Query fan-out is the step where an AI engine rewrites your buyer's question into several narrower searches, then retrieves pages for those searches rather than for the prompt itself. In ChatGPT, the sub-queries routinely name specific vendors, add words like "official" and "pricing", and since August 2026 some are scoped to a single domain. In one captured session about AI live chat software, ChatGPT ran fast|Zendesk AI agents pricing 2026|30|zendesk.com, a search aimed at one vendor's site with a 30-day freshness preference.
If your domain never appears in a line like that, your well-matched page never gets its turn. That is the short answer to a question many B2B SaaS SEO leads keep asking: the page matches the prompt, so why isn't it retrieved? The engine never searched the prompt. This guide covers what the sub-queries contain, what changed in August, why the measurements disagree, and what to optimize instead.
What query fan-out is, and how much of it ChatGPT does
Google named the technique. At I/O on 20 May 2025, Elizabeth Reid, VP and Head of Search, wrote that "AI Mode uses our query fan-out technique, breaking down your question into subtopics" (Google). Google's developer documentation now says AI Overviews and AI Mode may issue "multiple related searches across subtopics and data sources" to build a response (Google Search Central).
ChatGPT does the same, and how much depends heavily on where you measure. Ahrefs analysed 66.6 million responses from free, logged-out GPT-5.6 Luna in August and September 2026. Only 14.4% searched the web at all, and the ones that did averaged about 1.6 fan-out queries each (Ahrefs). Nectiv, testing roughly 4,000 prompts through the API, saw something much busier: the average climbed from 2.17 searches per prompt in 2025 to 7.61 in 2026, and software prompts drew the most of any vertical, 10.7 on average (Nectiv).
Which number applies to you? For a B2B SaaS team, prompt type matters more than the headline average. In a sample of 80 real-user prompts, 18 of 23 commercial prompts triggered fan-out (78.3%), against 2 of 55 informational ones (3.6%) (arXiv 2609.23162). The sample is small and not SaaS-specific, but the direction is clear: the buyer-intent prompts you track are the ones most likely to be rewritten before anything is retrieved. If you've read our breakdown of how often ChatGPT searches the web at all, this is the other half: when it searches, it doesn't search what was asked.
So what does it search instead?
What the sub-queries add that your buyer never typed
The first addition is brand names. Suganthan Mohanadasan captured ChatGPT's searches on one ChatGPT Plus account in Dubai on 24–25 July 2026. In 21 of 27 conversations, the very first search query already named brands the user never typed (Mohanadasan). Two captures from software categories:
best AI note taking apps 2026 official pricing features Granola Notion AI Otter Fireflies Fathom Mem Limitlessofficial pricing Xero QuickBooks Zoho Books FreshBooks Wave Sage
No buyer types seven competitor names into a chat box. The model supplies the shortlist itself, then searches for it.
The second addition is a set of modifiers that pull the search toward current, first-party, commercial facts. In Ahrefs' dataset of 15 million fan-outs, "2026" appeared in 17.6% of English fan-outs in August and 22.9% in September. "Official" rose from 10.1% to 13.1%, "price" from 6.0% to 8.9%, and "reviews" by about a third, while "meaning" and "definition" each fell by about a fifth (Ahrefs). The citations follow the same pull: 84.4% of answers to factual questions cited at least one official source, and 68.3% of answers to comparison questions did.
Those queries are also invisible to keyword research. AirOps expanded 15,000 prompts into 43,233 ChatGPT fan-out queries and found that 95% had zero monthly search volume by traditional metrics (AirOps). The same study found comparison queries were split into sub-queries 38.4% of the time, the highest of any intent type, while definition queries stayed near-verbatim 51.6% of the time. The "X vs Y" prompts closest to a purchase drift furthest from their wording, and your keyword tool cannot see where they drift to.
Until August, nearly all of this ran as open web searches in the consumer app. Then a new kind of search appeared.
August 2026: ChatGPT started searching inside specific domains
On 8 August 2026, Promptwatch saw the share of ChatGPT fan-outs using the site: operator jump "from 0.37 % to 16.8 %" in the consumer interface, roughly 1 in 6 queries. On the same day, fan-outs per response rose from about 1.08 to about 1.83 (Promptwatch). Promptwatch's author, Klaas Foppen, noted that "the shift happened within a single day," pointing to a rollout on OpenAI's side. He reads the timing as a sign the scoped searches were added on top of generic ones, not swapped in, and flags that as inference.
Then the call format changed. Between 16 and 20 August, Mohanadasan saw ChatGPT's search tool call switch from JSON, where a scoped search looked like {"q":"site:intercom.com Fin AI Agent pricing 2026"}, to a compact pipe-delimited line (Mohanadasan). He reads the fields as call type, query, a recency window in days, and an optional domain, and stresses that this is his interpretation, not documented behaviour. The domain slot scopes a search to one site without the literal "site:" string in the query text. Our inference, untested by any study so far: counts that only look for "site:" may undercount domain-scoped searches made after the change.
