What is query fan-out, in plain terms?
So when a buyer asks ChatGPT about our kind of business, it is not searching for what they typed?
Mostly, no. When an AI answer engine decides it needs the web, it rewrites the question into a handful of shorter searches, runs them, reads what comes back and writes one answer from the pile. That rewriting step is query fan-out. The buyer typed one sentence. The engine searched for three or four different things, and those things decide which pages it ever sees.
Google says this openly. When it launched AI Mode it described a query fan-out technique that issues "multiple related searches concurrently across subtopics and multiple data sources" (Google, 5 March 2025). At I/O it repeated that AI Mode breaks a question into subtopics and issues "a multitude of queries simultaneously", and that Deep Search can issue hundreds (Google, 20 May 2025). Its own optimization guide, last updated 10 July 2026, now defines a fan-out as a set of concurrent, related queries generated by the model to request more information (Google AI optimization guide).
ChatGPT does the same thing, it just does not publish the queries. We capture them. Our engine runs a fixed set of buyer prompts in logged-out temporary chats and records every hidden search the model fires.
The numbers differ because the setups differ. Seer forced Gemini 3 to search on every prompt through the API, so its count shows how wide a fan-out can get, not how often the consumer app searches on its own (Seer Interactive, 21 November 2025). Our ChatGPT figure comes from the real chat interface, logged out, with buyer questions. Both point the same way: the searches are many, short and mostly invisible to keyword tools.
What the evidence says, and how sure we are
Fan-out is one of the few AI search ideas with a primary source behind it. What it means for your pages is less settled, so we grade each claim before we act on it.
| Claim | Who says so, and when | How sure we are |
|---|---|---|
| Google AI Mode and AI Overviews run several related searches per question. | Google, AI features and your website, last updated 10 December 2025 | Works Primary source |
| Making a separate page for every fan-out query, mainly to manipulate rankings or AI answers, is scaled content abuse. | Google AI optimization guide, last updated 10 July 2026 | Works A rule, not a theory |
| Many pages ChatGPT cites do not rank on Google or Bing for the fan-out queries it ran. | Grow and Convert, 100 buying prompts, updated 30 March 2026: 27 percent of cited sources ranked on Google for the fan-out queries, 23 percent on Bing, roughly 40 percent in the first 10 pages of either | Depends Small sample from one agency, but measured |
| Most AI Overview citations come from outside the organic top 10. | Conductor, about 168 million citations, updated 25 September 2026: 57 percent outside the top 10 | Depends Vendor data, correlation. Fan-out is one plausible reason, not proven |
| For long-tail questions, roughly 80 percent of AI-cited URLs do not rank in Google at all for the original query. | Ahrefs, about 15,000 long-tail queries, 11 August 2025 | Depends Correlation, mid 2025 data |
Read together, this says one practical thing. Ranking for the sentence the buyer typed is not enough, because the engine never searched for that sentence. You want your page to be a strong answer to the shorter searches it actually runs.
No study shows that copying a fan-out query into a title causes a citation. The overlap studies are correlations, and the engines change their behaviour every few months. We treat fan-out words as the best available description of what the engine looks for, and we measure every change we make.
The rule that keeps this safe: one page per intent
If the engine searches for ten variations, should we not just make ten pages?
No, and this is where most fan-out advice goes wrong. Google's guide says plainly that creating separate content for every query variation or fan-out query, done primarily to manipulate rankings or AI responses, is scaled content abuse and does not work long term (Google, 10 July 2026). The spam policies call scaled content abuse many pages generated mainly to manipulate rankings, "no matter how" they are created (Google spam policies, last updated 28 August 2026).
Changed in 2026: Google added the AI optimization guide on 15 May 2026. It names separate pages per fan-out query as scaled content abuse, and the same month Google clarified that its spam policies apply to generative AI answers. The scaled content and doorway rules underneath are older. The guide simply applies them to fan-out, so nobody can call it a grey zone now.
