Skip to content
Chalk Labs / AI SEO
How-to guide

How to Pick Keywords for AI SEO

By Rahil Jain, founder of Chalk LabsUpdated How we research

The short answer

To pick keywords for AI SEO, start with the questions your buyers ask, expand each into the related searches an AI runs behind it, then keep only topics where a lookup is likely, you can offer something original, and a buyer would act on the answer. Drop questions an AI answers from memory. Score what is left before you write.

  • Start from the buyer's question, then list the related searches an AI would run to answer it.
  • Many AI answers involve no web search, so skip questions an AI can answer from memory.
  • Nobody can see true AI prompt volume. Use Google volume as a rough guide and buyer value as the tiebreaker.
  • You do not need a top-10 rank to be cited, but your page must be indexed and eligible to show.
  • Score each target on buyer value, lookup likelihood and what you can add, then write the top ones first.

How is picking keywords for AI SEO different from classic keyword research?

The unit changes from a short phrase to a question with a job behind it. Classic research asks which phrase has volume and how hard it is to rank. AI SEO asks which buyer question will make an assistant search, and whether your page can be the source it picks.

Two things carry over. You still need pages that are crawlable, indexed and useful. Google says AI Overviews and AI Mode have no extra technical requirements beyond being indexed and eligible to show in Search with a snippet 2. And you still need to pick topics your business can speak about with authority. For the wider picture, read what AI SEO is.

Do it step by step

0 of 7 done0%

  1. Pull real questions from sales calls, support tickets, enquiry forms and reviews. Write each as a full sentence, in the buyer's words.

    Tip: Ask your sales team for the three questions prospects ask before they buy.

  2. Paste each question into one or two assistants and note the sub-topics and options named. Use a fan-out tool to see the related searches where you can.

  3. In a sheet, put each seed in column A and every related search in the rows below. Merge near-duplicates so each row has one clear intent.

  4. For each row, ask whether the assistant answers it fully and correctly without searching. Mark rows as memory, lookup or unsure.

    Tip: Mark pricing, local, comparison and recent-change questions as lookup by default.

  5. Check Google volume for the parent topic in a keyword tool or Search Console. Record it, but do not delete low or zero rows on volume alone.

  6. Give each row 0 to 3 on each of the three scores and total them. Be strict on 'what you add': a 3 needs first-hand data, a process or original work.

  7. Merge rows with the same intent into one page plan. Take the top 10 totals, write them first, and review the list again after 60 days.

Your ticks are saved in this browser only.

Copy-ready templates

Keyword scoring sheet (CSV)

text
seed_question,fan_out_search,type,answered_from_memory,google_volume,buyer_value_0_3,lookup_0_3,what_we_add_0_3,total,page_plan
"How much does AI SEO cost?","ai seo pricing",price,no,unknown,3,3,3,9,"Pricing page with dated numbers"
"How much does AI SEO cost?","what is ai seo",definition,yes,unknown,1,0,1,2,"Skip"

Total = value + lookup + add. Sort by total, highest first. Keep volume as a note, not a filter.

Prompt to expand a seed into related searches

text
I run [business type] in [place]. A buyer asks: "[seed question]".
List the 8 to 12 separate web searches you would run to answer this well.
For each, say whether it is about price, comparison, how-to, local options or a definition.
Mark which ones you could answer fully from memory without searching.

This is a guide to what an assistant might search, not a record of what it did. Check against a fan-out tool when you can.

Where should you start: the keyword or the buyer's question?

Start with the question, written the way a buyer would say it, then work outward. Take the seed “how to choose keywords for AI search optimization”. An AI answering it may search for keyword research methods, tools, how AI search differs from Google, and examples. Those related searches are the query fan-out: the extra searches a system runs to build one answer.

