How to Do Keyword Research for GEO: Finding the Prompts People Ask AI

ChatGPT alone has over 1 billion weekly active users, officially confirmed by OpenAI on August 6, 2026 (Source: OpenAI, via TechCrunch). Every one of those sessions runs on a question typed or spoken in plain language, not a keyword. If your content strategy still starts in a keyword tool built for Google's ten blue links, you are researching for a search engine that is no longer where a growing share of your buyers start looking.

Keyword research for GEO is not "keyword research, but for AI." It is a different research question. Traditional keyword research asks: which search terms have volume and low difficulty. GEO research asks: which questions do people actually put to ChatGPT, Perplexity, Copilot, and Google's AI Overviews, and what does an AI model need to see before it cites you in the answer. This guide walks through how to do that research properly, using methods that hold up, not a renamed version of your old process.

Why is keyword research for GEO a different job than traditional keyword research?

Traditional keyword research matches a search term to a ranked list of ten results, and you can pull volume and difficulty for that exact phrase from Google Keyword Planner, Ahrefs, or Semrush. GEO research matches a question to a synthesized answer, and the phrasing that actually gets typed into ChatGPT or spoken to Copilot is longer, more specific, and usually has no reliable volume data attached to it at all, because these platforms do not publish query-level search volume the way Google does.

The two also measure success differently. A keyword ranks or it does not, at a position you can screenshot. A prompt either gets your brand mentioned inside an AI-generated answer or it does not, and there is no position 4 or position 9 to climb toward. That changes what "targeting" a query even means: you are not chasing a ranking spot, you are trying to become one of the two to seven sources a model pulls from when it builds that specific answer.

The practical upshot: one seed topic in GEO research is worth more than one seed keyword in traditional research, because a single question can fan out into a cluster of sub-searches inside the model itself, most of which you will never see in any tool.

What makes a query worth targeting for AI search?

Not every question is a good GEO target. Based on what actually shows up in AI Overviews, per the query-length and question-format pattern documented in our AI Overviews guide, longer, more conversational queries and direct questions ("how does X work," "what is the difference between X and Y," "is X worth it") trigger AI-generated answers far more often than short, navigational searches ("brand name login"). If a query is two words and has an obvious single right answer, it is not a GEO opportunity. If it takes a paragraph to answer well, it usually is.

That reframes the starting question for research. Instead of "what has volume," ask "what would someone need explained to them before they'd trust a recommendation here." Comparison questions, cost questions, how-it-works questions, and "is it worth it" questions are where AI answers get built, because those are the questions a single search result rarely answers completely on its own.

Step 1: mine your existing Google Search Console data first

Before opening a new tool, look at data you already have. Google Search Console added an AI Overviews filter under the Search appearance report. It only shows click data, not what happened inside the AI Overview itself, but it is free and it is the most reliable starting signal for which of your existing queries already trigger an AI-generated answer.

Cross-reference that against your full Performance report, filtered to queries with four or more words and a question word (who, what, when, why, how, is, does). Those are your People Also Ask and AI Overview candidates hiding in data you already collect. Sort by impressions, not clicks. High impressions with low clicks on a long question-style query is a strong sign an AI Overview or featured snippet is absorbing the click before it reaches you.

Step 2: layer in rank-tracking tools that flag AI Overview presence

GSC only tells you about queries you already rank for. To find AI-Overview-triggering queries you have not touched yet, you need a tool that checks AI Overview presence for a broader keyword list. Ahrefs added AI Overview tracking into its existing rank-tracking suite through its 2025 to 2026 product updates: for every tracked keyword it can now show whether Google is surfacing an AI Overview for that query, and where detectable, whether a given domain is cited inside it (Source: this is the same tool covered in our best GEO tools roundup). That data is Google-only; it says nothing about ChatGPT, Perplexity, or Copilot, so treat it as one input, not the full picture.

Run your existing seed keyword list, and any adjacent questions your sales and support teams hear, through a tool like this before you write anything. Sort the output by AI-Overview presence rather than by search volume. A keyword with modest volume that reliably triggers an AI Overview is often a better GEO target than a high-volume keyword that never does.

Step 3: mine reddit, forums, and review sites for how people actually phrase questions

Search volume tools describe how people type into a search box. They do not describe how people talk to a chatbot, and the two are not the same sentence. The fastest way to hear real phrasing is to search Reddit, niche forums, and review sites (G2, Capterra, industry-specific communities) directly for the topic, and read the actual questions people ask each other, not just the ones a keyword tool infers from Google click data.

This is research into phrasing, not a bet on Reddit as a citation source. That distinction matters right now: Reddit's own share of ChatGPT Search citations fell from an average of 3.83% between July 18 and August 7, 2026 to just 0.52% by mid-August, an 86% relative drop tied to a change in how ChatGPT Search constructs its internal fan-out queries (Source: Promptwatch data, independently reported by Search Engine Land and Semrush, covered in full in our Reddit citation drop analysis). Reddit threads are still a genuinely useful place to read how real people phrase a question. They are a much weaker bet as a place to plant content and expect an AI model to cite it back.

