Query fan-out1.2M queries · 111 sites · ChatGPT search logs
AI agents are already searching Google for your pages.
We read 1.2M Google Search Console queries from 111 sites, looking for searches no human would type: chatbot prompts pasted into Google, search operators, templates with unfilled placeholders, the same keyword bag in two different orders. 79% of sites had them. Almost none of them ever click.
Original research by the Outrank team · Published October 1, 2026 · Updated October 25, 2026
79%
Of sites get machine-written queries
88 of 111 sites
95–98%
Of them never get a click
Human queries: 74%
88
Machine queries per site
Median, among sites that have them
−58%
Clicks on question queries
Versus other top-3 queries, same site
Key statistics
6 findings you can quote
Each line stands on its own. Copy it with its source, or link straight to it.
79% of websites studied already received machine-written Google queries, such as pasted chatbot prompts, search operators and unfilled templates.
1.2M Google Search Console queries from 111 sites.
95–98% of machine-written Google queries never get a click, against 74% of human queries.
1.2M Google Search Console queries from 111 sites.
Websites that received machine-written Google queries saw a median of 88 of them in Search Console.
88 sites with at least one machine-written query.
Question queries ranking in Google’s top 3 got 58% fewer clicks than other top-3 queries on the same website (median site studied).
Median site, 39 sites with enough top-3 queries of both kinds.
60 of the 71 web searches ChatGPT ran to answer 48 questions were site: searches on a single brand’s domain (85%).
48 questions sent to ChatGPT (GPT-6.1 Sol with web search) on October 20, 2026, every search query logged.
On 67% of websites studied, question queries in Google’s top 3 got a lower click-through rate than other top-3 queries.
39 sites with enough top-3 queries of both kinds.
Exhibit A
These are real Google searches that showed our customers’ pages
A selection of machine-shaped queries from the dataset, with their impressions, clicks and average position. Filter by type.
- PromptA full chatbot prompt, typed into Google word for word
›write a love poem about someone longing for a sandwich. make it a sonnet, and give it a tone that straddles a serious love poem and whimsy
- impr
- 10,017
- clicks
- 157
- pos
- 10.6
- TemplateUnfilled {placeholders}: a script ran this, not a person
›"dr. {name}" "{address}" "{address}" education residency fellowship board certification insurance hospital affiliations languages
- impr
- 2,583
- clicks
- 0
- pos
- 6.4
- Fan-outSame 13 keywords, two orders, both at #3, zero clicks
›commercial property operator residential developments management fee percentage rent success fee asset light
- impr
- 7,147
- clicks
- 0
- pos
- 3.0
- Fan-outThe reshuffled twin of the query above
›commercial property operator residential developments asset light management fee percentage rent success fee
- impr
- 6,867
- clicks
- 0
- pos
- 2.9
- OperatorA line of Python in quotes: a coding agent looking up docs
›"from langgraph.graph import stategraph, end"
- impr
- 6,085
- clicks
- 0
- pos
- 9.0
- PromptA textbook LLM demo prompt, sent to Google instead
›write a casual email to my colleague jamie letting her know i can’t make the 2pm meeting and asking to reschedule. make the body of the email 3 sentences
- impr
- 3,047
- clicks
- 657
- pos
- 4.0
- AI-style questionPhrased for an assistant, answered by an AI box, clicked by nobody
›are there any firms that specialize in writing standard operating procedures for new startups?
- impr
- 5,822
- clicks
- 0
- pos
- 5.7
- Fan-outYear stacking is a classic fan-out signature
›top us data center colocation providers 2025 2026 list
- impr
- 5,059
- clicks
- 0
- pos
- 9.2
- OperatorExclusion operators are almost never typed by people
›best books learn french beginners -site:blog.librarything.com
- impr
- 1,916
- clicks
- 0
- pos
- 5.6
- AI-style questionGenerated-sounding phrasing, the kind agents use to research a list
›are there any underrated tools for working with ahrefs alternative that seos love?
