Fan-out queries vs keywords: what is the difference?

Published 8 November 2026. Updated 8 November 2026.

A keyword is a phrase people type into a search engine, counted month after month. A fan-out query is a search an AI assistant writes for itself while it answers one prompt, so it has no search history and often names things the assistant found seconds earlier. Ask the same prompt again and the queries can change.

Keyword research shows what people want and the words they use for it. Fan-out queries show the searches a page has to come up for before ChatGPT, Claude or Grok can cite it. Our guide to query fan-out explains how assistants split a prompt into searches. Below, we put the two side by side, using real queries recorded with Only Fan Out.

What is the difference between a keyword and a fan-out query?

People write keywords and tools count them over months. An assistant writes fan-out queries in a few seconds, for one answer, and builds some of them from pages it has just read.

	Keyword	Fan-out query
Who writes it	People typing into a search engine	ChatGPT, Claude or Grok, while answering a prompt
Where you see it	Keyword tools, Search Console, autocomplete	Inside the assistant's response data, one turn at a time
Search volume	Measured every month	Most have little or none, Semrush says
Length	A few words	61 to 90 characters was the most common length in DataForSEO's sample of 100,249 ChatGPT fan-outs
Stability	The same phrase from month to month	27% repeated across runs of the same prompt in a Surfer test, and 66% showed up once
Settings	None, just the words	Search type, recency window, location, result count or search operators
Built from	What people already know to ask	The prompt, plus pages the assistant read earlier in the same answer

What do real fan-out queries look like next to the prompt?

They are longer and more specific than the request, and two assistants given the same request write different ones. On 10 and 11 September 2026 we asked Claude and Grok to search the web and compare secondary maths tuition centres in Jurong East, Singapore, and recorded each turn with Only Fan Out.

Engine and search	Query as the assistant wrote it	Words
Claude, places	secondary math tuition centre Jurong East Singapore	7
Claude, places	math tuition Jurong East MRT Singapore	6
Claude, places	tuition centre Jurong Gateway Road Singapore	6
Claude, web	secondary school maths tuition rates Singapore 2026 per hour tuition centre	11
Grok, web	secondary math tuition centres Jurong East Singapore	7
Grok, web	best secondary mathematics tuition Jurong East 2026	7
Grok, web	Jurong East math tuition reviews fees classes	7
Grok, web	"Jurong East" "math tuition" OR "mathematics tuition" secondary OR O-level	10
Grok, web	The Math Lab Jurong East tuition	6
Grok, web	AGrader Jurong East secondary math fees class size	8
Grok, web	Zenith Education Studio Jurong East math tuition fees reviews	9
Grok, web	Matrix Math Jurong East fees class size reviews	8

The request to Claude ran to 14 words. The two assistants wrote 12 searches between them, 6 to 11 words long, 7.7 on average, and no two were the same. Claude ran 4 searches and Grok ran 8, then opened 3 pages to read in full. Claude's first search and Grok's first search differ only by "centre" against "centres". A keyword tool would fold those into one keyword, but each was a separate search with its own results. Claude also wrote "centre" in its search though the request said "centers".

Why can't a keyword tool show you fan-out queries?

Many fan-out queries do not exist until the assistant has read its first results, and the rest change from one run to the next. Our recordings and two published studies show why.

• Some are built from results. Grok's first four searches were broad. Its next four each named one centre it had found, with "fees", "class size" or "reviews" added. A ChatGPT turn we recorded on 9 September 2026, for a local services search in Singapore, did the same: its second search step put three business names in quotes. No keyword database could hold those searches in advance.
• They add words the person never typed. Of the 12 tuition searches, 2 added "2026", 5 added fees, rates, class size or reviews, and Claude added a train station and a street name. In Ahrefs' study of 66.6 million ChatGPT responses, "2026" appeared in 17.6% of fan-out queries in August 2026 and 22.9% in September.
• They change between runs. In Surfer's 1,600 runs with Gemini, 27% of fan-out queries came back across repeated runs, 66% appeared once, and 0.6% appeared every time. One recording is a sample of what an assistant searches, not a fixed list.

What settings travel with a fan-out query?

A keyword is only words. A fan-out query can carry settings that change which pages come back, and none of them shows up in a keyword tool.

Engine	Setting	What we recorded
ChatGPT	Search type	Web, business and image searches in one turn: 7 web, 7 business and 3 image across 3 search steps (9 September 2026)
ChatGPT	Recency window	A number of days on each web search: 30 on the first step, 365 on the later ones
ChatGPT	Location	Business searches carried "Singapore" in a field of its own, apart from the query words
ChatGPT	Bundled queries	One business line held three queries separated by semicolons
Claude	Result count	Its places searches asked for 8 and 6 results
Grok	Search operators	Quote marks and OR in one query

The ChatGPT format above is the one we saw in August and September 2026, and OpenAI has changed it more than once this year. The recency window suggests a search can be limited to recent pages, so a page's date may decide whether it comes back at all. We have not tested how strictly ChatGPT applies it.

Does keyword research still matter for GEO?

Yes. Fan-out queries grow out of the prompt, and prompts come from the same people whose searches keyword tools count. DataForSEO found its fan-out queries shared 54% of their words with the prompt on average.

Keyword and Search Console data still tell you which topics have demand and which words people use. Fan-out data adds what they cannot show: the narrower searches an assistant runs on the way to an answer, including the ones that name a rival or carry a year, a price or a place.

How do you use fan-out queries and keywords together?

Use keywords to choose topics and fan-out queries to see what a page on that topic has to cover.

1. Take the keywords you already track and write each one as the question a person would ask an assistant.
2. Ask each question in ChatGPT, Claude and Grok, more than once, and record the queries with Only Fan Out.
3. Group the queries by what they look for, such as a price, reviews, a place or a named rival, not by exact wording, since the wording changes between runs.
4. Check each group against your pages. A group no page answers is a gap. A query that names a rival or a year your page leaves out is an edit.
5. Use keyword and Search Console data to put the gaps in order of demand.
6. Ask the same questions again after the changes and compare.

Our guide to GEO with fan-out data sets this out as a full workflow, from recording the queries to choosing what to fix first.

Related questions