your customers already have it.

The underlying idea is not new and not owned by anyone. It is a retrieval pattern: decompose the question, retrieve against the parts, synthesise the whole. What is new is that it now sits between your customer and your website on every major platform.

It is not just Google, and the platforms do it differently

This is the part most coverage skips, and the differences actually matter for how you write.

Google fans out sideways. Its own description is “simultaneously”. One question becomes a set of subtopics that all go off at once, and the answer is built from whatever each one brings back.

ChatGPT fans out forwards. OpenAI’s own support documentation says ChatGPT search “typically rewrites your query into one or more targeted queries” that it sends to its search partners. It then gives a worked example: a researcher asking about drugs targeting CCR8 for cancer might first produce the query “CCR8 immunotherapy drug development 2025”, and then, in OpenAI’s words, “after reviewing the initial results, ChatGPT search may send additional, more specific queries”, such as “CHS-114 conference 2025”.

Read that again, because it is the useful bit. ChatGPT’s second question is shaped by what it found in the first. That means the pages it picks up early influence what it goes looking for next. Being the source of a first-round answer is worth more than a single citation.

Perplexity fans out and shows its working. Its documentation says Pro Search “conducts multiple searches across the web” and that it “excels at breaking down ambiguous or multifaceted questions”. Deep Research goes further again, breaking a question into large numbers of retrievals run in parallel.

Microsoft Copilot sits on the same retrieval pattern. The names differ. The behaviour does not.

You will see numbers thrown around for how many sub-queries each platform generates. Treat them carefully. None of these companies publish the list, so any precise count is somebody’s sample, not a specification. Google’s own “a dozen” is the most solid figure available, and it was said in an interview, not a spec sheet.

The part the GEO industry skips, and it is all still search results

Here is the question nobody selling you a new acronym wants you to ask. Where do the answers to those sub-queries actually come from?

They come from search results.

OpenAI says it plainly in its own documentation: ChatGPT search “partners with other search providers” and rewrites your question into targeted queries that it sends to them. It is running searches on a search engine and reading what comes back. Google’s AI Mode issues its sub-queries into Google’s index, which is the same index that decides who ranks for what. Perplexity conducts multiple searches across the web.

So when a business owner asks me whether they need a separate AI strategy, a GEO consultant and a new budget line, the honest answer is no. If you do not rank for the sub-query, you are not in the pool the answer gets built from. Ranking is the entry ticket. You cannot be cited from a result nobody retrieved.

There is one honest caveat, and it matters. Ranking gets you considered, not quoted. These systems also pull from sources beyond the ten blue links, including Google’s own Knowledge Graph and shopping and product data, and once your page is in the candidate pool the writing on it decides whether anything gets lifted out. So ranking is necessary and not sufficient.

Which means the work splits cleanly in two. Rank for more questions than you used to, because there are now nine of them behind every one. Then write each answer so it can be taken out and used. The first half is SEO exactly as you already know it. The second half is the part this article is really about.

What one question turns into when it fans out

Easier to show than to describe. Pick one of these and look at what has to be answered before anybody could give a sensible reply.

Count the lines. That is eight or nine separate questions hiding inside one, and the uncomfortable part is that no single website answers all of them well. Including yours. Including ours.

So the machine takes a bit from here and a bit from there. Every line is a door, and most of them are currently being answered by somebody who is not you.

Watch a query fan-out happen in about two minutes

You do not have to take any of this on faith, and this is the single most convincing thing you can do before you change a word on your website.

Open Perplexity and run Pro Search on a question a real customer would ask about your business. Then open the steps it shows above the answer. Perplexity’s own documentation says you see “not just the summary, but also how the AI broke down your question and approached the research”. The decomposition is on screen, for free, in about two minutes.

Write those sub-queries down. That list is the thing everybody is selling you a tool to guess at, and you just read it off the screen.

Then look at which websites got cited for each part. That list is your actual competition now, and it is usually not the list you have been tracking. Do the same with your own main service and your own county, and you will have a content plan by the end of your coffee.

This is also the test to apply to anyone pitching you AI visibility work. Ask them to share their screen and show you a brand appearing for a specific prompt, live. If they cannot do it in front of you, they are selling a theory.

Why paragraphs now compete instead of pages

Here is the part that changes the work.

When a search engine ran one query, it looked for the best page. When it runs nine, it is looking for the best piece of evidence for each of those nine, and those pieces can come from nine different websites. The unit of competition has moved down a level. It used to be the page. Now it is the passage.

