If you want ChatGPT to surface your pages, cover the exact sub-questions its query fan-out breaks a prompt into — and do it on pages structured so each answer stands on its own. Ahrefs just analysed 15 million sub-queries and found ChatGPT only retrieves live web content in 14.4% of responses. That single number changes how you should think about writing for it.
Here's the short version: ChatGPT rarely goes shopping on the open web for a fresh answer. When it does, it isn't searching your target keyword — it's searching the smaller questions it invented to satisfy the user. Win those small questions and you earn a seat at the table. Ignore them and your keyword-perfect page never gets looked at.
This post turns the Ahrefs data into a concrete page-structuring move, with a worked example from a one-person site.
What query fan-out actually is
Query fan-out is when an AI model takes one user prompt and silently splits it into several narrower searches before answering. So a prompt like "best way to back up a WordPress site on shared hosting" doesn't become one lookup — it becomes a handful: how often to back up, where to store backups, free vs paid plugins, how to restore, whether the host already does it.
The model fires these sub-queries, pulls whatever it finds, and stitches a single reply together. You never see the fan-out. But it decides which pages get read, and a page that only answers the headline question gets skipped in favour of one that happens to nail a sub-question.
The Ahrefs study measured this at scale: 15 million sub-queries generated from real prompts. The headline finding — retrieval happens in only 14.4% of responses — tells you the model prefers its training memory. The second, quieter finding is the useful one: when it does retrieve, it's chasing specific sub-questions, not your main keyword.
Why the 14.4% number should change your plan
Because retrieval is rare, you can't rely on being "the fresh result" — you have to be worth retrieving when the model finally reaches out. Most pages are written for one keyword and one search intent. That's a bet on being the best match for a query the model may never run.
Flip the logic. The 14.4% of answers that do fetch the web tend to be the ones where the model lacks confidence: recent changes, niche specifics, pricing, step counts, version numbers, local or regulatory details. Those are exactly the sub-questions a good page can own. Generic overviews lose here because the model already "knows" the generic stuff from training.
So the move is not more content. It's content aimed at the gaps in the model's memory — the precise, fresh, specific things it has to look up. If your page is the clearest source for "how many WordPress backups to keep" or "what GDPR requires for EU-hosted backups," you're in the retrieval lane, not fighting over the saturated keyword.
How to find the sub-questions before you write
To find the fan-out, ask the model to show its work and mine the "people also ask" ecosystem around your topic. You don't need the Ahrefs dataset — you can reverse-engineer the fan-out for any topic in about ten minutes.
- Ask ChatGPT directly: "List the sub-questions you'd need to answer before replying to: [your prompt]." It will happily enumerate them. That list is your section plan.
- Pull the autocomplete long tail: type your topic into Google and note the suggestions, plus the "People also ask" box. These overlap heavily with fan-out sub-queries.
- Run the prompt a few different ways: change "best," "how to," "is it safe to," "cheapest." Each phrasing surfaces a slightly different sub-question set.
- Look for the specifics: numbers, steps, prices, versions, dates, locations. These are where the model's memory is weakest and retrieval is likeliest.
Write the list down. Each distinct sub-question becomes an H2 or H3 on your page — and the first sentence under it is a complete, standalone answer. That structure matters: the model lifts self-contained sentences, not paragraphs that only make sense after three others.
A small-site example that actually works
Here's the move applied to a solo baker's site competing against recipe giants — and winning the sub-questions the giants bury. Say the target is "sourdough starter troubleshooting." The big sites rank for the phrase. The baker can't outrank them, so she structures for the fan-out instead.
The fan-out for that topic looks roughly like this:
| Sub-question the model fans out to | How the small site wins it |
|---|---|
| Why is my starter not rising? | One H3, first line a direct cause-and-fix, then the detail |
| What does a healthy starter smell like? | Specific descriptors, not "pleasant and tangy" |
| How long before I throw a starter out? | An actual number: "after 10 days of no activity despite daily feeding" |
| Can I revive a starter from the fridge? | Exact steps with feed ratios and timings |
Each row is a short, self-contained block. The baker isn't writing a 4,000-word mega-guide — she's writing four tight answers the model can quote verbatim. Because each one carries a specific number or step the big generic pages gloss over, hers is the page worth retrieving when confidence dips. The same approach works for a plumber, a bookkeeper, or anyone selling to a local market: own the specifics the national players skip.
A page-structure checklist you can apply today
Structure every page so each sub-answer is independently quotable, specific, and placed near the top of its own section. Run your draft against this list before you publish:
- One sub-question per heading. If a heading answers two things, split it.
- First sentence answers fully. No "it depends" openers — take a position, then qualify.
- Front-load the specifics. Numbers, steps, versions, prices, dates. Vague text doesn't get retrieved because the model already has the vague version.
- Keep answer blocks short. Two to four sentences per sub-answer reads naturally and lifts cleanly.
- Use a table or list where comparison helps. Structured data is easy for a model to parse and reuse.
- Refresh the fresh bits. If your edge is current detail, keep it current, or the retrieval advantage evaporates.
None of this fights good writing for humans — a reader skimming for one specific answer wants exactly the same structure a model does.
Where hosting quietly helps
Fast, reliable, EU-hosted pages give the model fewer reasons to skip you on the rare occasion it does fetch the web. Retrieval tends to favour pages that load cleanly and aren't buried behind errors, and if your topic touches EU readers, being EU-hosted and GDPR-friendly is one of those specific details the model may surface.
That's part of why we run TPC Hosting the way we do: EU data centres, plain honest pricing, free migration if you're moving an existing site, and 30 days to back out if it isn't a fit. If something breaks while you're restructuring your pages, you can reach AI on chat at any hour, with real engineers following up during business hours — so you're not stuck waiting to get back online.
The content work is yours to do. But the plumbing underneath — fast load, no downtime mid-crawl, EU hosting you can point to — is worth getting right while you're at it.
FAQ
Does covering sub-questions mean writing longer pages?
No — it means writing more specifically, not more. A tight page that answers four sub-questions clearly beats a 4,000-word guide that answers them vaguely, because the model already holds the vague version in its training memory.
If ChatGPT only retrieves the web 14.4% of the time, is SEO for AI worth it?
Yes, because that 14.4% is where you have the best shot. Retrieval happens exactly when the model is unsure — on fresh, specific, niche or local details — which are the sub-questions a focused page can own outright.
How do I see the sub-questions ChatGPT generates for my topic?
Ask it directly: "List the sub-questions you'd need to answer before replying to [your prompt]." Cross-check that list against Google autocomplete and the People Also Ask box, and each distinct item becomes a heading on your page.
Can a small site really win against big publishers this way?
Yes, on the specific sub-questions the big sites bury or skip. You can't outrank them on the headline keyword, but you can be the clearest source for an exact number, step, or local detail they treat as an afterthought.
Does my hosting affect whether AI tools retrieve my pages?
It can, indirectly. Pages that load fast and stay up are easier to crawl and quote, and details like being EU-hosted and GDPR-friendly are exactly the kind of specifics a model may surface for the right query.