Google confirms that AI Overviews and AI Mode may run several related searches across subtopics and data sources 2. Its AI optimization guide defines fan-out queries as a set of concurrent, related queries generated by the model 1. Ahrefs found ChatGPT averaged about 1.6 fan-out queries on the responses that searched at all 3. So each seed gives you a small cluster, not one keyword.

List 10 to 20 seed questions from sales calls, support tickets and your own enquiry form. Then use the fan-out research guide or the Query Fan-Out Explorer to see the searches behind each one.

What do you do about search volume when AI prompt volume is hidden?

Use Google search volume as a rough guide for the parent topic and rely on buyer value to break ties. Ahrefs states that, apart from Microsoft Copilot, no major AI platform shares what users ask, and that its AI adjusted volume is a proxy to be used directionally 5. A precise AI volume for a long prompt is an estimate.

So do not reject a question because a tool shows zero volume. Long, specific prompts rarely show volume anywhere, yet they are how people talk to assistants. Ask instead whether a person with money and a need would type it. A question from a prospect who is choosing a supplier is worth more than a vague one with big numbers.

Which questions will an AI answer without looking anything up?

Simple, stable definitions and well-known facts are usually answered from the model's memory, and no page gets cited. Ahrefs found fan-out queries in only 14.4% of logged-out ChatGPT responses it tracked in August and September 2026 3. Our prompt list guide sets this beside Semrush's different figure; either way, many answers involve no lookup.

A lookup is more likely when the answer depends on fresh, local, priced or compared information. In a 20,000-prompt US test, where Ahrefs does not define "official", official sites got 63.4% of citations on fact questions such as pricing and features, and 46.4% on comparison questions 3. That is a reason to publish your own pricing, specs and comparison pages: buyers ask those questions, and the AI has to fetch the answer.

How do you judge whether you can win a keyword?

Ask whether you can add something the top results lack, not whether you can outrank them. In a study published in March 2026 (863K keywords, about 4M cited URLs), Ahrefs found about 38% of AI Overview citations came from the first 10 results for the same keyword, down from about 76% in its July 2025 study. It linked the drop to fan-out searches pulling from other results 4. A separate 2025 study of 15,000 long-tail prompts found about 12% of assistant citations matched Google's top 10, with Perplexity highest at 28.6% 6.

Treat those as vendor data from specific samples, not laws. They suggest that a page ranking 15th for a narrow sub-question can still be cited. They do not mean rankings are irrelevant. Page-one pages still supply many citations.

Google's own self-check questions help here. It asks whether content gives original information, research or analysis, and goes beyond the obvious 7. If you have first-hand data, a real process, pricing or photos, you have an edge. If you would only reword the top three results, pick another keyword.

What should you do with each type of keyword?

Sort every candidate into one of five types and apply the matching action. This table is our decision rule, not an industry standard.

Keyword types and the action for each
TypeExampleAction
Simple definitionWhat is a meta description?Skip, or answer briefly inside a bigger page
Buyer problemWhy are my leads dropping since AI Overviews?Write a full answer page
ComparisonFreelancer vs agency for AI SEOWrite a fair comparison with a table
Price or specHow much does AI SEO cost?Publish your own numbers with a date
Local or nicheAI SEO for dentists in LeedsWrite only if you serve that place or field

How do you score and rank your keyword list?

Give each target three scores from 0 to 3 and add them: buyer value, lookup likelihood, and what you can add. Write the highest totals first. A 9 is a clear start. A 4 or less goes to the back of the queue.

The next example is invented to show the method. It follows a small company that makes custom pool covers. Real scores will differ with your market.

Scoring example (illustrative): custom pool cover maker
  • Keyword | Value | Lookup | Add | Total
  • custom pool cover cost for odd-shaped pool | 3 | 3 | 3 | 9
  • safety cover vs mesh cover for toddlers | 3 | 2 | 2 | 7
  • pool cover installers near me | 3 | 3 | 1 | 7
  • what is a pool cover | 1 | 0 | 1 | 2

What goes wrong when you turn every variation into a page?

