Articles / Playbook
How to write for the sub-questions AI engines search: from searches to a page plan
Group the searches AI engines run into sub-questions, rank them by how many answers searched them, and write the page each needs. A worked sample.
- Published
- Reading time
- 8 min read
- By
- AmpliRank
- Tags
- Query Fan-out · Content Strategy · Playbook
To write for the sub-questions AI engines search, collect the searches they run across many answers, group them by what each one asks for, and rank the groups by how many answers searched them and left you out. Then give each group the one page it needs: a pricing search needs prices, a “best tools” search needs a place on someone else’s list, and a feature search needs a page about that feature.
In short
- Engines with web search on research a buyer’s question through several narrower searches. Your page for the buyer’s own words can miss all of them.
- Group searches by what they ask for, not by their wording, and count how many answers ran each group.
- Rank groups by answers searched, then by how rarely you’re named in them. Ranges on small counts overlap, so treat close groups as ties.
- Each kind of search needs a different kind of page, and some, such as “best tools” lists, can only be answered on other people’s sites.
- Searches are a sample like the answers: measure again on the same questions to see whether the searches and the brands named moved.
Why write for sub-questions rather than the buyer’s question?
Because the engines don’t search the buyer’s question; they search the parts of it. Google says AI Overviews and AI Mode “may use a ‘query fan-out’ technique”, running related searches across subtopics. Anthropic’s documentation for Claude’s web search says single-fact queries typically use one to three searches, while comparative or multi-entity research “can use 10 or more”.
A buyer asking for the best project management software for agencies is asking a comparative question. The engine may search for pricing, for client access, for alternatives to a named tool. Each search returns its own pages, and the answer is written from those pages, its cited sources. Google calls the step query fan-out, and what is query fan-out? explains it. This article covers what to write once you can see the searches.
The common advice is to write one comprehensive page for the head term. That page answers the buyer’s words, and an engine that searched “agency project management pricing per user” may never read it.
How do you turn recorded searches into sub-questions?
Record the searches with each answer, across many answers, then group them by the fact or page type each one is looking for. One answer’s searches are an anecdote: they change by engine, by day and by phrasing, the same way the answers do.
Four rules keep the grouping honest:
- Group by intent, not wording. “agency PM pricing per seat” and “project management cost per user agencies” are one sub-question: what does it cost per person.
- Count answers, not searches. An answer that ran three pricing searches counts once for pricing. Otherwise one thorough answer outweighs ten brief ones.
- Keep the brand-named searches apart. A search for “Workloft alternatives” asks for something different from a search with no brand in it, and it tells you which rival the engine treats as the reference.
- Note which engines report searches. Google AI Mode shows the sources it cited but not its searches, so its answers add to the sources, not to the search counts.
You need a fixed set of unbranded buyer questions to record searches against, asked the same way every time.
The 60 buyer questions template is a stage-tagged starting set for B2B software, from first problem to final decision.
Download the 60 questions as a CSV.
Which sub-questions should you write for first?
Start with the sub-question searched in the most answers that leave you out. A frequent search where you’re absent is where a page of yours, or a mention on someone else’s, has the most answers to change.
A sample, to show the arithmetic. Taskmoor is a fictional project-management tool for agencies. Its team asked one buyer question, “What’s the best project management software for a 40-person agency?”, 10 times to each of three engines that report their searches: 30 answers. The searches grouped into five sub-questions:
| Sub-question (sample) | A search in the group | Searched in | 95% range | Taskmoor named | Taskmoor page today |
|---|---|---|---|---|---|
| Best tools for agencies | best project management software for agencies 2026 | 26 of 30 | 70.3% to 94.7% | 5 of 26 | none (a list) |
| Price per person | agency project management pricing per user | 17 of 30 | 39.2% to 72.6% | 2 of 17 | pricing page, no per-seat price in text |
| Client access | project management client portal for agencies | 14 of 30 | 30.2% to 63.9% | 1 of 14 | none |
| Resource planning | agency resource planning software | 11 of 30 | 21.9% to 54.5% | 4 of 11 | one blog post |
| Alternatives to a rival | Workloft alternatives | 8 of 30 | 14.2% to 44.5% | 0 of 8 | none |
Three readings follow from the table:
- “Best tools” is searched in 26 of 30 answers, and Taskmoor is absent from 21 of those 26. It is the largest gap, and it isn’t one Taskmoor’s own site can fill (see the next section).
