Chapter three of ten
An engine breaks one prompt down into several searches. Those searches show you what to strengthen on your site and where to earn mentions off it.
Buyers ask What does the engine look up when someone asks about my category?
By Maeva Cifuentes · Updated October 2026
Best ATS for a healthcare company
What the engine looks up
The short answer
Query fan-out is the set of searches an engine runs during retrieval to answer one prompt. The exact wording changes on every run. The themes repeat. Read the themes, strengthen the pages you already have, earn mentions where the engine looks, then build something new only when a recurring theme has no home.
Chapter 2 covered retrieval, the live search an engine runs before it answers. But, behind the scenes, when LLMs retrieve information, they don't just rewrite your prompt. The engine rewrites the buyer's prompt into several shorter searches (that look like long-form keywordese), retrieves pages for each one, then writes a single answer from the lot.
Google named the technique. Its AI Mode announcement describes "issuing multiple related searches concurrently across subtopics". OpenAI describes the same behavior in its help center, where ChatGPT search "typically rewrites your query into one or more targeted queries".
Google's material describes many searches fired at once. OpenAI's researchers describe a model that searches, reads the results, then decides what to search next. The model learned how to search through training. It was rewarded for good research results, so it worked out on its own when to search, what to search next and when to stop. Nobody, OpenAI included, has a fixed rulebook of searches you could optimize against.
Your buyers will type one prompt in, the LLM decides what to search for to find the information required, and those searches are what we want to optimize for.
Use Google Search Console or DataForSEO to get fan-out queries.
We ran one B2B prompt through ChatGPT five times on October 9, 2026 and captured every search it ran. Click through the runs.
The prompt
Searches ChatGPT ran
Collected through the DataForSEO LLM Scraper, which reads the ChatGPT consumer interface. United States, English, web search forced on. Five runs of one prompt is a small sample, shown to illustrate the pattern.
searches per run on average, with one category search first in every run
12 searches across 5 runsruns cited a third-party "best ATS for healthcare" list, next to vendor pages and help docs
One list was cited in every runpages retrieved against pages cited across the five runs
About one page in three made it into an answerEvery run opened with a category search. The next search bundled the Workday integration with healthcare compliance, usually with vendor names written in. iCIMS was recommended in all five answers.
None of the 12 search strings repeated word for word. The vendors written into the searches rotated between iCIMS, Greenhouse, Lever, SmartRecruiters and UKG. Run 5 ran four searches where the others ran two.
"Credential" appeared in 4 of 12 searches, "official" in 5 and "2026" in 2. None of those words were in the prompt. The engine decided that compliance in healthcare hiring means credential tracking.
Zero searches asked for pricing, reviews or comparisons, so the engine spent its searches on fit and requirements. It read compliance as credential tracking, which matches what healthcare hiring teams need.
This matches a pattern we often see, where ChatGPT runs a category search first and then searches vendors by name.
One page showed up in all five runs. It was a "best ATS for healthcare" list published on the blog of an ATS company. ChatGPT cited that list every time as a source for its answer, yet it recommended other vendors every time and left the company that wrote the list off its shortlist. The company did the work of getting its page read, and its competitors got the recommendation. That is the gap between being cited and being mentioned from chapter 2.
Categories behave differently. In a second software category we tested the same day, ChatGPT averaged 4.2 searches per run, looked up one vendor per search and cited zero third-party lists. Run your own category before you assume a pattern.
In these runs, ChatGPT used "best ATS" lists on other websites to pick which vendors to name. It used the vendors' own websites to check details like integrations and credential tracking. Getting recommended takes having your own pages like this while also being mentioned on third-party lists.
Run your own category prompts several times. Log the searches, the cited pages and the brands named, then look for what repeats.
Across the patents and the large studies, the searches an engine writes fall into six kinds.
The prompt restated as a search, often with "best", the industry and the year added.
One search per brand the engine is considering, paired with the buyer's requirements.
Pricing, integrations, security, support and other facts that qualify or rule out a vendor.
Reviews, comparisons and "alternatives" searches that pull in third-party voices.
Location, earlier turns in the chat and saved memories can all be written into the search.
Thinking models read results and then search again, which adds narrower searches late in the process.
How many searches per prompt
All four bars share one scale from 0 to 12 searches. The first three come from Peec AI, about 5 million fan-outs in April 2026. The Gemini figure comes from Seer Interactive, 501 prompts with search forced on, which makes it a different method. ChatGPT's thinking models averaged 7.3 to 10.5 searches per prompt in a 50-prompt Writesonic test from April 2026.
of fan-out searches have zero monthly search volume in keyword tools
Seer, 501 prompts. AirOps, 15,000 promptsof fan-out searches run seven words or longer
DEJAN, 365,920 fan-outssearches appeared in all 13 runs, out of 11,029 unique searches
Seer, 100 prompts run 13 times eachKeyword tools miss nearly all of this demand. A team that plans from search volume alone is planning from a small slice of what engines look up, and missing a lot of opportunity.
Once you have collected the searches an engine ran for your prompts, sort each one into the six kinds above, such as a category search or a deal breaker. Then look at what repeats within each kind, for example "Workday integration" coming up again and again under deal breakers. Those repeated topics are the themes your pages and outreach need to cover.