Here are the captured lines from his live chat software session. They read like a buyer's evaluation process:
fast|Intercom Fin AI agent pricing 2026 live chat support|30|intercom.comfast|Gorgias AI Agent pricing 2026 customer support|30|gorgias.comfast|Zendesk AI agents pricing 2026|30|zendesk.comfast|best AI live chat support Intercom Gorgias Zendesk Ada Tidio Crisp reddit|365|reddit.com
ChatGPT has a shortlist in mind. It goes to each vendor's own domain for pricing with a 30-day freshness preference, then checks Reddit for opinion over a 365-day window. A vendor not on that shortlist doesn't get a |yourdomain.com probe at all.
Why the measurements disagree so much
How big is the domain-scoped slice now? It depends on who measured, on which surface, with which prompts:
| Share of ChatGPT fan-outs using site: | Who measured | Surface and scope |
|---|---|---|
| 0.14% (August) | DEJAN | OpenAI API, 196,692 fan-outs, March–August 2026 |
| 6.8% | Ahrefs | Free, logged-out GPT-5.6 Luna, 15M fan-outs, Aug–Sep 2026 |
| 16.8% (from 0.37%) | Promptwatch | Consumer ChatGPT, step change on 8 August |
| ~23% (from ~0.3%) | Peec AI, reported by Search Engine Journal | Consumer app, the day GPT-5.6 became default |
| 30.2% | MJ Cachón | Branded prompts only, GPT-5.5 via API |
| 64% | Nectiv | ~4,000 prompts, GPT-5.6 Sol via API |
That is a spread from 0.14% to 64%, and every row is accurate about something. DEJAN's API data is the sharpest counterpoint: its site: share fell from 6.43% in March to 0.14% in August, and 0 of the 371 OpenAI fan-outs it captured on 8 August contained site: (DEJAN). Its step changes track model version more than calendar date: gpt-5.2 scoped 5.88% of the time, gpt-5.4 5.32%, gpt-5.5 just 0.15%. Research-style "citation mining" prompts scoped 6.29% of the time, fixed monitoring prompts 0.44%. As Dan Petrovic puts it: "Surface, prompt wording, and prompt mix are all candidate causes of the gap, and our data does not separate them."
The defensible summary: since August, domain-scoped searches went from a rounding error to a meaningful slice in the consumer app, and how big the slice is depends on surface, model, and prompt type. Branded prompts scope heavily. In Cachón's 189 branded prompts, 55.1% of site: searches targeted the brand's own domain, and 64.4% of prompts triggered at least one own-domain site: search, which she describes as "a form of fact-checking against your own content."
Which domains get targeted? In Ahrefs' data, 75.9% of the most common site: searches pointed at a brand site. The published top targets are consumer brands and government sites, with no SaaS domain in the top 15. Nectiv's software prompts show the B2B pattern: ChatGPT "starts with non-site searches, then shifts to site searches of key vendors," often placing "official" right after the operator. Lily Ray reads the behaviour as "functioning (at least in part) as a spam filter" (Search Engine Journal): the engine goes to the source it trusts for a fact rather than to whoever ranks for it.
This is where the "well-matched page" puzzle starts to make sense.
Why a well-matched page still isn't retrieved: the shortlist comes first
If ChatGPT names brands in its first query and then probes their domains, the decisive moment comes before retrieval. Brands named in ChatGPT's own search query reached the final answer 68.9% of the time (119 brands). Brands that were fetched during the search but never named in a query reached it 2.1% of the time (515 brands), roughly 33 times less (Mohanadasan). Mohanadasan later softened his framing, and his section heading is the right takeaway: "Being in the query is the entry ticket, not the win." The data comes from one account in one location, and he calls every percentage directional.
A separate dataset points the same way from the other side. Benjamin Tannenbaum's September 2026 paper measured how well a page's content matched the prompt as a predictor of own-domain citation. On GPT, the AUC was 0.545, barely better than a coin flip at 0.5. On Gemini it was 0.641 (arXiv 2609.23162). When a brand's own domain was cited, GPT mentioned the brand in all 30 of 30 runs, against 16 of 169 (9.5%) when it wasn't. His line sums up the mechanism: "Excellent selection propensity is irrelevant if the source is never exposed." The author is founder and CEO of Aiso Boost, which sells AI-search measurement software, and the paper discloses that conflict.
Even when a page is retrieved, the odds stay thin, because retrieval is wide and citation is narrow:
- Mohanadasan: 110 of 3,554 retrieved pages were cited across 57 labelled conversations, or 3.1%.