That is not a guess about how it works. Perplexity’s own engineering writing describes its retrieval infrastructure dividing documents into fine-grained units that are surfaced and scored individually. Passages compete, not pages.

Think about what that does to the two kinds of content most Irish businesses have.

The first is the long guide that touches everything and settles nothing. Three thousand words that mention price, timelines, options and mistakes, but never answer any of them cleanly in one place. That page can rank perfectly well and still lose every sub-query, because there is no self-contained passage to lift out of it.

The second is the short, specific page that answers one question completely. It might never have ranked for a competitive head term. But it can win a sub-query outright, because it contains exactly the thing the machine went looking for.

Nine paragraphs that each answer one question properly will beat one page that half-answers nine. That is the whole shift in a sentence.

Why you can sit on page one and never get cited

This is the bit that frustrates business owners who have already invested in SEO, and it is worth being straight about.

Ranking well for your main keyword does not guarantee you appear in an AI answer, because the system is not searching for your main keyword. It generated its own questions and it is matching against those. If none of your writing answers any of them in a liftable form, you can be sitting third on a results page fewer people are reading.

The reverse is also true, and it is the good news. A small business with genuinely specific pages can get cited in answers where it has no hope of outranking the big players on the head term. The fan-out spreads the opportunity around, because it rewards depth on narrow things. Which means you can end up named in the answer to a question your biggest competitor is too broad to have bothered writing about, on a page that took you an afternoon.

Google also builds in a get-out. It has said that where there is not high confidence, users see a set of ordinary web results instead of an AI answer. Vague, hedged, unsupported writing does not just fail to get cited. It makes the whole answer less confident.

Does query fan-out kill keyword research?

No, and anyone telling you otherwise is usually selling the replacement.

Keywords still tell you what people want and roughly how many of them want it. That has not changed. What has changed is that the keyword is now the starting point of the research rather than the target of the page.

The practical difference: you used to pick a keyword and build a page for it. Now you pick a keyword, work out the questions that sit underneath it, and make sure each of those has a home. Sometimes that is a section. Sometimes it is a page of its own.

Most of the sub-questions a system generates are not things anybody types into a search box, so they will never show up in a keyword tool with a volume beside them. That does not make them worthless. It makes them invisible to the tool, which is a different problem.

Which is why the best research for this is not a tool at all. It is your own phone calls, and it is Reddit. Search Reddit for your sector and read how people actually phrase the problem when they are not talking to a salesperson. The questions they ask each other are far closer to the sub-queries a machine generates than anything in a keyword report, because both are phrased the way a person thinks rather than the way a person searches.

How to write a passage that survives a query fan-out

Four things, and none of them require new software.

Answer first, then explain. Put the actual answer in the opening sentence of the section, then earn it over the paragraphs below. If a reader has to get through a hundred words of setup before you say anything, so does the machine, and it will take the passage that got there first.

Make each passage stand on its own. Read any paragraph in isolation. If it depends on the one above it to make sense, it cannot be lifted out. That means naming the thing rather than saying “it”, and saying “a Shopify migration” rather than “this process”.

Be specific enough to be checkable. Real figures, real timescales, real names, real places. “Costs vary depending on your requirements” is unusable. A number with the reasoning behind it is quotable. If you cannot support a figure, do not invent one, because a claim that falls apart is worse than no claim.

Give each question its own heading, in your customer’s words. Headings are how a passage gets found and framed. Write them as the question somebody would actually ask, not as a clever label. Comparison tables work unusually well for the same reason: a table of prices or features is already a self-contained answer. Our page on header tag structure covers how to build that properly.

To be clear about what that is and is not: it is a structural habit, not a writing style. You are not writing for a machine and you should not start. Nobody needs keyword-stuffed sentences or robotic phrasing, and that sort of thing reads badly to the person who does land on the page. All you are doing is making sure each answer is complete where it sits, which is what a good editor would have told you anyway.

Name things the way your customers name them

The fan-out works on meaning, not on exact strings. To take part in it, your writing has to contain the recognisable things a question is about: the service, the product, the town, the problem, the software, the sector. These are entities, and they are how a machine connects your page to a question that never used your words.

That is a real argument against vague marketing language. “Bespoke digital solutions for ambitious brands” contains nothing to attach a question to. “WooCommerce site speed for a furniture retailer in Louth” contains four things, all of which map onto something in Google’s Knowledge Graph.

A word on schema, because there is a genuine argument in the industry about it and you deserve the honest version rather than the one that suits an agency.

There is no evidence that schema markup makes an AI cite you. Nobody has shown a mechanism, and anyone selling structured data as an AI visibility tactic is ahead of the evidence. On that, the sceptics are right.

