You type one question. ChatGPT runs five searches you never see.
You ask ChatGPT "which CRM is good for a small sales team." You expect it to search that phrase. It does not. It rewrites your question into a handful of separate searches, adds words you never used, runs them in parallel, and stitches the results back together before it writes a single word of the answer.
Most brands optimize a page for the question they think the buyer typed. ChatGPT is searching for something else. That gap is why a page can read perfectly and still never get cited.
This post shows what ChatGPT actually searches for, using a 5 million fanout study, and what to do about it. Half of it explains the mechanism. Half of it is the fix. The full operating loop lives in our AEO 101 playbook, refreshed daily from the research we publish.
The 40-second answer
ChatGPT does not retrieve your literal prompt. It expands one question into roughly two parallel searches, injecting modifier words like "best," "review," and "2026" that the user never typed. It then merges the results with reciprocal rank fusion, so a page that appears across several of those hidden searches outranks a page that appears in only one. To get cited, match the words ChatGPT injects and answer multiple angles on one page.
How ChatGPT actually runs your query
When you send a prompt, ChatGPT runs a step most people never see. It treats your question as an intent to interpret, not a search string to paste into a box.
On May 5, 2026, Peec AI published "Patterns we see in ChatGPT query fanouts", based on 5 million fanouts collected between April 1 and 21, 2026 across ChatGPT, Perplexity, and Grok. A fanout is one of the hidden searches a model spins up behind your single prompt.
The headline numbers:
- •ChatGPT runs an average of 2.1 fanouts per prompt
- •Perplexity runs 1.4
- •Grok runs 6.8
So your one question becomes two or more separate retrievals. Each one is a fresh search with its own phrasing. Each one is a separate chance for your brand to be present or missing.
You are not competing for one search. You are competing for the two or three searches ChatGPT writes on your behalf.
There is a second wrinkle. Peec and Search Engine Journal both found ChatGPT frequently rewrites non-English prompts into English fanouts. A buyer can ask in German and the retrieval that decides the answer runs in English. If your proof only exists in one language, you can lose the search before it starts.
The 10 words ChatGPT injects most
ChatGPT does not just rephrase your question. It adds vocabulary. Peec measured which words show up in fanout searches despite never appearing in the user's prompt, and the pattern is consistent enough to plan around.
Peec AI — 5 million query fanouts, Apr 1 to 21, 2026
The words ChatGPT adds to your search that you never typed
Share of affected ChatGPT responses where each word was injected into a hidden fanout query despite being absent from the user prompt.
“best” is injected more than twice as often as any other word. Listicles are built around it (“10 best”, “top 7”), which is why they keep winning the answer.
ChatGPT runs 2.1 of these rewritten searches per prompt, then merges the results with reciprocal rank fusion. A page that answers several of these angles at once scores higher than a page that answers only one.
Here are the ten most-injected words, by share of affected ChatGPT responses:
- •best at 15.33%
- •what at 8.72%
- •review(s) at 6.84%
- •2026 at 5.44%
- •top at 5.24%
- •comparison at 4.48%
- •vs at 4.27%
- •company / companies at 4.02%
- •service(s) at 3.99%
- •software at 3.47%
Read that list as a buyer's decision path written by the machine. It starts broad, adds "best" and "top" to rank options, adds "review" and "comparison" to validate them, and adds "2026" to demand current proof. The model is doing the shortlisting work a buyer used to do by hand.
"best" is the single word ChatGPT adds most. Listicles are built around it. That is not a coincidence, it is the mechanism.
This is also the cleanest explanation we have seen for why listicles dominate AI answers. We covered the trap version of that in why self-promotional listicles are nearly invisible. The short version: "best" and "top" pages match the question ChatGPT is really asking, but only the ones that read as neutral and well-sourced get pulled.
Why your single-angle page loses
The injection list explains the vocabulary problem. Reciprocal rank fusion explains the structure problem.