- impr
- 2,086
- clicks
- 1
- pos
- 38.2
- OperatorOur own site, audited by someone’s bot
›site:www.outrank.so
- impr
- 2,357
- clicks
- 1
- pos
- 1.1
- Fan-outNo verbs, no stopwords, nine concepts in a row
›academic tool product hunt launch timing visibility optimal day
- impr
- 5,169
- clicks
- 0
- pos
- 6.1
Source: Google Search Console via Outrank. Queries that could identify a customer were left out.
Look at the template with the empty {name} and {address}: a script built that query and forgot to fill it in, then ran it thousands of times. Look at the commercial property query that appears twice, the same thirteen keywords in a different order, both ranking third, neither clicked. That is what an AI agent researching a question looks like from the other side: it fans one question out into many searches and reads the results itself.
And the sonnet. Somebody, or something, asked Google for a love poem about a sandwich in the exact words they would give ChatGPT. Search is becoming a prompt box, and Search Console is where you can watch it happen.
Exhibit B · the other side
We logged ChatGPT’s searches. 85% were aimed at one brand’s own site.
Search Console shows the queries arriving. To see them leave, we sent 48 questions to ChatGPT with web search on and recorded every query it typed. Four of them, word for word.
User asked
“best CRM for small businesses”
- search 1
best CRM small business 2026 Zoho Bigin HubSpot Pipedrive pricing official - search 2
site:zoho.com/crm zohocrm pricing USD standard 14 free three users
- search 1
User asked
“best email marketing tools”
- search 1
best email marketing software Mailchimp Brevo Kit Klaviyo ActiveCampaign MailerLite pricing official - search 2
site:kit.com pricing newsletter plan 10000 subscribers
- search 1
User asked
“best noise cancelling headphones”
- search 1
best noise cancelling headphones 2026 Sony WH-1000XM6 Bose QuietComfort Ultra 2nd generation - search 2
site:sony.com WH-1000XM6 $449.99
- search 1
User asked
“best VPN for privacy”
- search 1
site:mullvad.net audit 2025 2026 no logging cash account - search 2
site:mullvad.net pricing 5 euro month no email
- search 1
Source: Outrank ChatGPT deep dive, October 20, 2026. GPT-6.1 Sol through the OpenAI API, with web search.
These are the same fingerprints as Exhibit A: brand names, prices, years and constraints stacked into one string, and site: operators. Across 48 answers, ChatGPT ran 71 searches and 60 of them targeted a single domain. Then it opened about 36 pages per answer.
Look at the first search for each question. On 9 of 16 “best X” and “alternatives” questions, it already names the brands before it has read anything. The shortlist comes from what the model knows; the search is there to check prices and facts on each brand’s own pages. If your brand is not in that first query, no amount of on-page work gets you into the answer. What gets you there is being mentioned, linked and covered before the question is asked. The AI citation study shows which pages it ends up citing.
Machines do not click
95–98% of machine-shaped queries never produce a click
Share of queries in each group that got zero clicks over 16 months, across all sites. Count shows how many sites had at least one.
Search operators & quoted strings: 98% · 84 sites · CTR 0.90% · average position 12.0 Hover a bar for details.
| Search operators & quoted strings | 98% | 84 |
|---|---|---|
| Long conversational questions | 96% | 81 |
| Chatbot-style prompts | 95% | 78 |
| Normal human queries | 74% | 111 |
| Keyword-stuffed strings | 46% | 75 |
Source: 111 sites, 1.2M queries.
When an agent reads your page, it does not click a blue link the way a person does, so Search Console records an impression and nothing else. Keyword-stuffed strings are the exception that proves the rule: many come from non-English searchers and tool sites where a human did type a list of words, and they rank high (average position 4.6) and get clicked.
The prompt tail
The longer the query, the higher you rank and the less anyone clicks
Search Console metrics by query length, all queries pooled.
| Query length | Share of impressions | Click-through rate | Avg position | Zero-click queries |
|---|---|---|---|---|
| 1–2 words16% of queries | 37.7% | 5.25% | 6.9 | 78% |
| 3–4 words45.7% of queries | 52.9% | 6.61% | 8 | 72% |
| 5–6 words23.3% of queries | 7.9% | 8.39% | 8.3 | 74% |
| 7–9 words9.9% of queries | 1.1% | 5.44% | 9.2 | 83% |
| 10–14 words3.3% of queries | 0.3% | 3.15% | 5.6 | 78% |
| 15+ words1.7% of queries | 0.2% | 3.75% | 4.6 | 65% |
Source: 111 sites. Pooled, so the largest sites weigh more.