So our rule is simple. One page per buying intent. A buying intent is the job the buyer is trying to get done, like "find a dentist who can fix a cracked tooth today". The ten ways of phrasing it are not ten intents. They are vocabulary for one page: its title, its first sentence and its section headings.
Ask whether the same page would satisfy both buyers. "Emergency dentist Bangkok" and "dentist open Sunday Bangkok" can share a page if your emergency service runs on Sundays. "Dental implants cost" and "emergency dentist" cannot: a different buyer, a different decision, a different page.
How we harvest the hidden searches
Can we not just guess what it searches for?
You can guess, and you will guess wrong in a particular way: you will guess in your own words. The engine uses the buyer's words, sometimes in another language, and often in a shorter, flatter form than anyone would write. That is why we harvest instead of brainstorm.
- Write the prompt file. Start with ten unbranded questions a real buyer would ask, one per line, tagged with the language, and grow to 15 to 50 once you measure. We take them from Search Console question queries, from real enquiries and from sales calls. Never include your brand name: you want to know what happens when the buyer does not know you yet.
- Run them logged out. A fresh, temporary, logged-out chat for every prompt. In our October 2026 harvest a personalised, logged-in chat put the brand name straight into the hidden query. That tells you about the account, not the market.
- Capture three things per prompt. The hidden queries, the pages the engine retrieved, and the pages it actually cited. Retrieved but not cited is a different problem from never retrieved at all.
- Record the conditions. Date, model, plan, language and location. A harvest describes one population: logged-out users of one model in one locale on one day. Say so in the notes.
- Repeat every 28 days. Our engine reruns the same file monthly. We add prompts over time, but we do not rewrite the old ones, or the comparison breaks.
Where the browser route is blocked, we fall back to a data provider's fan-out field, inside an existing research budget. Either way the output is the same file: prompt, language, hidden queries, retrieved, cited, and whether your domain appeared.
A model asked to "list the searches ChatGPT would run" invents plausible ones. A harvest records the ones it did run. Only the second is evidence.
Mapping the words onto the page
Once the queries are in, we cluster them by intent and write one line per intent: the verbatim queries with their language, the page that should answer it, whether that page was cited, retrieved or absent, and the edit. Then the words go onto the page, cheapest surface first.
| Surface | Our rule |
|---|---|
| Title tag | The commercial phrase the engine used, 60 characters at most, brand last. |
| URL slug | Natural words. Change it only if the current slug misses the query badly, and then with a 301 redirect, an updated sitemap and language links in the same release. |
| H1 | The human version of the phrase. A headline a person would say out loud, not a keyword string. |
| First sentence | Answers the query directly and carries the phrase once, in the page's own voice. |
| Section headings | The sub-questions written as questions, each answered in the first 40 to 60 words beneath it. |
| Visible Q and A | Real sub-queries answered on the page for readers. FAQ rich results stopped showing on 7 May 2026 (Google documentation updates), so this is for people, not for markup. |
| Meta description | Delivers what the title promises. It feeds the snippet, which is what both people and AI features read first. |
| Internal links | Two to five links from related pages, with anchors that use the phrase. See our guide to internal linking. |
Keep the order. A title and a first sentence cost minutes. A new page costs days, needs links and needs time to be indexed. Google also says there is no need to break content into tiny chunks for AI features, so do not rebuild a good page into fragments to chase a query (Google, 10 July 2026).
Google's AI optimization guide (last updated 10 July 2026) says AI systems understand synonyms, so "you don't have to worry that you don't have enough 'long-tail' keywords or haven't captured every variation of how someone might seek content like yours." That is right, and it is why we never write a page per phrasing. We use the harvest for something narrower: to find pages that describe themselves in words the engine does not search with, and fix that wording on the one page that already owns the intent. Whether a wording change moves citations is a claim we test per change, not a law.
