How to Do Keyword and Topic Research for AI Search and GEO

Keyword research has long been a foundational part of search engine optimization. It helps marketers understand what their audiences search for, how often they search for it, and which terms may present opportunities to attract organic traffic. However, the growth of AI-powered search experiences has changed how people find information and how search platforms interpret their requests.

AI Search, expands keyword research beyond short phrases and conventional search queries. It considers natural-language questions, conversational prompts, named entities, and the additional searches that AI systems may perform when constructing an answer. An effective research process for AI search should therefore combine established SEO methods with newer techniques designed to reveal how users and AI assistants explore a topic.

Traditional search engines have historically encouraged users to express their needs as compact keyword phrases. Someone researching business software, for example, might search for “best accounting software” or “accounting software pricing.” These phrases remain valuable, but they represent only part of how people now search.

Users often interact with AI assistants in a more conversational manner. Instead of entering a short phrase, they may ask, “What is the best accounting software for a small consulting firm with five employees?” They might then refine the request by asking about integrations, pricing, security, or implementation. These longer prompts contain more context and reveal more about the user’s situation, goals, and decision criteria.

AI search systems may also break a question into several related searches before generating an answer. This process can expose subtopics, comparisons, entities, and supporting questions that would not necessarily appear in a conventional keyword report. As a result, AI SEO research must examine not only what users type, but also how an AI system may interpret, expand, and investigate their requests.

This does not mean traditional keyword research is obsolete. Search volume, ranking difficulty, intent, and existing search results remain useful sources of insight. The difference is that these inputs should now form the foundation of a broader topic-research process rather than represent its final output.

Begin with traditional SEO keyword research. This is the process of identifying the words and phrases people enter into search engines when looking for information, products, services, or solutions. Its purpose is to reveal audience demand, understand search intent, assess competition, and identify subjects that deserve coverage.

Tools such as SEMRush, MOZ, and Ahrefs can help you discover relevant terms and evaluate metrics such as estimated search volume, keyword difficulty, ranking results, and related queries. Start with broad terms that describe your organization, products, services, customer problems, and areas of expertise. Expand those seed terms into more specific variations and group related phrases into recognizable topics.

Avoid limiting the research to generic keywords. Include known entities associated with the subject or company, such as brands, products, organizations, locations, and people. Entities are especially important in AI search because they help systems understand the relationships surrounding a topic. If an audience frequently connects a product category with certain companies, experts, technologies, or standards, those relationships should be represented in the research.

The result of this first step should be a strong baseline of conventional search terms and relevant entities. Although this list will resemble the output of a standard SEO project, it will also supply the seed terms needed for the conversational and AI-focused stages that follow.

The next step is to identify conversational terms. These are natural-language searches that resemble the way a person would speak to an expert or AI assistant. They are usually longer and more specific than conventional keywords, and they often take the form of complete questions, requests, or descriptions of a particular situation.

For example, “CRM software” is a traditional keyword, while “What CRM should a small sales team use if it needs email automation?” is a conversational query. The second version provides context about the user, the desired feature, and the decision being made. This makes it particularly useful for GEO research.

Use the terms gathered during traditional keyword research as starting points for discovering these questions. A tool such as TermSuggest (I built this, LOL) can help expand seed terms into relevant questions and other natural language searches. Run several variations of each important term because even small changes in wording can reveal different needs, concerns, and stages of the customer journey.

Google Search Console can provide another valuable source of conversational queries. Export all available queries for the website rather than looking only at the highest-volume terms. Longer searches may receive relatively few impressions individually, but they can reveal specific audience needs that broader keywords conceal.

After exporting the queries, sort them in descending order by character length. This brings longer phrases to the top of the sheet and makes likely conversational searches easier to find. Length alone does not prove that a query is conversational, but it is a useful initial filter.

You can then use a LLM to classify the exported searches. Ask the model to determine whether each query is conversational based on factors such as natural sentence structure, the presence of a question, the inclusion of personal context, and the expression of a specific goal. Human review is still important, especially when a short query has conversational intent or a long query is simply a string of unrelated keywords.

At the end of this step, you should have a set of real or plausible audience questions that can be used to investigate how AI assistants approach your topics.

A query fan-out is the collection of related searches or subqueries that an AI system may use to investigate a user’s request. Instead of treating a prompt as a single keyword, the system may divide it into several information needs and search for supporting facts, comparisons, definitions, examples, or sources.