- Client access and price per person are the clearest own-page gaps. Both are frequent, and Taskmoor is named in 1 of 14 and 2 of 17 of the answers that searched them.
- Client access and resource planning can’t be ranked against each other. Their ranges, 30.2% to 63.9% and 21.9% to 54.5%, overlap widely. With 30 answers, treat them as a tie and pick the one with the weaker page.
The ranges are Wilson intervals, the same ones the AI visibility calculator computes for your own counts.
What kind of page does each sub-question need?
The page type follows from what the search asks for. Some searches want a fact, some want a comparison, and some want a list that only a third party can credibly write.
| The search asks for | Example | What to publish | Where |
|---|---|---|---|
| A list of tools | best project management software for agencies | A complete, accurate entry on the lists the engines cite; a homepage that states your category and audience in one sentence | others’ sites, then yours |
| A price | pricing per user | Prices, limits and what each plan includes, in text | your pricing page |
| A capability | client portal for agencies | One page per capability, named in its title, saying who it’s for and what it does | your site |
| A how-to | how to plan agency resources | A guide that answers the task, with your product as one option | your site |
| An alternative or comparison | Workloft alternatives | A factual comparison: what each does, prices with sources, who each suits | your site |
| A security or compliance fact | SOC 2 project management | The certifications you hold and the documents behind them | your security page |
On lists, our data note on cited sources found 130 of 346 sources cited in 45 answers on 13 November 2026 were “best tools” roundups, in one category. For a “best tools” search, getting onto the cited lists is the work; the playbook guide gives the outreach checklist.
The same data note found no cited page titled as a comparison, 0 of 346. That has two readings: few such pages exist in that category, or the engines preferred other pages. Either way, a comparison page is a bet worth measuring, not a sure win.
How should the page itself be written?
Write it so the sentence an engine needs is the first one on the page. A page that answers its sub-question in a full sentence, with the number or fact stated in text, gives an engine something to quote.
For a page written for a sub-question:
- One sub-question per page. A pricing page that is also a feature tour answers neither search plainly.
- Use the searched words in the title and first heading. “Client portal for agencies” is the page; “Collaborate seamlessly” is not.
- State the fact in the first two sentences. For the sample brand: “Taskmoor’s client portal gives each client a read-only view of their projects, included on every plan.”
- Name the audience. Searches carry their audience (“for agencies”, “for a small team”); the page should say it too.
- Date facts that change. Prices and limits with an “as of” date are easier to trust and to keep current.
The rest is ordinary hygiene, and Google lists it for its AI features: crawling allowed, internal links, important content in text, and structured data that matches the visible text. The playbook guide’s page checklist puts those points with the ones above.
What can’t the searches tell you?
The searches show what the engines looked for, not why they chose the brands they named. A page that answers a frequent sub-question raises the chance that an engine reads you; it doesn’t make the engine name you.
Three limits matter when you plan from them:
- They are a sample. Thirty answers rank the large groups, not the close ones. Small groups need more answers before they deserve a page.
- They move. Engines change their search behaviour without notice. A plan built on one month’s searches needs the next month’s to confirm it.
- A change after you publish is observed, not caused. If Taskmoor is named more often after its client-portal page goes live, the same run also saw the engines, rivals and cited pages change. Compare on the same questions and engines, and call a change only when its range excludes zero (is that change real?).
The product tour shows this on a sample run: the searches the engines ran, and each question matched to the brand’s best page, marked as answering it, partly, or not at all.
Where to start
This week, take three unbranded buyer questions and ask each one ten times in the engines your buyers use, recording the searches where the engine shows them. Group the searches by what they ask for, count the answers behind each group, and write the page for the most-searched group where you’re absent and your own site can answer.
Sources
- Google Search Central, AI features and your website, accessed 14 November 2026.
- Anthropic, Web search tool, accessed 14 November 2026.
- Edwin B. Wilson, Probable Inference, the Law of Succession, and Statistical Inference, Journal of the American Statistical Association 22 (158), 209–212, 1927, accessed 14 November 2026.
- AmpliRank, Why “best tools” roundups decide AI answers, 45 answers from Gemini, Claude and Google AI Mode, 13 November 2026.