What schema does is help search engines understand what your page is, what your business is, where it operates and what it sells, which helps you turn up in the results those sub-queries are answered from. Under this article’s own argument, that is the whole point. Schema is not an AI tactic. It is an SEO tactic, and SEO is what feeds the AI. Worth doing, for the reason it was always worth doing, and not because it has been rebranded.

Cover the topic, not just the page

If one question fans into nine, the site that answers six of them is worth more than the site that answers one very well.

That is the argument for topical authority, and the fan-out is the clearest mechanical reason for it anyone has produced. Build a cluster: a main page on the subject, and separate pages underneath it that each own one real question, linked to each other so the relationship is obvious. Your migration guide links to your pricing page, which links to your page on what happens to rankings, and each one is a complete answer on its own.

Done properly the same cluster can be cited several times inside a single answer. Done badly it becomes pages repeating each other, which helps nobody. The test is simple: does each page answer a question a customer actually asks, in a way the others do not. Our internal linking strategy page goes into how to wire that together.

One kind of page earns its place more than any other here, and most Irish businesses refuse to write it: the honest comparison. X versus Y. What we do that the cheaper option does not. When you would be better off with somebody else. Comparison questions are among the most common things a fan-out produces, because a person weighing a decision is always implicitly comparing, and almost nobody publishes a straight answer to them. If you are willing to name your competitors and be fair about them, you will own a whole class of sub-query that your rivals have left empty out of nerves.

What this changes for an Irish business specifically

Local questions fan out just as hard as national ones, and often harder, because a local question carries more assumptions. Someone asking about an accountant in Drogheda is implicitly asking about proximity, about price, about whether the firm handles their sector, and about whether anyone has had a good experience. Four sub-questions minimum, from one short query.

Most local business websites answer one of those, badly, on a page called Services.

The opportunity is that very few Irish businesses are writing at this level of specificity yet. The gap between “we provide accountancy services” and a page that answers, in plain words, what a Drogheda retailer with two employees should expect to pay and how long onboarding takes, is not a technical gap. It is a willingness to be specific in public. That is still wide open, and it is the same thinking behind how AI is reshaping local SEO.

What not to do about it

Do not rewrite your whole website. The fan-out rewards good answers, and most sites already have a handful of pages that are close. Fixing the structure of what you have will move you further than starting again.

Do not buy a tool because it has AI in the name. No tool can show you the sub-queries Google generated, and Perplexity will show you its own for nothing.

Do not buy a separate AI or GEO retainer on top of your SEO. If the sub-queries are answered out of search results, then the thing that gets you into an AI answer is the thing that got you onto page one, and paying twice for one job is just paying twice. Ask anyone quoting you for GEO what they would actually do that a good SEO would not. If the answer is a list of tactics you are already paying for, you have your answer.

Do not treat this as a discipline that replaces SEO either. The pages that get cited are crawlable, fast, well structured and genuinely useful, which is the same list as it was five years ago. What changed is that vagueness is punished harder, because a machine is extracting a passage rather than a person skimming a page. For the wider picture, we have written about competing in AI search and about getting an Irish business listed in ChatGPT and AI Overviews.

Where to start this week

Run your main service and your county through Perplexity Pro Search and read the breakdown. Write down the sub-questions it produced and note which websites got cited for each.

Then take your most important service page and write down every question a customer asks you on the phone before they buy. Not the questions you wish they asked. The real ones, including the awkward one about price.

Check whether your page answers each of them in a single paragraph that would make sense to someone who had never seen the rest of the page. Most pages answer two or three and gesture at the rest.

Write the missing ones. Give each a heading in your customer’s words. Put the answer in the first sentence.

That is the work. It is not glamorous and there is no tool for it, which is exactly why it is still available to you.

The short version

Query fan-out is not a Google feature. It is how AI search works now, on Google, ChatGPT, Perplexity and Copilot, and all of them stopped looking for pages that match a question and started looking for paragraphs that answer one part of it.

But every one of those parts is answered out of search results. That is why AI visibility is not a new discipline sitting beside SEO. It is SEO, with the target moved: you now need to rank for nine narrow questions instead of one broad one, and you need each answer written so it can be lifted out and used.

You do not need new software for that. You need to be the business that gives a straight answer in public, on the record, in the words your customers actually use, while your competitors are still writing about their passion for excellence.

If you want someone to go through your site and work out which passages are close and which are doing nothing, that is the kind of thing we do every week. Tell us what you are trying to rank for and we will tell you what is standing in the way.