Consider a user who asks, “What is the best project-management platform for a remote creative agency?” An AI assistant may explore separate questions about collaboration features, pricing, integrations, customer reviews, file sharing, and support for distributed teams. These fan-out queries help reveal the supporting information that may influence the assistant’s final response.

Use the conversational terms collected in the previous step as prompts in AI assistants. Test them individually and record the query fan-outs associated with each prompt. Running prompts across multiple assistants may uncover meaningful differences in how each system interprets and investigates the same topic.

Log out of your account before conducting this research whenever possible. Account history, saved preferences, location information, or earlier conversations may personalize the response and affect the results. A logged-out session will not eliminate every source of variation, but it can reduce the influence of account-level personalization and produce a more neutral research sample.

Browser extensions can help expose and extract the fan-out queries generated during an AI search. Options include the ChatGPT Search Fan-Outs Chrome extension and the RainMojo Query Fan-Out Chrome extension. Use the appropriate extension for each assistant and capture the fan-outs produced for every prompt in your research set.

AI-generated results can vary between sessions, so avoid treating a single run as definitive. Repeating important prompts can reveal which subqueries appear consistently and which are occasional variations. The recurring fan-outs are often the strongest candidates for content planning because they suggest stable relationships within the topic.

Once the traditional keywords, conversational terms, and query fan-outs have been collected, combine them in a single spreadsheet. A unified dataset makes it easier to remove duplicates, identify patterns, group related concepts, and find gaps between existing content and the information audiences or AI systems may need.

Add a column named “type” and classify every term as “traditional SEO,” “conversational,” or “Query Fan Out.” This classification preserves the role each term played in the research process. Traditional SEO terms reveal established search demand, conversational terms capture more detailed audience needs, and query fan-outs show how AI systems may expand those needs into supporting searches.

Add another column named “source” to record where each term originated. Source values might include SEMRush, MOZ, Ahrefs, Google Search Console, ChatGPT, or Claude. Recording the source makes the research easier to audit and update. It also allows you to compare patterns across platforms instead of assuming that every tool or assistant views a topic in the same way.

Maintain separate rows when identical terms come from meaningfully different sources if source comparison is important to the analysis. Otherwise, duplicates can be consolidated while preserving all relevant source information. The goal is to produce an organized topic dataset rather than an unnecessarily large list.

After classification, group closely related terms into topic clusters. Each cluster should represent a central subject and its supporting questions, entities, comparisons, and subtopics. These clusters can guide briefs, page updates, FAQ sections, original research, product comparisons, and other content intended to serve both human readers and AI search systems.

The completed spreadsheet should not be treated as a list of phrases to repeat throughout a page. Instead, use it as a map of the subject. Identify the main question behind each cluster, the audience context revealed by conversational searches, and the supporting information suggested by query fan-outs.

Strong AI SEO content should answer the primary question clearly while also covering the related details a reader or AI system may need to evaluate the answer. Relevant entities should be identified accurately, comparisons should explain their criteria, and important claims should be supported with credible evidence. Content structure should make relationships between the main topic and its subtopics easy to understand.

Prioritization should reflect more than conventional search volume. A low-volume conversational query may still matter if it represents a valuable customer need or repeatedly causes AI assistants to investigate the same supporting questions. Combining traditional SEO data with conversational and fan-out insights makes it possible to balance measurable demand with emerging AI-search behavior.

~

Keyword and topic research for AI search begins with familiar SEO practices, but it should not end there. Traditional keywords establish the demand surrounding a subject, conversational queries reveal how people express detailed needs, and query fan-outs show how AI assistants may expand those needs while building an answer.

A practical workflow is to collect conventional keywords and relevant entities, mine natural-language questions, use those questions to uncover AI query fan-outs, and combine everything in a classified spreadsheet. The resulting dataset offers a richer picture of the topic than search volume alone can provide.

AI search will continue to evolve, along with the systems and signals that shape generated answers. Organizations do not need to predict every change to make progress. By studying how audiences ask questions and how AI assistants investigate them, marketers can create content that is clearer, more complete, and more genuinely useful. That is not only a sound AI search strategy; it is also the foundation of content worth discovering.

Context Window: The amount of text an AI model can consider at one time.

Learn More

One more thing: AI is only as good as its operator, and if you are reading this newsletter, you’re better than most!

Till next time,

PS: Let me know what you think of this issue, or anything else here: [email protected]