How to Research Prompts for AEO

If you want your website to appear in answers from ChatGPT, Gemini, Claude and other AI platforms, you need to understand which questions your potential customers ask.

That is not just general advice. In my own research into 1,680 AI answers, how closely a page matched the question was the strongest page-level factor associated with being cited. When I compared pages from the same website, the pages that matched the question more closely were cited considerably more often.

In other words: before you can write the answer AI assistants cite, you need to know which question is being asked. That is where prompt research comes in.

With traditional SEO, we research keywords to understand what people search for on Google. With Answer Engine Optimization (AEO), we also want to understand the complete questions people ask AI assistants. (If you are unsure how AEO relates to GEO, I explain the difference between AEO and GEO separately.)

But how do you find those questions? And how do you decide which ones are worth targeting when there is no reliable search volume data for individual AI prompts?

In this article, I explain how I approach prompt research, which sources and tools you can use, and how to turn your findings into useful content. Along the way, I share more results from the study, including the 1,830 web searches the AI platforms ran to find their sources.

Why does prompt research matter for AEO?

Here is how I reached that conclusion, and what it does and does not show.

For the study, I collected 1,680 answers from ChatGPT, Claude, Gemini and Google AI Mode. In one of the follow-up analyses, I compared pages that AI assistants cited with other pages from the same websites that they retrieved but did not cite.

I measured how well each page matched the question, using semantic similarity between the question and passages from the page.

The difference was substantial. In the statistical model, a one-standard-deviation increase in the match score was associated with roughly 1.7 to 2.2 times the odds of being cited, depending on which websites were included.

Other factors often mentioned in AEO and GEO guides, such as structured data, FAQ sections, author information and content length, did not show the same consistent advantage when I compared pages on the same website.

These are associations, not proof that changing a page will get it cited, and the result was clearest for ChatGPT. I go into the limits at the end of this article.

Still, the findings suggest something useful: before worrying about special AEO formatting techniques, make sure your content answers the questions people actually ask. And to do that, you first need to find those questions.

Prompts vs. keywords: what is the difference?

A keyword is usually a relatively short phrase. A prompt can be a complete question with considerably more context.

For example, someone might search Google for best CRM software. In ChatGPT, the same person might ask:

“What's a good CRM for a small consulting business with five employees? We currently use spreadsheets and want something easy to implement without spending too much.”

Both concern CRM software, but the second question tells us much more about what the person needs: the type of business, the number of employees, the current problem and the requirements for a suitable solution. An AI assistant can use that information to recommend products that fit those circumstances.

A few more examples:

These are examples of possible prompts, not verified questions from real AI conversations.

Of course, people also type complete questions into Google, and plenty of ChatGPT users write short prompts. The distinction is not absolute. The important difference for your research is that AI conversations often include extra context, conditions and follow-up questions.

Does that mean keyword research is becoming less important?

No. I would actually use your existing keyword research as one of the starting points for AEO. As you will see in the section on query fan-out, AI assistants turn questions into keyword-style searches themselves.

Keyword research shows which topics people are interested in. With tools such as Semrush, you can also investigate search volumes, related questions and commercial intent. What you cannot assume is that a keyword with 10,000 monthly Google searches also generates 10,000 prompts in ChatGPT. They are different platforms with different user behaviour.

I see keyword research and prompt research as complementary. Keyword research helps identify topics with existing demand. Prompt research helps you understand the specific questions, problems and decisions behind those topics.

Where can you find the prompts people use?

This is probably the most difficult part of AEO prompt research. Unlike Google searches, where tools estimate the popularity of individual keywords, we have much less visibility into the exact questions people submit to AI assistants.

Some AI visibility platforms now offer prompt databases and estimated AI search demand. These estimates depend on the data available to the provider and its methodology, so they should not be treated as complete counts of real prompts. That is why I recommend combining several sources rather than relying on one tool.

1. Questions from your own customers

If you sell products or services, your customers are probably your most valuable source. Look at the questions people ask during sales calls, in support emails, through contact forms and in product reviews.