Production systems can differ, which makes these a guide to intent more than a description of today's product.
| What it describes | In plain words | |
|---|---|---|
| Google, Search with stateful chat (application, 2024) | The model writes synthetic queries, including rewritten and drill-down versions. It carries the user's state across turns. | The engine remembers what the buyer asked earlier in the chat, so the same question can lead to different searches for different people. |
| Google, Generative summaries for search results (granted 2023) | The query is expanded with related, recent and implied queries. Each part of the answer is checked against a source before a link is attached. | Google writes the answer, then links a page next to each statement that page confirms. Your page gets linked for backing up a specific fact. |
| Google, Thematic search (granted 2024) | Top-ranking pages are summarized and grouped into themes. A theme can become a narrower new query. | The engine reads the top-ranking pages and uses what they talk about to decide what to search next. |
| Google, Generating query variants (granted 2023, filed 2018) | A trained model produces eight variant types, including follow-up, more specific, more general and clarification. | The engine can reword one question many ways, such as narrower, broader or as a follow-up, and search each version. |
| Microsoft, Deep search using large language models (application, 2025) | The model works out intents first, writes alternative queries for each intent, then scores results. | The engine first works out what the person is trying to do, then searches for that. |
| OpenAI | We found zero OpenAI patents on search query generation. Its published research and help center describe a model that rewrites prompts and learns its own search strategy. | OpenAI keeps its search rules private. The model learned how to search during training. |
Readings rest on each patent's abstract and description. Sources are listed at the end of the chapter.
We collect ours through DataForSEO. Several other routes exist, each with a trade-off.
| What you get | Engines | Good for | |
|---|---|---|---|
| DataForSEO LLM Scraper | The searches from the consumer interface, plus every page retrieved and every page cited | ChatGPT | Bulk runs that match what buyers see, at a fraction of a cent per run |
| Tracking platforms | Fan-outs on the prompts you already track. Scrunch shows them for ChatGPT, and Ahrefs Brand Radar shows them for ChatGPT and Perplexity. Profound and Semrush Enterprise also offer them. | Varies by tool | Ongoing monitoring without extra setup |
| Model APIs | The searches a model ran, returned in the API response from Google, OpenAI or Anthropic | Gemini, GPT models, Claude | Controlled tests, with a caution that API results differ from the consumer apps |
| Browser extensions | The searches from your own logged-in ChatGPT session | ChatGPT | Quick spot checks, one prompt at a time |
| Simulators such as Qforia | Predicted fan-outs written by a model | Google surfaces | Idea generation where observed data is missing |
One test of 3,645 conversations found zero shared fan-outs between the GPT-5.2 API and ChatGPT.com. The test came from Genezio, which sells interface tracking.
Single runs mislead. Seer advises trusting themes, along with searches that appear in 80% or more of runs.
DEJAN caught Gemini 3 inventing fan-out queries when asked to list them. Take searches from the raw response.
In DataForSEO's own study, 47.5% of 100,000 ChatGPT prompts triggered a fan-out. The rest were answered from memory.
The goal is to win a category of prompts. The prompts themselves are only the surface. What you can work on is everything the engine looks up to answer them, which feeds your content plan and your off-page plan together.
Please don't feel tempted to create one page per fan-out query. We're not here to play topic whack-a-mole, and that's an approach that will hurt your search performance. Google's guidance treats separate pages for each query variation as scaled content abuse. AirOps also found that covering every sub-topic on one page added only 4.6 points to citation rate.
Run each category prompt several times. Keep every search, cited page and brand named.
Group the fan-outs into themes. Keep the themes that recur across runs.
Decide where the answer to each theme should live, on your site or off it.
Strengthen first. Build new only when a recurring theme has no home.
Try it
Pick the kind of theme you found. The examples come from the ATS runs above.
Rankings and keyword volume are still two of the most important factors. Surfer found a 0.77 correlation between the number of fan-out searches a page ranks for and its citation in AI Overviews, across 10,000 keywords. A page that ranks for the category search plus several of its fan-outs is in a strong position.
In the ATS runs, the engine's first search pulled up third-party "best ATS" lists. In four of the five runs, it then took vendor names from what it found and searched for those vendors by name. A brand left off those lists has a much harder time getting into the vendor searches that follow. Getting onto them is the job of PR, reviews, partnerships and third-party lists.
Most third-party listicles are pay-to-play, or expect to swap placements with competitors. Press releases and traditional PR will not have the same impact that you think.
Work on the business case in this guide to get your executives bought in on creating and doing outreach to listicles. You can also work on your product and brand pages for support.
Surfer and Marie Haynes favor broad topic coverage, built as hubs and clusters, that answers many fan-outs. AirOps and Kevin Indig favor focused pages of 500 to 2,000 words. We read it as breadth across the site with focus on each page.
Seer prioritizes fan-outs that recur and also carry search volume. Kevin Indig argues for targeting zero-volume fan-outs when they fit the buyer.
DEJAN found Reddit at 19.5% of OpenAI's retrieved pages and 0.93% of its citations. The same study found Google cites Reddit far more readily than OpenAI does, so the picture changes by engine.
Google says its AI features need zero extra optimization and that no outside tool sees its internal fan-outs.
Choose your next stop
We use analytics cookies to see which pages help people. They load only if you accept. Cookie policy