Keyword Research for AI Search: From Keywords to Questions and Topics

Does your keyword list still look like a column of two- and three-word phrases with a volume number next to each? Then it is missing how more and more people now search. Keyword research for AI search means collecting the full questions your customers ask. You group them into topics and entities, then build pages that answer the whole set, not one phrase.

The good news is that the core skill hasn’t changed. You are still trying to understand what people want and give it to them better than anyone else. What changes is the unit of research. It moves from the single keyword to the question set. If you are new to the basics, start with my keyword research guide for beginners.

Why Keyword Research for AI Search Looks Different

When someone types into a classic search box, they squeeze their need into a few words: “ebike battery life”. When they ask an AI system, they tend to write the way they talk: “how many years will an ebike battery last if I charge it every day and ride 20 km?”

That longer question carries far more context. It includes the situation, the limits and often the decision the person is trying to make. A tool that only reports volume for “ebike battery life” hides all of that.

This applies across the main AI search tools:

• Google AI Overviews appear above normal results for some queries. They sum up an answer with links to sources.
• Google AI Mode is a chat-style tab where people ask follow-up questions in a thread.
• ChatGPT and Perplexity run their own web searches for many questions and cite the pages they use.

Each one picks sources a little differently. We covered that in how AI Overviews and AI Mode pick their sources. But they share one habit that matters for research. They break a question into smaller pieces before answering it.

Query Fan-Out: The Idea Behind the Shift

Google’s own help page on AI features and your website says that AI Overviews and AI Mode may use a “query fan-out” technique. That means running several related searches across subtopics and data sources to build one answer.

In plain terms, one question from a user becomes several searches behind the scenes. Take “what’s the best ebike for a hilly 15 km commute?” A system might look up several things separately:

• motor types and which suit hills
• battery size and real-world range
• bike weight and how easy it is to carry
• price ranges
• servicing and reliability

Your page doesn’t need to rank for the original long question word for word. It needs to be a strong answer to one or more of those sub-questions. Content that covers a topic fully, with each subtopic answered clearly, gives the system more reasons to use it.

So when I research, I think in two directions. First, which questions do people ask? Second, for each question, what sub-questions would a careful expert need to answer to respond well? That second list is your fan-out map.

Google doesn’t publish the exact sub-queries it runs, and no outside tool can see them directly. Treat any fan-out list as an informed guess about the topic, not a record of what Google actually did.

From Keywords to Question Sets, Topics and Entities

I find it helps to work at three levels at once.

Question Sets

A question set is a group of questions that share the same need. “How long does an ebike battery last”, “when should I replace my ebike battery” and “why is my ebike range dropping” all belong to one set about battery life and care. One good page can answer the whole set.

Topics

Topics are the broader themes that hold several question sets: buying an ebike, looking after one, commuting with one. Topics help you plan your site structure and decide which sections or hubs you need.

Entities

Entities are the specific things, places, brands, standards and ideas that a topic involves. For ebikes, that means mid-drive and hub motors, watt-hours, torque sensors, the brands you stock and your town or region. AI systems and Google’s own systems understand topics through these entities. If an expert page on the subject would naturally mention them, yours should too, explained in plain language.

Classic keywords still matter. People still type short queries, and those queries still have volume worth chasing. The change is that keywords become labels on a group of questions, not the whole plan.

Free Sources of Real Questions

You don’t need a paid tool to build a strong question list. Some of the best sources are free, and several are already inside your business.

People Also Ask

Search your main topics in Google and open the People Also Ask boxes. Each one you open usually loads more. Write down the questions. Also note the wording, because it shows how people describe the problem.

Google Autocomplete

Start typing a question opener such as “how long does an ebike”, “can you ride an ebike” or “is it worth” followed by your topic. Try different openers: how, why, can, should, which, is, vs. Autocomplete reflects real searches. It is a quick way to spot comparison questions and worry questions.

Search Console Long Queries

This is where I usually start. Your own Performance report already holds long, chat-style queries that showed your pages. Open the Search results report and add a Query filter. Choose “Custom (regex)” and use a pattern that matches queries of seven or more words:

^(S+s+){6,}S+$

To pull question-style queries, try a pattern like this:

^(how|what|why|when|which|can|does|do|is|should|will)b

Sort by impressions. Long queries with impressions but few clicks are often questions you are close to answering, but not quite.

One warning: Search Console hides some very rare queries to protect user privacy. Many long, unique questions fall into that group. So treat what you see as a sample, not the full picture.

Search Console’s AI-powered Configuration

Search Console now lets you describe the report you want in plain language. It then sets up the filters, metrics and date comparisons for you. Google added it as a test in December 2025, and it is now open to everyone. You might ask for “queries containing how or why for pages in the /guides/ folder, last 3 months compared with the previous 3 months.”

It only works on the Search results Performance report, not Discover or News. It sets up the view but doesn’t sort or export the table. Always check the filters it applied before you trust the numbers, because it can misread a request. Our guide to using Search Console’s AI report builder walks through it step by step.

Forums and Communities

Reddit, Quora, niche forums, Facebook groups and YouTube comments are full of questions in people’s own words. They often include the same context AI search users give. Look for threads where people describe their situation before asking. Those are the long, specific questions you want.

Customer Support and Sales Logs

In my experience, this is often the richest source, and competitors can’t see it. Pull questions from:

• support emails and help desk tickets
• live chat transcripts
• sales call notes and quote requests
• your site’s internal search log
• reviews, both yours and competitors’

Remove personal details and keep the question wording. If your team answers the same question every week, it belongs on your site.