Look for the exact words people use to describe their problem before they know the vocabulary your industry uses. That gap, between how a prospect describes something and how your industry names it, is usually where the highest-value untargeted prompts live.

Step 4: use fan-out thinking to expand one seed topic into a full query set

Once you have a seed topic, expand it the way an AI model would expand it internally: break it into the sub-questions someone would need answered along the way to a full answer. Take "keyword research for GEO" as an example. A model answering that question is likely to pull in sub-questions about what tools to use, how it differs from traditional keyword research, whether search volume data exists for AI queries, and how to validate a prompt list once you have one. Each of those sub-questions is a legitimate H2 or FAQ entry in the content you build, and each is a query worth checking individually in Step 1 and Step 2's tools.

This is manual work; there is no tool on the market today that reliably replicates a specific model's internal fan-out logic from the outside. Google names the mechanism in its own documentation without publishing the exact sub-queries it generates, and the same is true for OpenAI, Perplexity, and Microsoft. Doing this expansion by hand, grounded in what a genuinely thorough human answer would need to cover, is currently the closest working substitute.

Step 5: classify what you have found before you write anything

By this point you should have a messy list: existing GSC queries, AI-Overview-flagged keywords from a rank tracker, raw phrasing pulled from forums, and a set of fan-out sub-questions you wrote by hand. Sort that list by intent before prioritizing it. Our free Keyword Intent Classifier groups a pasted list into informational, commercial, transactional, and navigational buckets using the same signal categories search engines use to match intent, which is a fast way to see where your list actually sits before you commit writing time to it.

Informational and comparison-style questions are where AI Overviews and chat answers show up most often. Transactional and navigational queries rarely trigger a generated answer at all, because there is usually one obvious destination (your pricing page, a specific login) and no real question to synthesize an answer to. Weight your content calendar toward the informational and comparison end of that list.

Step 6: validate the highest-priority prompts by actually asking the AI

Run your top 15 to 20 candidate prompts through ChatGPT, Perplexity, and Copilot directly, using the exact phrasing you found in Steps 1 through 4. Record, for each: does an AI Overview or generated answer appear at all, does your brand or a competitor get mentioned, and which sources does the answer cite. This is slower than pulling a report, but it is the only way to see the actual output, since none of the tools above show you the finished answer, only signals that one might exist.

Repeat this check periodically rather than once. Query fan-out behavior changes when platforms update it, as the August 2026 ChatGPT Search shift showed, and a prompt that produced no AI answer in one testing round can start triggering one a few weeks later with no announcement.

Where does this fit with the rest of your GEO work?

Keyword research for GEO is the input, not the whole strategy. Once you have a prioritized, intent-classified list of real prompts, the work moves into content structure, entity clarity, and technical setup, covered in depth in our guides on how to rank in ChatGPT and the broader GEO vs SEO distinction. Research tells you what to build. It does not, on its own, get you cited.

Frequently asked questions

Is there a keyword volume tool for AI search prompts?

Not a reliable one. Google, OpenAI, Perplexity, and Microsoft do not publish query-level search volume for the way people phrase questions to their AI products, unlike Google's traditional Keyword Planner data for classic search. The closest available signals are AI Overview presence flags from rank-tracking tools like Ahrefs, and Google Search Console's AI Overviews click data, neither of which is a true volume figure.

Does traditional keyword research still matter if I'm doing GEO?

Yes. Traditional keyword research and GEO prompt research overlap heavily, since many AI Overviews are triggered by queries that also have a traditional search volume. Start from your existing keyword list and Search Console data, then expand it with the AI-specific methods in this guide rather than replacing one process with the other.

Should I target long-tail conversational phrases instead of short keywords?

For GEO specifically, yes, weight your research toward longer, question-format phrases. Short, navigational queries rarely trigger an AI-generated answer because there is usually one obvious destination and nothing to synthesize. Comparison, cost, and "how does X work" style questions are consistently where AI Overviews and chat answers appear.

Is Reddit still worth researching for GEO?

Worth researching for phrasing, yes. Worth betting on as a citation source right now, less so. Reddit's share of ChatGPT Search citations fell roughly 86% in a matter of days in August 2026 after a change to ChatGPT's internal query-handling. Reading Reddit threads to learn how people phrase a question is still useful; publishing content there and expecting it to get cited back is a weaker bet than it was a month earlier.

What is query fan-out and why does it matter for keyword research?

Query fan-out is the mechanism, named directly in Google's own May 2026 documentation, by which an AI system breaks one query into several related searches before assembling a single answer. It matters for research because it means the sub-questions a model generates internally, which you cannot see directly, often decide what gets cited, not just the exact phrase you originally targeted. Manually expanding a seed topic into its likely sub-questions, as described in Step 4, is the closest available substitute for seeing that process.

How often should I redo GEO keyword research?

Quarterly at minimum, and immediately after any documented platform change, such as the August 2026 shift in ChatGPT Search's fan-out behavior. AI Overview and chat answer behavior changes without advance notice, and a prompt set that was accurate three months ago can be stale without any change on your end at all.

Sources: Google Search Central, TechCrunch, Search Engine Land, Semrush, Ahrefs