A client rang me in March, furious. His site had held a top-three position for "best CRM for small law firms" for two years. Traffic: steady. Then Google's AI Mode started rolling out more broadly to US users and, within six weeks, his organic clicks fell 34%. He hadn't lost the ranking. The blue link was still there. But a big AI-generated answer box was sitting above it, citing three other sites, none of which were his. One of them had a Domain Authority of 31. His was 58. I pulled up the page that got cited and spent about twenty minutes reading it. Honestly? It was better structured for the specific question Google was now asking, even if it wasn't "better" in any traditional sense.
That's when I started paying serious attention to query fan-out.
What Query Fan-Out Actually Is
When you type something into Google AI Mode, or when Gemini processes a query inside Search Generative Experience, it doesn't pass your single question to one retrieval process. It breaks the query into multiple sub-queries, fires them in parallel, retrieves results for each, then synthesises everything into one response. Google's own engineers have described this as a "multi-step reasoning" process, and Liz Reid's keynote at Google I/O 2024 gave the clearest public look at how the system thinks.
The number of sub-queries varies. Simple questions might fan out into five or six. Complex research questions, the kind where someone's actually trying to make a decision, can generate 20 to 50 parallel retrievals. Each sub-query targets a slightly different angle, a different entity, a different comparison, a different objection. The system is trying to build a complete picture, not just retrieve one answer.
Here's the thing: most SEO advice still treats Google like it's asking one question. It isn't. Not anymore.
Why This Breaks Traditional Keyword Optimisation
I've built content strategies for over 12,000 sites through Seahawk. And the single biggest mistake I see, still, in 2025, is pages optimised for a keyword rather than a topic cluster with explicit sub-questions answered inline.
Traditional optimisation said: put your keyword in the H1, the meta title, the first 100 words, get some backlinks, done. That worked because the retrieval was relatively linear. Match the query, surface the page.
Fan-out breaks this because the AI isn't looking for the page that matches the query. It's looking for the page that answers one specific sub-question that it generated internally. You never see that sub-question. It's not in Search Console. It doesn't appear anywhere in your analytics. Google created it, silently, as part of decomposing what the user actually asked.
The Retrieval Is Competitive Per Sub-Query
Think of it like this. Imagine Google splits "what accounting software should a UK freelancer use" into sub-questions like:
- What are the Making Tax Digital requirements for sole traders?
- Which accounting tools integrate with UK bank feeds natively?
- What's the price difference between FreeAgent and QuickBooks for freelancers?
- How do freelancers handle VAT submissions digitally?
- What do UK accountants actually recommend to their clients?
Each of those is a separate retrieval competition. You could win two of them and lose the other three. If you lose more than half, you probably don't get cited at all. If you win four out of five, you might become the primary source the AI leans on.
The implication: a single long-form page that vaguely covers "accounting software for freelancers" is not competing effectively. You need to answer each sub-question explicitly, clearly, with enough specificity that a retrieval system can extract a clean answer to that exact question.
How to Reverse-Engineer the Sub-Queries
This is where the work actually happens. I'm not going to pretend there's a tool that surfaces Google's internal fan-out queries directly, because there isn't. But you can approximate them.
Here's the process I use, and used specifically for a fintech client at Seahawk in late 2024:
- Type your target query into Google AI Mode (or the AI Overview if you're in a region where that's live) and read the generated response carefully.
- Every paragraph in that response is answering a different sub-question. Write out what that implied question is.
- Cross-reference with AlsoAsked, which maps the "People Also Ask" graph better than any other tool I've tried.
- Run the same query through Perplexity and Claude. Both use similar multi-step retrieval. The sub-questions they ask explicitly, written out in their reasoning traces, are a reasonable proxy for what Google is generating internally.
- Use Ahrefs' Content Gap tool to find pages outranking you that cover angles your page doesn't touch.
Once you have 8 to 12 implied sub-questions, write content that answers each one as a discrete, extractable unit. Not buried in prose. Not hedged. A clear question, a clear answer, then supporting detail.
What Makes a Page "Citable" in a Fan-Out System
Back in 2019 a client handed me a brief that I thought was ridiculous at the time: write a glossary page for every legal term on their conveyancing site. Not a blog. Not a guide. A proper glossary with 200+ definitions. I pushed back. They insisted. That page now drives roughly 18,000 organic sessions per month, and since AI Overviews started rolling out, it's been cited in AI responses more than any other page on their domain.
Why? Because each definition is a self-contained, extractable unit. Google can pull one definition to answer one sub-query without needing to interpret surrounding context. That's citability.
The signals I've come to think matter most for citability:
- Explicit question-answer structure. A heading phrased as a question followed by a direct answer in the first sentence of that section. Not buried three paragraphs down.
- Named entities and specifics. "FreeAgent costs £19/month for sole traders" is citable. "Some tools are more affordable" is not.
- Short, quotable sentences within longer sections. The AI is pulling snippets. Make the snippet obvious.