ChatGPT uses reciprocal rank fusion to merge its parallel fanouts. In plain terms: a source that appears across several of the hidden searches scores higher than a source that appears in only one. The model rewards pages that show up again and again across the angles it tried.
A narrow page that answers exactly one phrasing can rank well for that one fanout and never appear in the other five. A page that covers the category, the comparison, the use case, and the current-year evidence can surface across several fanouts at once and win the merge.
What you typed:
- •"is Acme good for small teams"
What ChatGPT actually searched:
- •best CRM for small teams 2026
- •Acme reviews
- •Acme vs [competitor]
- •small business CRM comparison
- •Acme pricing and features
Your product page might win the first search. It almost never wins all five. That is the part most GEO programs miss. They measure one well-ranked page as if it carries the whole answer, when the answer is assembled from a web of searches the page was never built for. The same split shows up in our piece on ranking pages versus grounding answers.
Want to see the fanouts ChatGPT runs for your category?
Cite Solutions maps the hidden searches behind your priority prompts, shows which ones name your brand and which route around you, and builds the multi-angle pages that win the merge. Done as part of every discovery call.
Book a Discovery CallWhat this changes about how you write
Once you accept that ChatGPT searches its own rewrites and not your phrasing, the content job changes. You are no longer writing for one keyword. You are writing for the cluster of searches the model generates around it.
Traditional SEO asks:
- •What keyword does this page target?
- •How many backlinks point at it?
- •Are we in the top ten?
AI fanout optimization asks:
- •Which words does ChatGPT inject for this topic?
- •Does one page answer several of those angles cleanly?
- •Will the model find us across more than one of its parallel searches?
You can win the left column and still be absent from the answer. The brands that get cited answer the right column.
Step 1: Map the words ChatGPT injects for your category
Take your ten highest-value prompts and run each one a few times in ChatGPT with search on. Watch the cited sources and the phrasing of the answer. You will see "best," "review," "comparison," and the current year appear even when your prompt did not include them. Write down the modifiers that recur. That list is your content brief.
Step 2: Build pages that answer several fanouts at once
Stop publishing one page per narrow keyword. A page that covers the category definition, a comparison, the use case, and current-year proof can surface across multiple fanouts and win the reciprocal rank fusion merge. One deep, multi-angle page beats five thin ones aimed at single phrasings.
Step 3: Add the comparison and review layer ChatGPT keeps searching for
"review(s)," "comparison," and "vs" together account for more than 15% of injected words. Those fanouts route to comparison pages, "alternatives to" articles, and third-party review surfaces. If your brand is not named in any of them, you are absent from the searches where buyers decide. Our guide to comparison pages that get cited covers the format.
Step 4: Put a current year on your proof
"2026" is injected in 5.44% of affected responses. The model is screening for recency on its own. A page with a visible current date, a recent stat, and a clear last-updated signal reads as safe to cite. A page that looks three years old gets skipped even when the facts still hold.
Step 5: Make each angle extractable as a clean passage
ChatGPT lifts a short passage, not a whole page. Each angle on your multi-angle page needs its own claim-format heading and a direct answer in the first sentence under it. If the model has to dig, it cites a competitor who answered in line one. The full pattern is in passages beat pages.
Match the words the machine injects, answer several angles on one page, and date your proof. That is the whole fanout playbook.
How this differs by engine
The fanout mechanic is not the same everywhere, so a single playbook will not work across platforms.
ChatGPT generates new sub-queries on almost every run, which means its fanouts are wide and varied. Perplexity barely rewrites the prompt at all, so its 1.4 fanouts stay close to your literal phrasing. Grok runs the most fanouts of any engine, 6.8 per prompt, and leans hard on site: searches against trusted domains, which we broke down in should B2B brands optimize for Grok.
The practical takeaway: optimizing for ChatGPT means covering many angles, while optimizing for Perplexity means nailing the exact phrasing of the question. We mapped the three separate playbooks in why the same content strategy fails across engines.
