Keyword Research for AI Search: How It Works When There's No Keyword to Rank For

Google Keyword Planner, Ahrefs and Semrush are built around a fixed idea: a discrete "keyword" with a monthly search volume, matched (loosely) against how people type into a search box. That model works because Google search behaviour is genuinely clustered - thousands of people type "google ads management cost" a month, in more or less that phrasing.

AI search doesn't cluster the same way. A person asking ChatGPT or Perplexity types a full sentence, often a genuinely unique one: "what should I budget for google ads if I run a small plumbing business in Perth". There's no keyword database that has that exact string with a volume attached, because it's rarely typed the same way twice. What repeats isn't the phrasing, it's the underlying question.

That's the shift: traditional SEO research asks "what terms do people type?" AI search research asks "what questions does my industry get asked, regardless of how they're phrased?"

• Google's "People also ask" boxes - the actual follow-up questions Google already associates with a topic, in real question form
• ChatGPT and Perplexity themselves - put your seed topic in and read the follow-up questions they suggest, or ask them directly what people commonly ask about it
• Reddit and Quora threads in your niche - both are heavily cited source pools for Perplexity and ChatGPT Search, and the phrasing in genuine questions there is a strong signal for how people actually ask
• Your own inbox and call notes - the questions prospects and clients actually ask before buying are the highest-quality signal available, and none of it shows up in a keyword tool
• Google Search Console's Queries report, filtered to who/what/why/how/when - your existing impression data already contains the question phrasing Google associates with your pages

The goal of all five is the same: build a list of real questions, not typed search terms. A single question can absorb dozens of keyword-tool variants once you're writing to answer it directly rather than to rank for a phrase.

Question-mapping tools

AlsoAsked and AnswerThePublic are both built around this exact gap: instead of a keyword-and-volume list, they map out the tree of related questions people ask around a seed topic, pulled from Google's autocomplete and "People also ask" data. Ahrefs and Semrush also have a "Questions" filter inside their standard keyword tools, which is a reasonable proxy for the same thing if you already pay for one of them.

Google Search Console

Search Console's Performance report lets you filter Search appearance by AI Overviews, showing exactly which of your existing queries triggered an AI Overview. High impressions with a low click-through rate on a query is the AI Overviews fingerprint: Google is answering the question directly on the results page, and your listing is being seen but not clicked. That's a clear signal of which questions are worth writing a direct, complete answer for.

Keyword Planner and volume tools, used differently

Standard volume tools aren't useless here, they're just answering a different question than "should I write this?" Use them to size a topic (is this a niche question or a common one?) rather than to find the exact phrase to target. If a seed term has meaningful volume, that's a signal the underlying question is common enough to be worth a dedicated answer, not that you should optimise for that literal string.

Finding the questions is half the job. The other half is structuring the answer so an AI system can lift it cleanly, which is a separate discipline covered in full in what GEO actually involves: answer-first paragraphs, one question per FAQ entry, and enough topical depth that your site reads as a credible source on the subject rather than a single page that happens to mention it.

A practical workflow: pull 15-20 real questions using the methods above, group them by the underlying question rather than by phrasing, then write one direct, complete answer per question, either as its own section or as part of a broader article. Each answer should stand alone: an AI system will lift a paragraph or an FAQ entry out of context, so it has to make sense without the rest of the page around it.