AI search changes keyword research in one fundamental way: instead of ranking for a single query, you now need to cover the cluster of sub-queries an AI engine generates from one conversational question. When someone asks ChatGPT or Google's AI Mode "how do I get more clients for my consulting firm without cold outreach," the engine does not run that sentence as one search. It fans the question out into multiple smaller lookups (referral systems, content marketing for consultants, email nurture, positioning) and assembles an answer from whichever pages answer each piece best.
That means the unit of keyword research is no longer the keyword. It is the topic and its plausible sub-questions. A page that answers one sub-question completely can get cited inside an answer for a much bigger, messier conversational query it never explicitly targeted. Your research process needs to map those sub-questions deliberately instead of chasing exact-match volume.
What Actually Changed About How People Search?
Two shifts matter for anyone doing keyword research in 2026. First, AI Overviews now appear on roughly half of Google searches, and around 60 percent of searches end without a click to any website. The classic model, where ranking position translated predictably into traffic, has weakened. Being the source an AI answer cites matters as much as holding position three on a results page.
Second, the queries themselves are getting longer and more conversational. People talk to AI assistants the way they would talk to a knowledgeable friend, in full sentences with context attached. A business owner who would have typed "email automation funnel" into Google now asks an assistant "I run a design agency and leads go quiet after the first call, what should I automate first?" Those conversational queries are roughly three times longer than classic search queries, and they carry far more intent signal.
The traffic that does arrive from AI referrals is small in volume but unusually qualified. A visitor who clicks through from a ChatGPT citation has already read a synthesized answer and decided your take was worth the extra click. Keyword research that ignores this surface is optimizing for a shrinking share of a shrinking pie.
How Does Query Fan-Out Work?
When an AI engine receives a conversational question, it decomposes the question into sub-queries, retrieves candidate pages for each one, and synthesizes an answer with citations. Google has described this retrieval behaviour in its AI features documentation, and you can observe it directly in Google's AI Mode and in Perplexity, which shows the searches it ran.
Take a concrete example. The question "should my accounting firm publish prices on our website" might fan out into sub-queries like "pricing page best practices service business," "pros and cons of publishing prices," "how accounting firms present fees," and "pricing anchoring psychology." Four different pages could be cited in the final answer, each because it answered one sub-query cleanly.
The practical implication is that a single 3,000 word page trying to cover everything usually loses to a set of focused pages that each own one sub-question. The engine is not looking for the most comprehensive page. It is looking for the clearest extractable answer to each piece of the puzzle. Depth still wins, but depth per question, not length per page.
If you want to know which questions in your market you currently answer and which ones competitors own, start with a real audit. Request your free marketing audit and we will map your topic coverage against how AI engines actually decompose your buyers' questions.
The Keyword Research Process for AI Search
Here is the process we use, adapted from classic keyword research but reorganized around fan-out behaviour.
Step 1: Start from buyer conversations, not from a keyword tool. List the ten questions prospects actually ask you on sales calls, in their own phrasing. These conversational questions are the closest proxy you have for what they type into an assistant. Keyword tools still matter, but they come second because they only measure classic search demand.
Step 2: Decompose each conversational question into sub-queries. For each question, ask what an engine would need to look up to answer it well. Definitions, comparisons, steps, costs, risks, examples. Paste the question into Perplexity and note the searches it runs. Do the same in Google's AI Mode and record which pages get cited. This takes about 20 minutes per question and produces a real map of the retrieval landscape.
Step 3: Validate classic demand for each sub-query. Now bring in Semrush, Google autocomplete, and People Also Ask. Some sub-queries have measurable search volume, which means a page targeting them earns both classic rankings and AI citations. Prioritize those. Sub-queries with no measurable volume can still be worth a section or an FAQ entry, just not a standalone post.
Step 4: Assign each sub-query a home. Every sub-query gets mapped to either an existing page, a new page, or an FAQ entry on a related page. One idea per section, phrased as a question-form heading where natural, answered in prose an engine can lift whole. This is the same structural discipline covered in our guide on how to show up in Google AI Overviews.