- AirOps: of 548,534 pages ChatGPT retrieved, only 15% were cited (AirOps).
So the funnel runs like this. The model picks a shortlist, writes those names into sub-queries, probes their domains, then reads many pages and credits a handful. Page-level relevance only gets a vote at the last stage. A buyer-matched landing page on a domain that never entered the shortlist is optimizing for a stage it never reaches.
The domain slot adds one more failure mode, because it is filled from the model's memory. Mohanadasan caught ChatGPT probing profound.ai for a GEO vendor that actually runs on tryprofound.com. The scoped search returned 0 pages. A broader search later found the right domain, and Profound was still named six times in the answer (Mohanadasan). The brand survived the mention, but the scoped probe of its own pages, the one most likely to produce a citation, came back empty.
That distinction maps onto two numbers you can track. Getting onto the pre-search shortlist drives mentions. A scoped probe that lands on your pricing page drives citations. In Geoptimizer's own scan of notion.so, Notion was named in 92% of answers but cited in just 3%: consistent with a brand firmly on the shortlist whose own pages rarely get the credit.
What to optimize instead of the prompt text
The fix follows the funnel, in three parts.
Get onto the shortlist before the search starts
The shortlist is assembled from what the model already associates with your category, and from the third-party pages it reads first. Lily Ray lists G2, Clutch, Capterra, Consumer Reports, and Wirecutter as common retrieval targets for software and service comparisons, based on her prompt testing rather than a measured share. No study has yet measured how often ChatGPT runs site:g2.com, so treat any precise figure for that with suspicion.
Comparison roundups and "alternatives to" coverage are where shortlists form. In Mohanadasan's captures, an "alternatives to zendesk" prompt put Help Scout into the search itself. For a challenger brand, the job is to appear on the pages ChatGPT reads before it decides whose domain to check. That includes sources that rarely get credited: in the captured live chat session, 84 of the 221 entries in the retrieval pool were Reddit threads, and none was cited. One conversation proves little, but a source can shape what the model reads without appearing as a citation you can count.
If you publish your own comparison page to earn that spot, its fairness affects whether AI engines quote it, and our guide to fair comparison pages for AI search covers what makes one citable.
Press releases are one way to get a launch or pricing change onto the third-party pages that shape a shortlist, and we tested whether press releases improve AI visibility with 99% and zero results.
Make the scoped search succeed on your domain
Once you are on the shortlist, ChatGPT may come to your domain looking for a specific fact: pricing, plan limits, integrations, a feature claim. Ahrefs' advice is blunt and correct: "those facts need to be on your site, in plain text, on a page it can find." A practical checklist:
- Put pricing, specs, and integrations in server-rendered HTML. A scoped probe can only return a page that is indexed, crawlable, and readable. A pricing table assembled by JavaScript, or a "contact sales" wall, leaves nothing to retrieve.
- Keep commercial pages visibly current. In Mohanadasan's captures, every brand probe on commercial questions used a 30-day window. David Konitzny at Peec AI, working from roughly 250,000 chats, found 30 days was the most common freshness window and noted that "the freshness value influences retrieval, but does not strictly constrain it." About 65% of URLs retrieved under 30-day requests fell inside the window (Konitzny). How undated pricing pages fare is an open question, so a visible "last updated" date and a real monthly review are cheap insurance.
- Make your canonical domain unmistakable. The profound.ai misfire is the cautionary case. If your product name, company name, and domain differ, state the mapping consistently across your site, your profiles, and the third-party pages that describe you.
- Let the retrieval bots in. Promptwatch stresses allowing OAI-SearchBot and not blocking ChatGPT-User at the CDN. robots.txt alone won't reveal a CDN or bot-protection block, so test it with a live fetch. Our free AI crawler access checker reads robots.txt and llms.txt, then fetches the page as the retrieval bots do.
- Write sections that answer sub-queries on their own. The scoped probe lands on a page, but the engine scores passages. Our guide to passage-level retrieval and chunking covers how to shape each section so it stands alone when quoted.
Cover recurring sub-query themes, not one page per fan-out
The tempting response to a list of sub-queries is a page for each. Lily Ray warns that chasing individual long-tail fan-outs risks scaled content spam, and recommends tracking recurring core topics instead. Surfer's author makes the same case for topical authority, because fan-outs vary by user (Surfer).
The ranking data supports a measured approach. On Google AI Overviews, Surfer found a Spearman correlation of 0.77 between ranking for fan-out queries and being cited. Yet 67.82% of AI Overview citations ranked in the top 10 for neither the main query nor any fan-out, and the author notes that correlation is not causation.
So look at the sub-queries across many runs, group them into themes such as pricing, integrations, comparisons, and use cases, and give each theme one strong, clearly titled page. Google's own position is a fair counterweight: its documentation says no "special optimizations" are necessary to appear in its AI features. For Google, good SEO on the fan-out SERPs is most of the job.
