Ask ChatGPT with search turned on: "What's the best project management tool for a 10-person agency?"
It doesn't send that sentence to a search engine. It breaks it into several narrower searches, collects results for each, and writes one answer from the pile. That process is called query fan-out.
If you build docs, content sites or SaaS landing pages, fan-out explains a frustrating thing: a page can rank first on Google for your main keyword and still never show up in an AI answer.
What fan-out looks like
Here's a plausible fan-out for the prompt above:
best project management tool for a 10-person agency
├── best project management software for agencies 2026
├── project management tools for small teams pricing
├── Asana vs Monday vs ClickUp for agencies
├── project management tool client portal features
├── agency project management reviews
└── project management tool time tracking integration
Notice what gets added: a year, "pricing", "vs", "reviews", and a feature that matters for the use case. Each branch is a separate retrieval step. The final answer is built from whatever ranked well for those branches, not for the original sentence.
Why this matters for your content
Retrieval happens per sub-query. So the question isn't "do I rank for my keyword?" It's "do I have the best page for any of the searches the assistant actually runs?"
That leads to three practical rules:
- Write pages that answer one sub-query completely. A pricing page, a comparison page and a feature page each catch a different branch.
- Use the modifiers the assistant adds. If fan-outs keep adding "for small teams" or "2026", your page should say who it's for and show when it was last updated.
- Don't rely on one giant guide. A 6,000-word ultimate guide competes for one branch at best.
This is retrieval-augmented generation (RAG) in practice: the model can only cite what the retrieval step hands it. We walk through the whole pipeline in how ChatGPT generates answers.
Simulate a fan-out in 20 lines of Python
You can't see ChatGPT's internal planner, but you can approximate it by asking a model to play search planner:
`import os
from openai import OpenAI
client = OpenAI() # reads OPENAI_API_KEY from your environment
MODEL = os.environ["MODEL"] # any chat model your account can use
PLANNER = (
"You are a search planner for an AI assistant. "
"Rewrite the user's question into 6-8 web search queries "
"you would run to answer it well. One query per line, no numbering."
)
def fan_out(question: str) -> list[str]:
resp = client.chat.completions.create(
model=MODEL,
messages=[
{"role": "system", "content": PLANNER},
{"role": "user", "content": question},
],
)
text = resp.choices[0].message.content
return [line.strip() for line in text.splitlines() if line.strip()]
for query in fan_out("best project management tool for a 10-person agency"):
print("-", query)`
Run it with MODEL= python fanout.py. Run it five times and watch the sub-queries shift. That variation happens in production too, which is why checking an AI answer once tells you very little.
This is an approximation, not ChatGPT's real planner. But the pattern it produces (modifiers, comparisons, pricing, reviews) is close to what you see in practice.
Does ChatGPT just use Google?
Not directly. ChatGPT's searches go through OpenAI's own search stack, and the sources it cites often differ from Google's top 10. We looked at the overlap in does ChatGPT use Google results to generate an answer?
A 15-minute exercise for your own site
- Take your most important keyword and phrase it as a question a buyer would ask.
- Run it through the fan-out script above, or paste it into the free query fan-out generator.
- For each sub-query, write down the single page on your site that best answers that exact search.
- Any sub-query with no matching page is a content gap. That's your next page to write.
Most teams find that their pricing and comparison pages are the weakest links. Those are exactly the sub-queries AI assistants run most.
TL;DR
• AI search turns one prompt into many searches.
• You get cited by winning one of those searches, not the original prompt.
• Build focused pages for the sub-queries buyers trigger: pricing, comparisons, use cases, reviews.
Have you tried mapping fan-outs for your own product? Share what surprised you in the comments.
Written by the team at Track My Visibility, where we track how brands show up in AI search.

Top comments (1)
Have you tried running the fan-out script on your own brand name yet? The bit that stuck with me was the 6,000-word guide competing for one branch at best — we've seen the same thing from the other side: YouTube Search drove 23.9% of our views in the last 7 days, and it's almost always one very specific query doing the work, not the video's broad topic. Makes me wonder how much of the 'one branch' problem is really about pages trying to cover too much.