Keyword Research for AI Search: A New Playbook

Keyword research for AI search still starts with what people want, but it captures full questions and prompts instead of stripped-down keywords. You mine the conversational language people type into ChatGPT and Perplexity, group it into topic clusters that prove expertise, and judge "AI traffic potential" by whether your answer gets cited, not just ranked.

How is keyword research different for AI search?

Classic keyword research optimised for clicks. You found a term like "best CRM," wrote a page, and chased the #1 spot for the traffic. AI search changes the destination. Answer engines read the question, pull facts from several sources, and write one synthesised reply. Often nobody clicks at all.

So the unit of research shifts from the keyword (a 2–4 word fragment) to the prompt (a full, messy, conversational question). Real AI prompts tend to run far longer than search queries; tools like Semrush and other analysts have observed that conversational queries often run to a dozen words or more, versus the three or four typical of traditional search. That length carries context: budget, location, constraints, intent. Your job is to capture that context and answer it directly.

Three things change in practice:

• You target questions, not fragments. "How do I choose a CRM for a 10-person agency?" beats "best CRM."
• You win citations, not just rankings. Being one of the sources an engine quotes matters more than position #4.
• You build topical depth, not single pages. AI engines reward sites that cover a subject thoroughly, because depth signals authority.

If you want the full picture of how AI engines pick sources, our AI search optimization guide covers the citation mechanics in detail.

How do you find the questions people ask ChatGPT and Perplexity?

You cannot pull AI prompts from a keyword tool the way you pull search volume. The prompts live in the wild. You mine them. Here is where the real language is:

1. Google Search Console question queries. Filter your impressions with a regex for ^(who|what|where|when|why|how|is|can|does|will|should). These are real questions your audience already asks search, and they map almost one-to-one onto AI prompts.
2. Reddit, Quora, and niche forums. This is gold for two reasons: it shows exact phrasing, and ChatGPT leans heavily on Reddit and Wikipedia when choosing what to cite. Read the threads, note the follow-up questions.
3. Sales and support transcripts. Your sales calls and support tickets are a record of the precise objections and questions buyers have. Nobody else has this data. Use it.
4. Autocomplete and People Also Ask. Still useful for surfacing the sub-questions clustered around a topic.
5. The engines themselves. Ask ChatGPT and Perplexity your seed question, then ask "what else do people ask about this?" Log the fan-out of related sub-questions they generate.

Once you have a seed prompt, expand it into 5–10 sub-questions. AI engines bundle answers across a "query fan-out," so a page that answers the main question and its satellites is far more likely to be the source the engine pulls from.

What is a topic cluster and why does it build authority?

A topic cluster is one comprehensive "pillar" page on a broad subject, surrounded by focused "spoke" pages that each answer a narrower question, all internally linked. Instead of one thin page per keyword, you build an interconnected map of a subject.

This matters more in AI search than it ever did in classic SEO. When an engine decides whether to trust your site as a source, it weighs topical authority, how thoroughly and consistently you cover a domain. A scattered site with one post each on fifteen unrelated topics reads as shallow. A site with a pillar plus twelve interlinked spokes on one subject reads as an expert.

A simple cluster for "AI search optimization" might look like this:

• Pillar: What is AI search optimization (the broad guide)
• Spokes: How to get cited by ChatGPT · How AI Overviews choose sources · AEO vs traditional SEO · How to measure AI citations · Schema for answer engines

Each spoke answers a specific prompt. Each links up to the pillar and across to siblings. The cluster, not the page, is now your unit of competition. We build these maps for clients as part of SEO strategy.

Do keywords still matter in the age of AI answers?

Yes, more than the hype suggests. Keywords are still how you discover demand and still how Google ranks the pages that feed AI Overviews. The keyword has not died; it has been promoted into a building block of a larger structure.

Here is the honest split. Google's AI Overviews cite the brands' own websites a large share of the time, so optimising your own pages still pays directly. ChatGPT leans more on third-party sources like Reddit and Wikipedia, so there you also need brand mentions and digital PR off your own site. Keywords anchor the first; entity presence wins the second.

The table below shows how the discipline evolves rather than disappears:

Dimension	Classic keyword research	Keyword research for AI search
Unit of research	Short keyword (2–4 words)	Full prompt / question (12+ words)
Source of language	Keyword tools, search volume	GSC questions, Reddit, sales/support transcripts
Success metric	Rankings & clicks	Citations, answer inclusion, AI-assisted traffic
Content unit	One page per keyword	Topic cluster (pillar + spokes)
Page structure	Keyword in title & body	Answer-first, question H2s, tables, schema
Opportunity sizing	Search volume	AI traffic potential (citation likelihood)
Off-site lever	Backlinks	Backlinks + brand mentions / entity presence

The skills carry over. The targets and structure are what change.

How do you judge AI traffic potential?

Search volume sized the old opportunity. For AI search you need a fuzzier but more honest measure: AI traffic potential, the likelihood that answering a prompt earns you a citation and a qualified visit or recommendation. Judge it on four factors:

1. Citation gap. Run the prompt through ChatGPT, Perplexity, and Google AI Overviews today. Who gets cited? If the current sources are weak, generic, or outdated, there is room. If three authoritative competitors own every answer, weigh the cost.
2. Business potential (0–3). Same gate as always. A prompt where your service is the natural answer scores a 3. A high-volume prompt with no path to revenue scores a 0, skip it. We care about revenue over vanity metrics, and so should your research.
3. Answer suitability. Can you produce a genuinely better, more extractable answer than what is cited now, with original data, a clear table, or real expertise? If not, move on.
4. Entity fit. Does the prompt sit inside a cluster you can own, reinforcing your topical authority, or is it a one-off?

A quick way to prioritise: AI Priority = (Citation Gap × Business Potential) ÷ Effort to Out-Answer. High gap, high business fit, low effort goes first.

How do you prioritize topics for both Google and AI?

You do not run two research projects. You run one and structure the output to serve both. Here is the workflow:

• Seed with prompts, not fragments. Start from real questions you mined.
• Cluster by topic. Group prompts into pillars and spokes so one cluster proves authority on a subject.
• Qualify each cluster twice. Score classic traffic potential and business potential, then layer on AI traffic potential (citation gap + answer suitability).
• Map one cluster to one structure. Pillar page plus spokes, internally linked.
• Structure every page for retrieval. Answer-first in the first 100 words, question-style H2s, 3–4 sentence paragraphs, a table or list, and Article plus FAQPage schema so engines can lift clean passages.

This dual structure is exactly what an SEO audit should check for, and what a done-for-you SEO programme should produce month after month.

Where should you start?

Pick one cluster that matters to revenue. Mine 20–30 real prompts around it from Search Console and Reddit. Map a pillar and five spokes. Check who AI cites today for each, and write the answer that deserves the citation. One well-built cluster will teach you more than a spreadsheet of 500 keywords ever did.

Want a research plan built around prompts, clusters, and AI traffic potential for your business? Book a free strategy call and we will show you the exact topics worth owning, so you get found by Google and recommended by AI.