For context on how fast this is moving, Peec's earlier study found ChatGPT fanout length doubled in four months. The gap between what you typed and what the model searches is widening, not closing.
FAQ
Does ChatGPT search the exact words I type?
No. ChatGPT rewrites your prompt into an average of 2.1 separate searches and injects modifier words you never used, like "best," "review," and "2026." Peec AI's 5 million fanout study found these words appear in hidden searches despite being absent from the original prompt. The answer is built from those rewrites, not your literal phrasing.
What is a query fanout?
A query fanout is one of the hidden searches a model runs behind your single prompt. Instead of retrieving your question once, ChatGPT expands it into several parallel searches, each with its own phrasing, then merges the results. Peec measured an average of 2.1 fanouts per prompt on ChatGPT, 1.4 on Perplexity, and 6.8 on Grok.
Why do listicles keep showing up in ChatGPT answers?
Because "best" is the word ChatGPT injects most often, at 15.33% of affected responses, and "top" adds another 5.24%. Listicles are built around exactly those words, so they match the question the model is really asking. Neutral, well-sourced listicles get pulled. Self-promotional ones usually do not.
What is reciprocal rank fusion and why does it matter?
Reciprocal rank fusion is how ChatGPT merges its parallel fanouts into one ranking. A source that appears across several of the hidden searches scores higher than one that appears in a single search. It matters because it rewards pages that answer multiple angles at once and penalizes narrow pages built for one phrasing.
How do I optimize for ChatGPT's query fanouts?
Map the modifier words ChatGPT injects for your category by running your priority prompts and watching the cited sources. Then build pages that answer several of those angles at once, add comparison and review coverage, put a current year on your proof, and make each angle extractable as a clean passage. Optimize for the cluster of searches, not the single keyword.
The bottom line
The most useful thing the fanout data tells you is also the simplest. ChatGPT is not searching your question. It is searching its own rewrite of your question, several times, with words you never typed, and merging the results.
If you optimize one page for one phrasing, you are showing up in one of the searches that decide the answer and missing the rest. If you map the injected modifiers, answer several angles on one well-structured page, and keep your proof current, you start appearing across the fanouts instead of in just one of them.
Stop writing for the keyword. Start writing for the searches the machine writes for you.
Your buyers ask one question. ChatGPT runs five searches. Are you in them?
Cite Solutions maps the fanouts behind your category, finds the injected words and angles you are missing, and builds the pages that win the merge. The discovery call is free.
Book a Discovery CallRelated reading
- •AEO 101 living playbook: refreshed daily from our research
- •Why self-promotional listicles are nearly invisible in AI search
- •AI search is splitting into ranking pages and grounding answers
- •Why the same content strategy fails across ChatGPT, Perplexity, and Copilot
- •Passages beat pages: how to structure content for AI citation
Continue the brief
Did ChatGPT Just Kill Zero-Click for Brand Search?
On May 7 OpenAI made ChatGPT brand mentions clickable. Profound's monitored basket shows referrals up 60% overall and B2B SaaS daily traffic up over 200%.
Has ChatGPT Search Demand Already Peaked?
ChatGPT search demand hit 14,800/mo Jan-Mar 2026 then dropped 33% in April. AI overview and AI search queries are absorbing the demand.
Will Google's Core Update Hurt Your AI Citations?
Google's May 2026 core update finished rolling out this week. Here is why a ranking drop does not automatically cut your AI citations, and when it does.
Framework
Learn the CITE framework behind our GEO and AEO work
See how Comprehend, Influence, Track, and Evolve turn AI visibility into an operating system.
Services
Explore our managed GEO services and AEO execution model
Audit, prompt discovery, content execution, and ongoing monitoring tied to AI search outcomes.
Audit
Start with an AI visibility audit before execution
Understand prompt coverage, recommendation gaps, source mix, and where competitors are winning.