- Structured data. FAQ schema and HowTo schema still work, and they work specifically because they help Google parse discrete Q&A pairs. Use them. Google's structured data documentation is still accurate and worth re-reading in the context of AI Mode.
- Author credibility signals. Bylines with real credentials, About pages that are specific about expertise, links to the author's professional presence. Not just "by admin".
The Depth vs. Breadth Trap
Here's a mistake I made personally. Around mid-2023, when SGE was still in Labs, I wrote a guide for one of our own Seahawk service pages that was 4,800 words long. Comprehensive coverage of a topic. Every angle addressed. I was proud of it. It got cited exactly zero times in AI responses, and I watched a 1,100-word competitor page get cited repeatedly.
The difference: their page answered one specific sub-question with surgical precision. Mine answered everything so broadly that the retrieval system couldn't extract a clean, quotable answer to any individual sub-query.
Depth is not the same as length. Depth means going far enough on each sub-question that your answer is unambiguously useful. Length means writing a lot of words. These are different things.
If your topic generates 10 sub-queries, you're often better off with 10 focused sections of 150 words each than with one 2,000-word essay that addresses all 10 loosely. The structure signals that each section is a discrete answer. The AI can retrieve section 7 without needing sections 1 through 6.
A Note on Internal Linking
Because fan-out retrieves across many pages, not just one, your internal linking strategy becomes a way to signal topic authority across the sub-query space. If you have five pages that each answer two or three related sub-queries well, and they're properly interlinked with descriptive anchor text, you increase the chance that you win multiple retrieval competitions within the same fan-out event.
I use Screaming Frog to audit internal link anchor text across large sites. The number of sites I audit where 40% of internal links say "click here" or "read more" is depressing. Descriptive anchors are free wins.
How Fan-Out Affects Your Content Calendar
So practically: what does this mean for how you plan content?
It means the old "one keyword, one page" content calendar model is too coarse. You need to map topics to question clusters, then decide whether each cluster belongs on one well-structured page or split across several interlinked pages.
For a new project I'm running right now, a B2B SaaS client in the HR tech space, we mapped 40 target keywords to 180 implied sub-questions using the process above. Of those 180 questions, 60 were already answered somewhere on their site but buried in long-form text without clear extraction points. We didn't write new content for those. We restructured existing pages: added H3s phrased as questions, pulled the answer into the first sentence of each section, added FAQ schema. That took two weeks. Within six weeks, AI Overview citations for that domain went from three to nineteen tracked instances.
Not "write more." Structure better.
Monitoring Whether You're Getting Cited
There's no native Google Search Console filter for AI Mode citations yet, as of mid-2025. So you're working with proxies.
The most reliable method I've found: search your target queries manually, in incognito, across different regions using a VPN, and note which domains appear in the AI response. Do this weekly for your top 20 queries. It's manual. It's annoying. It's also the most honest signal you have.
Tools like Semrush's AI Toolkit and Ahrefs' AI mentions tracking (in beta at time of writing) are starting to surface citation data, but treat those numbers as directional rather than precise. The underlying access to Google's retrieval logs isn't there for third parties.
The metric I'd watch in Search Console as a proxy: impressions for question-phrased queries (queries starting with "how", "what", "why", "which", "can") combined with a drop in clicks despite stable impressions. That pattern usually means an AI Overview is satisfying the query before the user clicks. You're winning the impression, losing the visit.
FAQ
Does having more backlinks help you get cited in AI Mode responses?
Backlinks still influence which pages Google considers authoritative enough to pull from. But the relationship is weaker in AI Mode than in traditional ranking. I've seen pages with 12 referring domains get cited over pages with 400, because the lower-authority page answered a specific sub-question more extractably. Citability structure is increasingly the differentiator.
Should I restructure old content or write new pages to target AI Mode?
Start with restructuring. In my experience, about 60% of what's needed for AI Mode citability is structural, not topical. Add question-phrased H3s, pull answers to the first sentence of each section, add FAQ schema. If after restructuring a page still doesn't cover key sub-questions, then extend or supplement it.
Does page speed affect AI Mode citations?
Indirectly. Google still needs to crawl and index the page. Slow pages that trigger crawl budget issues or generate poor Core Web Vitals may be deprioritised in the index Google is pulling from. But page speed alone won't make or break citability. Content structure matters more.
Is AI Mode roll-out limited to certain regions?
As of mid-2025, AI Mode with query fan-out is most fully deployed in the US, with limited availability in other English-speaking markets. UK users see AI Overviews more than full AI Mode. Google's Search status dashboard is not particularly useful for tracking this. Your best signal is just searching in target regions manually.
---
The client from the opening? We restructured four pages on his site in April. Added explicit Q&A sections, FAQ schema, specificity on named tools and pricing. By mid-May he was appearing in AI responses for two of his five target queries. Not five. Two. This stuff takes time and it isn't guaranteed. But the mechanism is real, and ignoring it because you can't see the sub-queries directly isn't a strategy. It's just hoping the old rules still apply. They don't, quite.