What do they want to know before buying? What makes them hesitate? Which problems do they mention again and again?

For example, a security system installer might often hear the question whether access control keeps working during an internet outage. That is a more valuable subject to research than a generic question about the history of access control systems.

The advantage of customer questions is that you already know they are relevant to your business.

2. Google Search Console

If your website already receives organic traffic, Search Console shows which questions bring people to it. Open the Performance report and look at the queries that generate impressions and clicks.

To see only the questions, add a query filter, choose Custom (regex) and paste:

^(what|how|why|which|when|where|who|is|can|should|do|does)\b

Long, specific queries are often worth a look as well. This regex shows queries of six words or more: ^(\S+\s+){5,}\S+$

For a website about running, this might surface queries such as:

• Can I run with knee pain?
• How often should beginners run?
• What running shoes are best for flat feet?

These are useful starting points. You can then investigate related questions and consider how someone would ask the same thing in an AI conversation. Remember that Search Console reports Google searches; it does not show what people ask ChatGPT or other AI assistants.

3. Keyword research tools

Tools such as Semrush help you discover additional questions and topics. I usually start with a broad keyword related to the products or services, then look for:

• Question-based searches.
• Long-tail keywords.
• Comparison searches.
• Problems people want to solve.
• Searches with requirements such as price, location or intended use.

For example, starting with access control systems might lead you to searches about installation costs, mobile credentials, cloud management and apartment buildings. These can become the basis for several AI prompts.

Semrush also offers a dedicated Prompt Research report with AI prompts, topics and estimated demand. Tools like this can save time, but I would still compare their suggestions with other sources before deciding which prompts matter to your audience.

4. Google autocomplete and People Also Ask

Google's autocomplete suggestions and People Also Ask questions show what people want to know about a subject. Start typing a topic into Google and look at the suggestions, and at the questions Google shows in the results. Tools such as AlsoAsked help you explore related questions more systematically.

These questions are especially useful for finding topics you had not considered. Someone researching electric cars might start with charging costs and discover related questions about battery life, home charging and electricity tariffs.

5. Reddit, Quora and specialist forums

Forums are valuable because people often explain their situation before asking a question. Instead of asking “best CRM”, a forum user might explain that they tried two systems, found them too complicated and now need something for a small team.

Look for recurring questions and problems. Pay attention to why people reject certain solutions, which alternatives they consider and which details they need before deciding. Not every forum question is also asked in ChatGPT, but these discussions are good evidence of real customer needs.

6. Use AI to generate more prompt ideas

Finally, you can ask ChatGPT, Gemini or Claude to help expand your list. Give the model information about your audience, products, services and common customer problems, then ask it to suggest realistic questions. For example:

I run a company that installs access control systems for apartment buildings. Our customers are property managers responsible for buildings with 50–500 units.

Generate 20 realistic questions these property managers might ask an AI assistant when comparing access control solutions.

Include questions about costs, installation, compatibility, security, everyday management and replacing existing systems. Group the questions by intent and avoid unnecessary variations of the same question.

This is a useful way to discover angles you had not considered. But an AI-generated prompt is not evidence that people actually ask that question. Treat these suggestions as ideas to investigate, not as verified demand.

Combining real searches and AI: the AI Prompt Finder

The free AI Prompt Finder combines sources 4 and 6. It collects real Google and Bing autocomplete searches for your topic and turns them into the questions people are likely to ask an AI assistant, grouped by intent. Every prompt shows whether it is based on a real search, and which one, or is a suggestion.

How to prioritise AEO prompts without reliable search volumes

Once you have collected questions from different sources, you will probably have more than you can realistically target. The next step is deciding which ones are most valuable.

With traditional SEO, search volume is often a major consideration. For AEO, it is more complicated: some commercial tools estimate AI prompt demand, but we do not have the same established, comparable volume data that exists for Google keywords. I would therefore not rank prompts purely by estimated popularity. Instead, consider these factors:

Imagine you sell accounting software for small businesses and you find these two questions:

1. “What is accounting software?”
2. “Which accounting software is best for a Dutch freelancer who needs to send invoices and file VAT returns?”

