Stop optimising for keywords. Start optimising for sub-questions

One query in, many searches out

Google describes query fan-out in its own AI optimisation documentation: "Both AI Overviews and AI Mode may use a 'query fan-out' technique — issuing multiple related searches across subtopics and data sources — to develop a response." The system generates a set of concurrent, related queries and synthesises the results into a single answer.

Google's worked example makes the mechanics concrete. A visitor asks how to fix a lawn that's full of weeds. The AI doesn't search for that phrase. Instead it fans out into at least three parallel sub-queries:

Each sub-query retrieves content independently. The AI cites sources that answered sub-queries, not the parent query. If your page only targets the parent phrase, you are competing for a keyword the AI never actually searches.

Keyword tools can't see ~95% of it

Traditional keyword research tools — Ahrefs, Semrush, Google Search Console — aggregate volume from logged searches. Sub-queries generated by AI fan-out are synthetic: they are composed at inference time, not typed by users. Analysis of Ahrefs data by ALM Corp found that approximately 95% of fan-out sub-queries register zero search volume.

The Conductor 2026 AEO/GEO Benchmarks Report — drawn from 13,770 enterprise domains, 3.3 billion sessions, and more than 100 million citations across 17 million AI responses — found a threshold effect: brands that published 12 or more optimised content pieces gained AI visibility 200 times faster than those below that threshold. Not proportionally faster — categorically faster. Content breadth that covers the sub-question space is the mechanism.

Why AI cites sub-question content, not landing pages

Google patents US11663201B2 and US20240289407A1 describe the underlying mechanism: "query variant generation" and thematic sub-query expansion. The system identifies the intent dimensions of a query — causal, procedural, comparative, preventive — and searches for each independently. A landing page that makes a general claim satisfies none of them with specificity. A piece that goes deep on one dimension — say, the chemical-free removal angle — is exactly what the AI needs for that sub-query.

The practical implication: AI retrieval rewards specificity of answer over density of keyword. A page answering "can I remove lawn weeds without herbicides?" in 600 focused words will outperform a 2,000-word piece that mentions weeds, herbicides, and lawn care but answers none of those questions directly.

From topic to sub-question inventory

The workflow is different from keyword research but not complicated. Start with the questions your audience actually has — not the terms they might search — and decompose them by intent dimension.

For each topic you want to rank for in AI responses, write down the parent question, then generate sub-questions across these five dimensions. Each sub-question that lacks a specific, focused answer on your site is a gap the AI will fill from someone else.

Where this logic breaks — and what Google actually penalises

Fan-out sub-question coverage is genuinely useful, but it comes with real failure modes that are worth naming before you brief your content team.

Google explicitly warns against content written to cover variation, not to help. Its scaled content abuse spam policy flags content created "for every possible variation… primarily to manipulate rankings or generative AI responses." If you publish fifty thin pages that technically "answer" sub-questions but don't genuinely add depth, that is a spam signal, not a strategy. The goal is topical depth, not page count.

Fan-out doesn't trigger on every query. Lookup queries — "capital of Spain," "what year was the Eiffel Tower built" — get a direct retrieval, not a fan-out. The technique fires on exploratory, comparison, and problem-solving queries where multiple angles are genuinely useful. Mapping sub-questions to queries that are already unambiguous is wasted effort.

Thin content that targets sub-questions is still thin. The 12-piece threshold in the Conductor data is almost certainly a proxy for genuine topical authority, not a magic number to hit by publishing twelve short listicles. Quality over quantity applies here exactly as it does everywhere else.

The frame that holds up: write content because you have something specific and useful to say on that sub-question. If you do, the AI citations tend to follow. If you don't, they won't — regardless of how many pages you publish.

Find the sub-questions you're not answering

The fastest diagnostic is to run a brand mention scan and look at what the AI is actually saying when your topic comes up — which sub-questions it's pulling from competitors, which it's leaving unanswered, and which it's attributing to sources you've never heard of. That's your gap list.

Find the sub-questions you're missing