Step 5: Fill gaps in priority order. Rank the unowned sub-queries by business fit first, demand second. A sub-query that maps directly to a service you sell beats a high-volume query that attracts readers who will never become clients.
What Should You Stop Doing?
AI search does not make classic SEO obsolete, but it does make a few habits actively wasteful. Stop building five thin variations of the same page to catch keyword permutations; fan-out retrieval treats them as duplicates and cites none of them. Stop optimizing titles for exact-match strings at the expense of clarity; engines match meaning, not strings. And stop measuring success purely in organic sessions, because a growing share of your influence now happens inside answers you will never see in analytics.
Where the disciplines differ and where they overlap is its own topic, and we break that down in SEO vs AEO vs GEO for service businesses. The short version is that the research layer converges: the same sub-query map feeds all three.
Rebuilding Your Keyword List: A Practical Order
If you have an existing keyword list, here is how to convert it in roughly one working day.
- Cluster your current keywords by buyer question (2 hours). Group every keyword under the conversational question a real prospect would ask. Orphan keywords that map to no buyer question get cut.
- Run the fan-out test on your top ten questions (3 hours). Use Perplexity and AI Mode, record sub-queries and cited pages, and note every citation a competitor holds.
- Score each sub-query on business fit and demand (1 hour). A simple high, medium, low rating on each axis is enough.
- Write the coverage plan (1 hour). New pages, page updates, and FAQ additions, each tied to one sub-query, scheduled over the next quarter.
Teams that run this exercise usually find that 30 to 40 percent of their existing content answers no sub-query anyone asks, while several high-fit sub-queries have zero coverage. Closing that mismatch is the fastest visibility gain available right now, and it compounds because every new focused page becomes a citation candidate for dozens of conversational queries.
Frequently Asked Questions
Is keyword search volume still useful for AI search?
Yes, as a secondary signal. Search volume still measures classic Google demand, and pages that rank well organically are also more likely to be retrieved during AI fan-out. But volume tells you nothing about conversational queries, which are longer, more specific, and often invisible to keyword tools. Use volume to prioritize among sub-queries you have already identified through fan-out mapping, not as the starting point for research.
What is query fan-out in plain terms?
Query fan-out is what happens when an AI engine takes one conversational question and breaks it into several smaller searches behind the scenes. It retrieves pages for each smaller search, then writes one combined answer with citations. For content strategy, it means your page competes at the sub-query level, so a focused page that fully answers one specific question can be cited in answers to much broader questions.
How do I find the sub-queries AI engines generate?
Ask your target question in Perplexity, which displays the searches it ran, and in Google's AI Mode, and record which pages get cited for each. Repeat the question with different phrasings, because fan-out varies with wording. Supplement with Google's People Also Ask results and autocomplete, which reflect the same underlying question space. Twenty minutes per question gives you a workable map of the retrieval landscape for that topic.
Should I write one long guide or many focused pages?
Both, in a hub and spoke structure. A pillar guide gives you topical authority and a place to link everything together, while focused pages each own one sub-query with a complete, extractable answer. In practice the focused pages earn most of the AI citations because engines retrieve at the sub-query level. The pillar earns classic rankings for the head term and passes internal authority to the spokes.
Does AI referral traffic actually convert?
In our experience, yes, at noticeably higher rates than average organic traffic. The volume is small, often single-digit percentages of total sessions, but visitors arrive having already read a synthesized answer that referenced you, so they show up pre-qualified and further along in their decision. Treat AI referrals as a high-intent channel worth dedicated landing experiences, not as a traffic line to grow for its own sake.
Read Next
Find Out Which Questions You Actually Own
The free audit includes a topic coverage review: the buyer questions in your market, the sub-queries AI engines generate from them, and where your site answers versus where competitors get cited instead. The gap between the questions buyers ask and the questions your site answers is where revenue disappears.