That leaves the practical question: how do you see the sub-queries in the first place?
How to see your fan-outs and your site: footprint today
There are three places to look, and each shows a different slice.
ChatGPT, directly. After the format change, Mohanadasan's method still works: open ChatGPT in Chrome with DevTools on the Network tab, ask a commercial question, filter for conversation, open the largest response, and search its Response pane for fast|. One caution: the legacy search_queries metadata field that extensions read is gone from the payload, and he notes the FanoutFox extension's fan-out card will be empty on new conversations until patched.
Google Search Console. Lily Ray found thousands of site: query impressions with near-zero clicks across major brands, including one site: query with about 197,000 impressions and exactly one click. She suspects bots, and calls that speculative. To check your own, filter the Performance report by "Queries containing" site:.
Bing Webmaster Tools. The AI Performance report, in public preview since 10 February 2026, shows "grounding queries": the key phrases the AI used when retrieving content referenced in its answers, sampled across Microsoft Copilot, AI summaries in Bing, and select partner integrations (Bing Webmaster Blog).
Whichever method you use, one run is not enough. Surfer reports that only about 27% of fan-outs stay stable across repeated searches (Surfer). MJ Cachón puts it from the other side: a single run shows you roughly a quarter of what the model might ask about a brand. Any "here are your fan-outs" export is a snapshot of one draw from a distribution.
Measure the outcome, not the fan-out list
Fan-out lists are diagnostic. They tell you why a number moved. The number you manage is still the outcome: are you named, and is your domain cited, on the prompts your buyers ask? Because both the sub-queries and the answers are nondeterministic, that outcome needs repeated sampling over time, engine by engine.
That is how Geoptimizer measures. Each tracked buyer prompt runs on ChatGPT, Gemini, Claude, and Grok with web search on. Mention rate and citation rate are reported separately per engine, and headline scores use a 7-day rolling window with a confidence band rather than a single run. The trend chart marks the day an engine swapped models, so a step change like 8 August shows up as a mark on the paper: a drop has a cause instead of a guess.
Engine-by-engine matters because fan-out behaviour differs sharply. Peec AI's April 2026 count of 5 million fan-outs put Grok at 6.8 per prompt, ChatGPT at 2.1, and Perplexity at 1.4 (Peec AI). A single blended number would hide that.
One honest caveat about our own method. Geoptimizer collects answers through each provider's official API with web search enabled, and DEJAN's data suggests the API scoped searches far less than the consumer app in August 2026. Our published scoring methodology states that API-grounded answers approximate but do not exactly equal what users see in the consumer apps. For the closest read we offer, Premium and Ultra checks run your pinned prompts on each provider's newest flagship models. We'd rather you know where the measurement bends than take our word for it.
FAQ
What is query fan-out in ChatGPT?
Query fan-out is ChatGPT rewriting a user's prompt into narrower search queries before it retrieves pages. The sub-queries often name brands the user never typed, add terms like "official", "price" and the current year, and since August 2026 can be scoped to a single domain.
Does ChatGPT use the site: operator?
Yes. In the consumer app, site: searches jumped in August 2026, but the share depends on who measured: from 0.14% (DEJAN, API) and 6.8% (Ahrefs, free logged-out app) to 16.8% (Promptwatch, consumer app) and 64% (Nectiv, API, GPT-5.6 Sol).
Can I find ChatGPT's fan-out queries in a keyword research tool?
Mostly no. AirOps found 95% of ChatGPT fan-out queries had zero monthly search volume. Better sources are DevTools captures, the "Queries containing site:" filter in Google Search Console, and grounding queries in Bing Webmaster Tools.
Should I create a page for every fan-out query?
No. Only about 27% of fan-outs stay stable across repeated searches in Surfer's data, and chasing each one risks scaled content spam. Group recurring sub-queries into themes such as pricing, integrations, and comparisons, and give each theme one strong page.
Why does ChatGPT mention my brand but not cite my site?
Mentions and citations come from different stages of the fan-out. Being on the pre-search shortlist earns the mention. A citation needs a search, often scoped to your domain, that finds a crawlable, current page with the fact the answer needs. A high mention rate with a low citation rate points at that second stage.
The short version
ChatGPT doesn't search your buyer's prompt. It picks a shortlist, rewrites the question into sub-queries around it, and increasingly probes specific domains for official, current facts. So earn a place on the shortlist through the third-party pages the engine reads first, make the scoped search land on crawlable, dated, plain-text pages, and measure mentions and citations by engine over enough runs to trust the trend.
To see where you stand today, run the free AI Visibility Check. It asks Claude Haiku 5.5 three buyer-intent prompts with web search on, and shows whether it names and cites you, and which domains it cited instead. It needs no signup.