Queries of 15 words or more ranked at an average position of 4.6, better than any other group, yet earned a click-through rate of 3.75%, less than half that of five- and six-word queries. Long, specific queries have few competing pages, so a focused article ranks easily. Then the answer box takes the click. That is the bargain of the prompt tail: easy visibility, little traffic, and a direct line into the answers assistants give.
The question tax
At the same top-3 position, question queries earned 58% fewer clicks
For 39 sites, the click-through rate of “how”, “what”, “why” and other question queries ranking top 3, compared with the same site’s other top-3 queries.
Source: 39 sites with enough volume in both groups, brand queries excluded.
On 26 of 39 sites, questions converted worse than other queries at the same rank, and the median site’s questions earned 0.42× the click-through. Questions are exactly what AI Overviews and AI Mode answer, and what agents fan out into. The content format study shows the same pattern at page level.
Our take
Stop writing for the keyword. Write for the prompt, because the thing reading your page is increasingly not a person.
For twenty years, SEO meant guessing the two or three words a person would type. The queries in this study say the searcher is changing: longer, stranger, more specific, and more and more often a machine working for a person who will never see your site.
You cannot optimize for each of those queries by hand. You can cover the questions in your niche so thoroughly that whichever branch an agent explores, it lands on you. The AI citation study shows that breadth is exactly what gets cited.
- 01
Mine your Search Console for prompts
Filter for queries of ten words or more. They are the questions people now ask assistants, and a content plan in plain sight.
- 02
One precise question per page
Fan-out searches are narrow and stacked with constraints. A page that answers exactly that gets picked for that branch.
- 03
Count impressions as reads
An impression with no click on a machine-shaped query is an AI reading you. Track it, do not prune the page for it.
The case against
4 reasons to discount these numbers
Every dataset has blind spots. These are ours, stated plainly.
We cannot see who searched
Search Console does not say whether a query came from a person, a script or an AI agent. The patterns here are strong fingerprints, not proof, and some long questions are typed by real people. The ChatGPT log shows the fingerprints are real, but it covers one assistant and 48 questions.
Google hides rare queries
Queries searched once or twice are anonymized and never appear in the data. Machine queries are often one-off, so their real volume is much higher than what we can count.
A smaller, uneven sample
111 sites had query-level data, and a few large ones dominate the pooled tables. That is why the headline numbers count sites, not impressions.
Language noise
Our rules were written for English. Some non-English queries land in the keyword-stuffed group simply because our stopword list does not cover their language.
Methodology
How we ran the numbers
- Data: query-level Search Console data (up to 25,000 queries per site over 16 months) for the 111 sites where Outrank had collected it: 1,179,780 queries and 734.4M impressions. Google hides rare queries for privacy, so long, one-off machine queries are heavily undercounted.
- Operators & templates: queries containing quotation marks, OR, site:, intitle:, inurl:, filetype: or a minus-exclusion. Templates are the subset with unfilled {placeholders}.
- Prompts: eight words or more that start with an instruction (write, make, create, explain, summarize, compare…), address the reader as “you”, or contain several sentences.
- Keyword-stuffed strings: nine words or more with no stopwords at all (no “the”, “for”, “how”…), in plain ASCII. Long conversational questions are ten words or more starting with a question word.
- Question tax: for the 39 sites with at least 1,000 impressions on both question and non-question queries ranking in the top 3, excluding brand queries.
Cite this study
Outrank (October 1, 2026). Query fan-out is already in your Search Console: 1.2M queries from 111 sites. https://www.outrank.so/case-studies/query-fan-out
Journalists, researchers and bloggers are welcome to quote these numbers and screenshot the charts with a link back to this page.