Five kinds of gap, and what to do with each
Every intent in the map gets a gap type. The type decides the work, and it stops us writing new pages when a sentence would do.
| Gap | What it looks like | What we do |
|---|---|---|
| A. Wording gap | You have the right page, but its title and headings use your words, not the buyer's. | Edit the surfaces above. Cheapest fix, always first. |
| B. Missing page | A real buying intent with no page that answers it. | One new page, after checking that no existing page already competes for it. |
| C. Evidence intent | Health, safety, science or legal questions. The engine goes to authorities and journals. | Wording fixes on existing pages only. A new page only if your product directly answers the question. Promise no citation. |
| D. Third-party intent | "Best X", reviews, comparisons, directories. The engine cites lists and review sites. | Off-site work: profiles, reviews, honest inclusion on real lists. Make sure your own domain is indexed and plain to find. |
| E. Already cited | Your page is in the answer. | Leave it alone. Add internal links to it. |
We sell in two languages. Does the AI search in the language our buyer types in?
Do not assume it does. An assistant may search in a different language from the question, so check the language of every hidden query in your own harvest. Then split the work by intent, not by market.
- Where the hidden queries for shop, price and delivery questions stay in the buyer's language, those intents need a native page. For a shop, work these first.
- Where they switch to English, check that the page answering that intent has an English version carrying those English words.
- Write each language from its own queries. Do not translate the English title. A native reader checks the wording before it ships.
One strange query in one run is noise. We only act on a phrasing that shows up in two or more prompts or two or more languages.
A worked example
This is the shape of a real map, with an illustrative business so no client data is involved. Say a Bangkok company rents and services office coffee machines. One of its buyer prompts is: "We are a 40 person office in Sathorn, what is the best way to get proper coffee without buying a machine?"
A harvest of that prompt, logged out, might return hidden searches like "office coffee machine rental Bangkok", "coffee machine lease vs buy office" and "office coffee service Sathorn". These are illustrative, written to show the pattern, not copied from a client file.
| Hidden search | Page that should answer | Status | Gap | Edit |
|---|---|---|---|---|
| office coffee machine rental Bangkok | Rental service page | Retrieved, not cited | A | Title becomes "Office coffee machine rental in Bangkok"; first sentence states the monthly rental model plainly. |
| coffee machine lease vs buy office | Rental service page | Absent | A | New H2: "Is leasing an office coffee machine cheaper than buying one?" answered in 50 words with the real cost inputs. |
| office coffee service Sathorn | Rental service page | Absent | A | Service area sentence names the districts actually covered. No district pages. |
Three hidden searches, one intent, one page, three edits, zero new URLs. That is a typical result. In our experience most of the value of a harvest is in gap type A, because businesses describe themselves in their own vocabulary and buyers do not use it.
We log one decision per shipped page: the baseline (how many prompts cited or retrieved the page), the metric we will read, and a check date at the second harvest after shipping, roughly 56 days later. One harvest is a sample. Two harvests start to be a trend.
What is hype, and what is a waste of time
| Tactic you will hear about | Verdict | Why |
|---|---|---|
| A page for every fan-out query or phrasing | Against the rules | Google names it as scaled content abuse in its July 2026 guide. |
| Stuffing all the hidden queries into one paragraph | Hype | Reads badly, helps nobody, and the engine reads the page like a person would. One natural use per phrase. |
| "Chunking" pages into tiny blocks for AI | Skip it | Google says there is no requirement to break content into tiny pieces. Clear sections answer first; that is enough. |
| Asking a model to invent the fan-out list | Hype | You get the model's guess, not the engine's behaviour. Harvest instead. |
| Ranking on Bing to get into ChatGPT | Depends | Only 23 percent of ChatGPT's cited sources ranked on Bing for its fan-out queries in Grow and Convert's 2026 test. Bing matters for Copilot; it is not a ChatGPT proxy. |
| Buying a fan-out tracking dashboard before fixing pages | Skip it | A monthly harvest of 20 to 40 prompts answers the question for a fraction of the cost. Buy daily tracking only when you have a retainer that acts on it daily. |
| Harvesting monthly and editing the existing pages that match | Works | Cheap, reversible, measurable, and it uses the words the engine actually searched for. |
Fan-out words help an engine find and match your page. They do not make a weak page worth citing. If the page has nothing original, no clear facts and no evidence, the wording fix just gets a weak page retrieved more often. Our guide to pages AI answers can cite covers the other half.