The first question is relevant, but the second comes from a potential customer who is considering a purchase. That does not mean you should ignore informational questions: they attract visitors earlier in their research and help establish your expertise. But if your immediate goal is generating customers, I would prioritise questions closely connected to the problems your product or service solves.

Group questions with the same intent

You do not need to treat every variation as a separate target. Consider these questions:

• How much does an SEO audit cost?
• What is the average price of a technical SEO audit?
• How much should I pay for someone to audit my website?

They all ask essentially the same thing. Instead of three articles, you could write one useful page about SEO audit pricing, what influences the cost and what customers can expect. The goal is to cover meaningful information needs, not every possible way of phrasing them.

How to test your prompts in AI platforms

Once you have selected your most important questions, see how AI assistants currently answer them. I recommend testing at least a representative selection in ChatGPT, Gemini and Google AI Mode; Claude and Perplexity are useful comparisons. Enter the question exactly as you wrote it down, then examine the answer:

1. Does the AI assistant recommend specific brands or products?
2. Is your company mentioned?
3. Which websites does it cite?
4. Which individual pages are cited?
5. What information appears to influence the answer?

Record the results in a spreadsheet so you can compare them later.

Do not rely on one answer

AI assistants do not always give the same answer to the same question. Results can change with the platform, model, previous conversation, location, language and the moment you ask. In my study, between 30% and 76% of the domains cited in a second round had also been cited in the first round for the same question, depending on the platform and language.

So do not draw conclusions from a single response. Run the same question several times under comparable conditions, and test a few natural variations to see whether the results hold. If your company appears once but is missing from the next five answers, that is very different from being recommended consistently.

Look beyond the original prompt: query fan-out

When you ask an AI assistant a question, it usually does not search for that question. It runs its own searches to gather information. Google confirms that AI Mode and AI Overviews can use query fan-out: issuing multiple related searches across subtopics.

In my study, I recorded the 1,830 web searches that ChatGPT, Claude and Gemini ran while answering the questions. Only one of them was the question itself. The platforms rewrote questions into keyword-style searches, and ChatGPT usually ran several: 2.8 on average when it searched, more than one in nearly nine out of ten answers, and up to 14 for a single answer. Claude usually ran one search, Gemini about two.

Here is what ChatGPT searched for when I asked “Which mortgage advisor is the best?” with the location set to Amsterdam:

• best mortgage advisor Amsterdam Netherlands expats reviews hypotheekadviseur 2026
• Consumentenbond hypotheekadvies beste hypotheekadviseur Netherlands
• Independer hypotheekadviseur reviews Amsterdam
• Advieskeuze hypotheekadviseur Amsterdam reviews

The searches added information that was not in the question, and that pattern held across the dataset:

ChatGPT often searched for specific sources by name, such as Consumentenbond, RTINGS or a product's official pricing page. For Dutch questions, at least one in six of its searches was in English. Gemini often added a year too, but usually the previous year rather than the current one.

These answers came through the platforms' APIs in October 2026; the consumer apps may search differently. The figures on languages and named sources come from automatic detection and are approximate.

What this means for prompt research: your page competes on these supporting searches, not only on the original prompt. That is why keyword research still matters, why sub-questions about pricing, rules and comparisons deserve clear answers, and why it helps to be present on the sources AI assistants look up by name, such as review sites, comparison sites and industry organisations. It is also another reason to investigate the information needs behind a question, rather than collecting dozens of slightly different prompts.

Turning prompts into pages that answer them

After researching and testing your prompts, compare them with the content on your website. For each important question, ask yourself: do I already have a page that answers this question clearly and completely?

If yes, you may not need to create anything new. Check whether the answer is easy to find, whether the information is accurate and whether you address the circumstances described in the question. If not, decide whether the information belongs on an existing page or deserves its own article.

Make the answer easy to find

Back to the accounting software example. Suppose you identified this question: “Which accounting software is suitable for Dutch freelancers who need to file VAT returns?”

