Keyword Research in the AI Era — Query Fan-Out, Conversational Intent and the Words Nobody Types

For two decades keyword research came down to one move: find a phrase with high volume and low competition, write an article for it, repeat. In the AI era that move stopped working — not because keywords disappeared, but because how the search engine understands a query changed. Google's AI Mode no longer matches one phrase to one page. It breaks your question into dozens of parallel sub-queries, searches them simultaneously and synthesizes an answer from many sources at once. This mechanism is called query fan-out, and it now decides whether you're visible.

Google's AI Mode doesn't search for one phrase — it breaks your question into dozens of parallel sub-queries (query fan-out) and synthesizes an answer from many sources at once. That upends keyword research: what wins is covering the whole intent, not a single phrase from autocomplete. The complete workshop: fan-out mechanics, prompt research instead of keywords, a conversational intent taxonomy, 6 sources of topics and mapping to clusters.

The consequence is brutal for the old approach: you no longer win with a single phrase, but with coverage of the entire intent behind the question. And if that's true, research stops being a hunt for words and becomes a mapping of the questions real people ask a machine in conversation. In this post I show how to run that research from scratch: how fan-out works, what users are really looking for, how to classify conversational intent, where to find topics that aren't in any keyword tool, and how to turn it all into content clusters.

What query fan-out is — how AI breaks one question into dozens

A classic search engine took your query and looked for pages best matching those specific words. AI Mode works differently: it treats the question as an intent to decompose. Ask a complex question and, behind the scenes, the system generates a bundle of related sub-queries, runs them in parallel, gathers results from different sources and only then assembles a single answer. That's query fan-out.

Let's see it on a concrete example. The user no longer types "construction CRM" — they ask in a full sentence, and the AI decomposes it into a fan of queries:

User query:  "which CRM for a small construction company in Poland"AI Mode fan-out into parallel sub-queries:  - best CRM for a small company  - CRM with a module for the construction industry  - Polish-language CRM with invoicing and tax filing  - CRM integration with cost estimates / accounting  - CRM for a team of up to 5 — pricing  - reviews and comparisons of popular CRMsSynthesis -> one answer with citations from many sources at once

The takeaway for research: it no longer makes sense to aim at one "primary keyword". Your page has a chance to be cited if it strongly answers any of these sub-queries — ideally several at once. That's why topic coverage beats optimizing for a single phrase. It's the same logic behind building topical authority and optimizing for AI Overviews.

Why classic keyword research stopped being enough

The old research isn't wrong — it's incomplete. It focuses on what a tool can measure (volume, difficulty) and ignores what the tool can't see: the entire long-tail, conversational layer of questions nobody types into a search box, because they ask a machine in conversation instead.

The most important row is the long tail. Queries with "zero volume" in the tool aren't dead — they're invisible to the tool but alive in conversations with ChatGPT and in AI Mode. These are the "words nobody types": full sentences, clarifying questions, context-laden variants. In the fan-out model, they build the answer.

From keywords to prompts — researching what people actually ask

If users converse with the search engine, the unit of research stops being a phrase and becomes a prompt — a real question in natural language. Prompt research means collecting the questions your customer asks at every stage of the decision, in the exact form they ask them: "is X suitable for Y", "how much does X really cost", "what's better, X or Z, if I have constraint W".

This isn't a cosmetic change of vocabulary. A prompt carries context a bare phrase doesn't: the user's situation, constraints, funnel stage. And the model — and AI Mode — selects sources precisely by that context. That's why well-done prompt research is simultaneously intent research.

A taxonomy of conversational intent

So that research isn't a bag of random questions, it helps to classify them by intent. In the conversational era, the classic split (informational / navigational / transactional) is too coarse. The practical taxonomy I use:

This classification has a practical purpose: each intent demands a different answer format on the page. A comparison question calls for a table, a procedural one for a list of steps, a verification one for hard data and E-E-A-T signals. If the content's format matches the intent, it wins retrieval; if not, even a relevant topic won't get cited.