You risk a spam problem and a thin site. Writing a separate page for every way people might search can look smart, but Google's AI optimization guide says doing it mainly to manipulate rankings or AI responses breaks its scaled content abuse policy 1. Group near-identical questions on one page with a heading for each, and give a new page only to a different intent.

It is also fine to skip a keyword. If you cannot add anything original, or the question is answered from memory, leave it out.

Frequently asked questions

Do I still need keyword research for AI search?

Yes, but the unit is the buyer question and its related searches, not a single phrase. You still need pages that are indexed and eligible to show in Search with a snippet, since Google lists no extra requirement for AI features 2. Volume becomes a rough guide, not the deciding number.

Can I find the volume of ChatGPT prompts?

Not directly. Ahrefs says no major platform except Microsoft Copilot shares what users ask, and its own AI volume is a proxy built from Google data, best used directionally 5. Treat any exact prompt volume you see as an estimate.

Do I have to rank in the top 10 to be cited by AI?

No, but ranking helps. Ahrefs' March 2026 study found about 38% of AI Overview citations came from the top 10 4, so many citations came from lower pages. Your page still needs to be indexed and eligible to appear 2.

Should I write a page for every fan-out query?

No. Google's guide treats this as scaled content abuse when done mainly to influence rankings or AI answers 1. Combine near-identical questions on one page and add a new page only for a different intent.

How many keywords should a small business target first?

As our rule of thumb, ten well-chosen questions is a sound start. Pick the highest-scoring ones, write each page to be the clearest answer on the web, then review after about 60 days (also our rule of thumb). A small set done well beats a long list of thin pages.

About the author

Rahil Jain

Rahil Jain is an AI SEO (answer engine and generative engine optimisation) consultant and co-founder of Chalk Labs.

LinkedIn · Method and credentials

Sources

  1. Optimizing your website for generative AI features on Google Search · Google Search CentralOpened October 20, 2026. Supports: Definition of fan-out queries; separate content for every query variation to manipulate rankings or AI responses violates the scaled content abuse policy (updated 2026-07-10).
  2. AI features and your website · Google Search CentralOpened October 20, 2026. Supports: No extra technical requirements beyond being indexed and snippet-eligible; AI Overviews and AI Mode may use query fan-out (updated 2025-12-10).
  3. How ChatGPT's query fan-out works · AhrefsOpened October 20, 2026. Supports: 14.4% of 66.6 million logged-out ChatGPT responses had fan-out queries, about 1.6 each; in a 20,000-prompt US test, official sites (not defined by Ahrefs) got 63.4% of citations on fact questions and 46.4% on comparisons (August to September 2026).
  4. AI Overview citations and the top 10 · AhrefsOpened October 20, 2026. Supports: Study published 2 March 2026: about 38% of AI Overview cited URLs also appeared in the first 10 results (863K keywords, about 4M URLs), down from about 76% in July 2025; linked to fan-out.
  5. How Ahrefs estimates AI search demand · AhrefsOpened October 20, 2026. Supports: No major AI platform except Microsoft Copilot shares query data; AI adjusted volume is a proxy for directional use.
  6. AI search overlap study · AhrefsOpened October 20, 2026. Supports: On 15,000 long-tail queries, about 12% of assistant citations matched Google's top 10; Perplexity highest at 28.6% (published August 2025).
  7. Creating helpful, reliable, people-first content · Google Search CentralOpened October 20, 2026. Supports: Self-assessment questions on original information, research or analysis and insight beyond the obvious (updated 2026-10-05).
How this page was made. Built on Rahil Jain's AI SEO method and drafted with AI research assistance. Every statistic links to the page we read it on (opened October 20, 2026 or nearby). Published October 20, 2026; last updated October 20, 2026. Chalk Labs sells AI SEO services, so we have a commercial interest in this topic. Spotted an error? Tell us and we will fix it.