Asking the AI Tools Themselves

Ask ChatGPT, Perplexity or Google AI Mode your core questions. Note the follow-ups they suggest and the subtopics they cover. That shows you what a full answer is expected to include. Treat it as a source of ideas, not as data about what people search. These tools don’t show search volumes, and their answers change from one run to the next.

Grouping Questions by Intent

Once you have a few hundred questions, group them by what the person is trying to do. This simple set of intent groups works for most businesses:

• Learn: understanding how something works (“how does a torque sensor work”).
• Compare: weighing options (“hub motor vs mid-drive for hills”).
• Decide: checking if something fits their situation (“is a 500 Wh battery enough for a 30 km round trip”).
• Do: carrying out a task (“how to store an ebike battery over winter”).
• Fix: solving a problem (“ebike motor cuts out on hills”).
• Buy or visit: ready to act (“ebike shop with test rides near me”).

Within each group, merge questions that would get the same answer. My test is simple: would the person asking be happy with the same page? If yes, they belong together. If one needs a comparison table and the other needs a repair guide, split them.

Watch for the context people add. “For a heavy rider”, “with a child seat” or “in the rain” can change the answer enough to need its own section, even if the main question is the same.

Mapping Question Sets to Pages

The mistake to avoid here is making one thin page per question. Lots of near-copy pages, each answering a slight variation, is exactly the kind of content Google’s spam policies warn about when it is done at scale. It also splits your authority across pages that compete with each other. I see this mistake a lot, and I’d always rather have one strong page.

Instead, map each question set to one page, and each topic to a hub or section:

1. Check whether an existing page already covers the set. If so, expand it instead of starting fresh.
2. Give each sub-question a clear heading and a direct answer in the first sentence or two below it.
3. Add the detail an expert would add: numbers, conditions, exceptions and what you would recommend.
4. Link related pages together so the topic is easy to follow for readers and crawlers.

First-hand detail matters more when AI systems sum up the basics for free. Details from your own work, such as what you see in your workshop or what customers actually choose, are hard to copy. Our guide to what experience, expertise and trust mean for your site covers how to show that.

A Worked Example: An Independent Ebike Shop

Imagine a single-location bike shop that sells and services ebikes. It has a basic website: a home page, a product range, a servicing page and a contact page. Here is how the process plays out.

Collecting the Questions

The owner spends an afternoon pulling questions from the shop email inbox, notes from the workshop counter, People Also Ask and a couple of local cycling forums. Search Console’s long-query filter turns up a few surprises. These include questions about storing batteries in a cold garage and whether ebikes can go on a car rack.

In total they gather around 150 questions. After merging duplicates and near-duplicates, about 40 separate question sets remain.

Grouping and Mapping

Here is a sample of how those sets map to pages:

Thinking Through the Fan-Out

For the commuting guide, the owner lists the sub-questions a good adviser would raise. These are distance and hills, rider weight, where the bike is stored, how much carrying is involved, budget, and how often it will need servicing. Each becomes a short section with a direct answer and a recommendation.

The guide names the entities that matter: motor types, battery sizes in watt-hours, the brands the shop stocks and the local routes customers ride. It ends with a clear invitation to book a test ride.

What They Avoid

They don’t create separate pages for “ebike for 10 km commute”, “ebike for 15 km commute” and so on. One strong commuting guide with scenarios covers all of them. It is also far easier to keep up to date.

Measuring Results

Measuring AI search is still imperfect, so I set expectations early. Here is what you can track honestly.

Search Console

Google’s help pages say traffic from AI features is included in the Performance report under the Web search type. There isn’t a separate filter for AI Overviews or AI Mode, so you can’t separate them. What you can do is watch impressions and clicks for each page you created or expanded. You can also watch the long and question-style queries filtered with the patterns above. Compare like-for-like periods, and allow for seasons.

Analytics Referrals

Visits from ChatGPT and Perplexity usually appear in your analytics as referral traffic from their domains. Set up a simple segment or channel group for AI assistants. Then you can see which pages they send people to and whether those visitors buy or enquire.

A Manual Prompt Panel

Pick 20 to 30 real questions from your list. Ask them in AI Overviews, AI Mode, ChatGPT and Perplexity once a month. Record whether your site is cited, which page, and which competitors appear. Answers vary between runs, users and locations. So look for trends over several months instead of reacting to one check.

Business Outcomes

In the end, track the things that pay the bills: enquiries, bookings, sales and branded searches. Someone may read an AI answer that cites you and then search for your name. That shows up as branded demand, not as a click on the original question. I always keep an eye on branded search for this reason.

Common Mistakes to Avoid

• Chasing made-up “AI search volume” numbers. No AI platform publishes query volumes the way Google Ads Keyword Planner does. Be careful with any figure that claims to be one.
• Stuffing pages with every question variation. Answer the question once, well. Repeating near-identical questions as headings makes pages worse for people.
• Dropping classic keyword research. Short queries still drive a lot of traffic. Use them to name and rank your topics.
• Ignoring technical basics. AI features draw from pages Google can crawl and index. If a page isn’t indexed, it can’t be a source. Our explainer on how Google Search works covers the basics.
• Researching once and stopping. Questions change with seasons, products and news. Refresh your list every quarter, and build that review into your yearly SEO plan.

Conclusion: Keyword Research for AI Search Starts With Your Customers

Keyword research for AI search isn’t a new skill. It is the old one done properly. Start with real questions and group them by need. Think through the sub-questions an expert would answer. Then build fewer, stronger pages around them.

Your next step: open Search Console, apply the long-query filter, and list the 20 longest questions your site already appears for. That list will tell you which pages to expand first.