Mistakes we see, and a checklist to work from
Mistakes
- Running prompts in a logged-in, personalised chat, then celebrating that the brand appears in the queries.
- Prompts that name the brand. That measures recall of your name, not discovery.
- Rewriting the prompt file every month, so no two harvests can be compared.
- Translating the English title into Thai instead of writing from Thai queries.
- Treating a single harvest as a verdict on a page.
- Chasing a query that only appeared once.
Checklist
- Prompts. 8 to 40 unbranded buyer questions per language, saved in one file you only add to.
- Harvest. Logged out, temporary chat, conditions recorded, every 28 days.
- Map. One row per intent: queries, page, cited or retrieved or absent, gap type, edit.
- Edit. Title, first sentence, headings first. New page only for gap B, after a duplicate check.
- Link. Two to five descriptive internal links to each edited page.
- Record. Baseline, metric and a check date at the second harvest after shipping.
Fan-out is step six in our Get Found track. It sits after the basics on purpose: none of this matters if the page is not indexed (get indexed) or if the engine cannot tell who you are (structured data and entity). Next, mine your long-tail searches, which applies the same thinking to the queries already in your own Search Console. If you want this run on your site, our AI visibility diagnostic includes a harvest.
Sources and dates
Every external fact in this guide was read on 20 October 2026. Each source shows its publisher and the date it was published or last updated, and we flag what changed in 2026. Studies are their authors' measurements, usually correlations, not promises of results. Our own numbers come from our own Search Console and tools and say so where they appear.
- AI Features and Your WebsiteGoogle Search Central · Last updated 10 December 2025 · Read 20 October 2026.
- Google's Guide to Optimizing for Generative AI Features on Google SearchGoogle Search Central · Last updated 10 July 2026 · Read 20 October 2026. Changed in 2026: Guide added 2026-05-15 (changelog); llms.txt clarification note added 2026-06-15; page last updated 2026-07-10.
- AI Mode in Google Search: Updates from Google I/O 2025Google (The Keyword) · Published 20 May 2025 · Read 20 October 2026.
- Expanding AI Overviews and introducing AI ModeGoogle (The Keyword) · Published 5 March 2025 · Read 20 October 2026.
- Latest Google Search Documentation Updates (FAQ rich result entries, May and June 2026)Google Search Central · Undated page · Read 20 October 2026. Changed in 2026: No standalone blog post found; announced only in the documentation changelog.
- Spam policies for Google web searchGoogle Search Central · Last updated 28 August 2026 · Read 20 October 2026. Changed in 2026: Site reputation policy section updated 2026-08-28 for the EEA; 2026-04-13 back button hijacking added to malicious practices; 2026-05-15 clarification that spam policies apply to generative AI responses.
- Growth Engine scoring, indexing and probe rulesDLVX (Growth Engine operating notes) · Our own measurement, data to 19 October 2026 · Read 20 October 2026.
- ChatGPT fan-out harvests, logged-out temporary chatsDLVX (Growth Engine harvest) · Our own measurement, data to 11 October 2026 · Read 20 October 2026.
- Only 12% of AI Cited URLs Rank in Google's Top 10 for the Original PromptAhrefs · Published 11 August 2025 · Read 20 October 2026.
- AI Overviews and Organic Rankings: 57% of Citations Come From Outside the Organic Top 10Conductor · Last updated 25 September 2026 · Read 20 October 2026.
- Only 40% of ChatGPT Sources Come from Google & Bing for Known Fan-Out QueriesGrow and Convert · Last updated 30 March 2026 · Read 20 October 2026.
- Initial Research: Gemini 3 Query Fan-OutsSeer Interactive · Published 21 November 2025 · Read 20 October 2026.
Want this run on your own site? Start with the AI visibility diagnostic, or see the whole Get Found track.