Six sources of topics you won't find in a keyword tool

A keyword tool shows you what others already type. The real edge is in sources that reveal questions before they become popular phrases. The six I reach for:

1. 1.Search Console — your real queries. The Performance report is the most underrated source. Export queries from 12–16 months and filter for question forms (how, why, does, when, how much). It shows what you already appear for — often questions you didn't know existed.
2. 2.People Also Ask and autocomplete. An expanding tree of related questions. Every click on "People also ask" generates more — a ready map of sub-intents around the topic.
3. 3.The models' own answers. Ask ChatGPT or Perplexity your customer's question and read which threads they raise in the answer. These are the sub-queries the model considers relevant — exactly the ones you must cover to get cited.
4. 4.Forums and user content. Reddit, industry groups, Q&A sections. Here people write questions in their own words, without "optimization" — and AI models happily cite these sources.
5. 5.The gap versus competitors. Not to copy their phrases, but to find sub-topics they don't cover. In the fan-out model, whoever answers a question the rest skipped wins.
6. 6.Sales and support conversations. The questions a customer asks a human before they ask Google. The cleanest source of transactional and verification intent — and usually entirely absent from tools.

Together, these six sources give you something no single tool can: the full language of your customer, from the first problem to the moment of purchase.

How to turn research into content clusters

Collected questions are useless as a flat list. Value emerges when you group them. The process:

• Group by intent and topic, not by phrase. Dozens of variants of the same question are one topic, not ten. Semantically close questions (via embeddings or by hand) form one cluster.
• Set a pillar and satellites. Each cluster is one pillar article covering the topic broadly and several supporting ones answering specific sub-questions. That's the hub-and-spoke model from the topical authority post.
• Map to existing content before writing new. Check in GSC whether you already rank for a question with another page — if so, update it instead of creating a second one and cannibalizing your own traffic.
• Write a brief with questions as headings. Questions from research become H2 headings directly. That way the page answers fan-out sub-queries in the exact form they're asked.

The result isn't a list of phrases but a topic map: what to write, in what order and how to connect it — the foundation on which GEO and AI visibility rest at all.

Tools — what actually works in 2026

You don't need an expensive stack to start. In order of real value:

• Google Search Console — free and most important. Your actual queries, impressions and CTR. The starting point of any research.
• Classic tools (Ahrefs, Semrush) — still useful for volume, difficulty and gap analysis, increasingly with AI modules. Treat them as one source, not the whole truth.
• Question-tree tools (AlsoAsked, AnswerThePublic) — fast mapping of People Also Ask around a topic.
• Google Trends — validate seasonality and momentum: is the question rising or fading. It doesn't replace intent research, but it helps set priority and publication timing so you don't write for a topic that's already declining.
• The models themselves (ChatGPT, Perplexity) — the best tool for researching sub-intents and customer language. Ask a question, analyze the answer's structure and the threads it cites.
• Embeddings for clustering — at a large question count, semantic grouping with vectors turns chaos into clusters in minutes; automating that belongs to the SEO automation workshop.

The most common mistakes

• Chasing volume instead of intent. A high-volume phrase with unclear intent converts worse than a precise question with "zero" volume in the tool.
• One article = one phrase. In the fan-out model that's a recipe for thin content that loses to one page covering the topic comprehensively.
• Ignoring the long tail and conversational questions. That's exactly the layer that drives AI citations.
• No mapping to existing content. Creating a new page for a topic you already rank for is self-cannibalization — a common and costly mistake.
• Skipping citation measurement. If you measure only positions, you'll miss half the effect; you need an AI visibility audit and measurement of traffic from models.

Step-by-step rollout plan

1. 1.Export queries from GSC (12–16 months) and filter for questions (how, why, does, how much, when).
2. 2.Expand with question trees from People Also Ask and autocomplete (AlsoAsked).
3. 3.Query the models with the customer's questions and gather the sub-threads from their answers — a ready list of fan-out sub-queries.
4. 4.Add the language of forums, UGC and support conversations — the cleanest intent, absent from tools.
5. 5.Classify by conversational intent using this post's taxonomy.
6. 6.Group into clusters (embeddings or by hand) — topic, not phrase.
7. 7.Map to existing content — refresh instead of a new page where you already rank (avoid cannibalization).
8. 8.Build hub-and-spoke and briefs with questions as H2 headings.
9. 9.Measure the effect on two tracks: positions in Google and citations in AI — via analytics of traffic from models, not volume alone.

---

I run AI-era keyword research from a GSC export to finished clusters and briefs with questions as headings — as part of AI optimization (GEO) and SEO content marketing. I teach this in the SEO & GEO course. Get in touch — I'll start with a map of your customer's questions and the gap versus competitors.

Worth reading next:

• Topical authority — building topical authority through content clusters
• Google AI Overviews and AI Mode — how to optimize your page so AI cites you
• How to get cited in ChatGPT — an LLM visibility strategy
• GEO — Generative Engine Optimization from the ground up