An article about accounting software that mentions freelancers only briefly may not be the most useful result. A more relevant page discusses the needs of Dutch freelancers, explains which features matter for invoicing and VAT administration and compares suitable options. It answers the main question directly, then gives enough explanation to help someone decide.

That does not mean repeating the exact prompt throughout the article. Write naturally, and use headings that help readers find the information they are looking for.

Do not create a separate page for every prompt

One mistake I would avoid is generating dozens or hundreds of articles because an AI tool suggested different questions. Questions often overlap, and a well-written page can answer several related ones. Google's guide to optimising for generative AI features also warns against mass-producing pages for variations of queries.

Group prompts by intent and decide whether existing content can satisfy them. Create a new page when a question represents a genuinely different subject or information need.

What about schema, FAQs and other AEO techniques?

These can be useful, but I would not make them the starting point. In my citation research, structured data, author signals and FAQ headings did not consistently distinguish cited pages from non-cited pages on the same website. Pages with tables showed a smaller positive association, but the clearest finding was still how closely the content matched the question.

That does not make schema or FAQs useless. Structured data helps search engines understand certain content, and FAQs can make a page more helpful. Just do not expect an FAQ section or schema markup to make up for content that does not answer the question well.

How to measure the results of your AEO prompt research

Once you have improved your content, you will want to know whether it made a difference. Measuring AI visibility is less straightforward than tracking Google rankings. I would look at three things.

1. AI citations and brand mentions

Keep a stable list of important prompts and test them periodically. Record whether your brand appears in the answer and whether the AI assistant cites your website. Keep these apart: an assistant can recommend your business without linking to your website, and it can cite your content without recommending your company.

You can test prompts manually or use an AI visibility monitoring platform. Either way, keep the prompts and testing conditions consistent, so changes over time mean something.

2. Traffic from AI platforms

Use Google Analytics or another analytics tool to monitor visits referred by AI assistants, where referral information is available. Not every interaction with your content results in a visit, and not every AI-influenced visit is attributed to an AI platform, so referral traffic is useful but incomplete.

3. Business results

Ultimately, the purpose of AEO is not just to appear in AI answers. It should contribute to inquiries, purchases, bookings or other conversions. If traffic from AI assistants grows but does not bring relevant visitors or customers, reconsider which questions you target. Combine AI visibility monitoring with your usual analytics, rather than treating citation counts as the only measure of success.

The limits of AEO prompt research

Before you build an AEO strategy around prompts, keep these limitations in mind.

We do not have a complete picture of what people ask AI assistants. Customer conversations, Google searches and AI-generated suggestions help identify likely questions, but they do not represent the full volume of AI conversations.

AI results change. A prompt that produces a citation today may return different sources tomorrow. Testing gives you observations, not guaranteed rankings.

The question is not the only factor. My research found a strong relationship between question match and citations when comparing pages on the same website. But the website itself also mattered: many differences between pages disappeared once I accounted for which sites the AI systems selected.

Correlation is not causation. Part of the relationship may come from how AI systems retrieve pages in the first place: they find pages partly because they resemble the question. The evidence was strongest for ChatGPT and less conclusive for Claude. My analysis does not prove that rewriting a page to match a question more closely will increase citations.

These limitations do not make prompt research pointless. They mean you should treat it as a way of understanding your audience and improving your content, not as a guaranteed technique for getting mentioned by AI assistants.

Frequently asked questions about AEO prompt research

Final thoughts

Researching prompts for AEO does not need to be complicated. Start with the questions your customers actually ask. Use Search Console, keyword research and online discussions to find more, and tools such as the AI Prompt Finder to turn real searches into likely prompts. Prioritise the questions that matter to your business, and check how AI assistants currently answer them.

Most importantly, use those findings to improve the information on your website. My research suggests that relevance to the question is one of the strongest page-level signals associated with AI citations. That is a good reason to spend more time understanding what people want to know, and less time chasing every new AEO technique.

You cannot predict exactly which pages AI assistants will cite. But you can make sure your website gives genuinely useful answers to the questions that matter to your customers.