Query fan-out: why AI search never answers the question you typed

When you ask an AI assistant with web search something like "best CRM for a 5-person agency", it usually doesn't run that exact search. It quietly splits the question into several smaller ones, searches for each, reads the results, and writes one answer from all of them.

This is called query fan-out, and if you care about showing up in AI answers, it changes how you should think about a page.

What actually happens

A simplified version of the pipeline:

1. Rewrite. The model turns your prompt into a handful of search queries.
2. Retrieve. Each query goes to a search index. The top results for each are fetched.
3. Read. The model pulls the passages that look relevant out of those pages.
4. Synthesize. It writes one answer and (sometimes) cites the pages it used.

For "best CRM for a 5-person agency" the fan-out might look something like:

best CRM for small agencies 2026
CRM pricing per user small team
CRM with client portal for agencies
HubSpot vs Pipedrive for agencies
CRM project management integration
simple CRM setup time

The exact queries differ between products and change with each run, and vendors don't publish them. But the pattern is consistent and is described publicly by Google for AI Mode and by others: one question in, many searches out.

Why this matters for your content

Classic SEO trained us to optimize one page for one keyword. Fan-out breaks that in two ways.

You can win without ranking for the head term. If your page is the best result for "CRM pricing per user small team", it can be read and cited for the big question, even if you're on page 5 for "best CRM".

The unit that gets used is a passage, not a page. The model needs a chunk it can lift: a clear definition, a short comparison, a number with context. A 3,000-word article where the pricing is buried in paragraph 14 is harder to use than a page with a section titled "Pricing for small teams" that answers in two sentences.

A practical way to use this

You don't need special tools to start. For each important page:

1. Write down the main question the page should answer.

2. Generate the likely sub-questions. Ask yourself (or an LLM) "what would someone need to look up to answer this properly?" Think in these buckets:

• Definitions: what is X?
• Comparisons: X vs Y, alternatives to X
• Constraints: X for small teams, X under $50, X without coding
• Cost and numbers: how much, how long, how many
• Steps: how to set up X
• Freshness: X in 2026, latest changes

3. Map each sub-question to a section of your page, or to another page on your site. If a sub-question has no home, that's a content gap.

4. Make each answer liftable. Put the direct answer in the first one or two sentences of the section, then add detail. Use the sub-question (or something close) as the heading.

A tiny script to keep yourself honest while auditing:

const subQuestions = [
  'what is a CRM',
  'CRM pricing per user',
  'CRM with client portal',
  'how long does CRM setup take',
];

const headings = [...document.querySelectorAll('h2, h3')]
  .map((h) => h.textContent.toLowerCase());

for (const q of subQuestions) {
  const words = q.toLowerCase().split(' ').filter((w) => w.length > 3);
  const covered = headings.some((h) => words.filter((w) => h.includes(w)).length >= 2);
  console.log(covered ? 'OK  ' : 'GAP ', q);
}

Paste it into DevTools on your page. It's crude (keyword overlap, not meaning), but it's enough to spot obvious holes.

Things that don't change

Fan-out doesn't replace the basics, it depends on them:

• The page has to be crawlable by the AI search bot (robots.txt, CDN, firewall).
• The content has to be in the raw HTML, because most AI crawlers don't run JavaScript.
• The answer has to be true and specific. Models are getting better at preferring sources with concrete numbers, dates and named sources over generic text.

What not to do

It's tempting to generate 40 thin pages, one per sub-question. That's the old doorway-page mistake in a new outfit. Search engines penalize it, and a thin page has no passage worth citing anyway. Better: one strong page with well-structured sections, plus separate pages only where a sub-question deserves real depth.

Try it

The Query Fan-Out Generator on BotReady takes a question and produces the kind of sub-queries an AI search engine is likely to run, so you can check them against your page. If you want to see whether assistants already mention you, the AI Citation Checker is the next step.

Have you noticed AI assistants citing a page of yours for a question you never targeted? I'd love to hear which one.