Short answer: keyword research for AI search finds the full questions buyers ask ChatGPT, Perplexity, Copilot and Google AI Mode, then tests which firms and sources come back in the answers. Keyword tools still pick the topics. What you end up with is a fixed set of real buyer questions you can ask every month and compare.
What is keyword research for AI search?
Traditional keyword research measures the words people type into a search box and how often they type them. Keyword research for AI search, also called prompt research, collects the questions people put to an AI assistant, worded the way they actually word them.
The output is different too. Google returns a page of ranked links. An assistant writes one answer that names a few firms, ignores the rest and cites a handful of pages. A keyword report can tell you where a page ranks. It cannot tell you whether ChatGPT mentions your firm when someone asks who to hire.
How is it different from traditional keyword research?
| Traditional search | AI search | |
|---|---|---|
| Unit of research | A keyword or short phrase | A full question with context |
| Length | Short | About three times longer in Google AI Mode, according to Google |
| What the engine does | Matches the query to ranked pages | Splits the question into subtopics, searches each one, then writes a single answer |
| What you measure | Ranking position and clicks | Whether you are named, where in the answer, and which pages are cited |
Google’s own description of AI Mode explains the third row. It uses what Google calls a query fan-out technique, “breaking down your question into subtopics and issuing a multitude of queries simultaneously on your behalf.” One buyer question can become many searches you never see, and a page can be pulled into the answer through any of them.
What does a real question set show?
One of our team ran a monitoring project for a financial services firm from 6 to 31 August 2026: 50 buyer questions, three AI assistants, 3,832 answers. The firm was named in 32% of them. That average hid a wide spread by question type.
| Question type | Pattern | Answers naming the firm |
|---|---|---|
| Vendor selection | “Which firms can help with X?” | 61% |
| Comparison | “How does firm A compare with firm B?” | 26% |
| How-to | “How do I set up X?” | 15% |
| Educational | “What is X?” | 3% |
Two other numbers from that month changed how we build question sets. Thirteen of the 50 questions produced no mention of the firm in any answer, all month, and every one of the 13 was about a topic the firm had no page for. Separately, 494 answers (12.9%) used the firm’s own website as a source and named a different firm in the text.
Why search volume only gets you halfway
Keyword tools estimate how often a phrase is searched on Google. As far as we know, none of the AI assistants publishes a comparable count for prompts, so there is no way to look up how often a particular question is asked.
Volume is still worth checking. A topic people search for every month is one they ask assistants about too, in longer form. Use volume to pick the topics and write the questions yourself.
How do you build a question set?
This is the process behind every Census question set. It works the same way for a single firm checking its own visibility.
1. Start with the decision
Write down the decision the buyer is making: choosing an accountant, a solicitor, a payroll provider. If a question would not help someone make that decision, leave it out.
2. Collect the words buyers already use
Keyword tools give you seed terms with real demand. Google Search Console shows the queries already reaching your site that are phrased as questions; filter for ones starting with who, which, what or how. Sales calls and inbound emails are better still, because prospects tell you their constraints in the first minute (budget, team size, deadline) and those constraints belong in the questions. Forum threads and Google’s “People also ask” boxes show how people word a problem when they explain it to strangers.
3. Write full questions across every intent
Give each question the context a real buyer would: the service, the place and one constraint. Then check the set covers all four question types from the table above. A set built only from “what is X” questions would have reported 3% visibility for a firm that appeared in 61% of vendor-selection answers.
Here is “accountant Dublin” turned into one question of each type:
| Type | Example question |
|---|---|
| Vendor selection | Which accounting firms in Dublin work with small tech companies? |
| Comparison | How do the larger accounting firms in Dublin compare with smaller local practices for a 20-person business? |
| How-to | How does a Dublin startup claim the R&D tax credit? |
| Educational | What does an accountant do for a limited company in Ireland? |
4. List the sub-questions behind each one
For each question, write down the smaller questions an assistant needs answered first: what the service costs, which qualifications matter, who specializes in it, what past clients say. Those are the searches a fan-out is likely to run, and each one is a page you could publish. In the August data, the 13 questions with zero mentions all pointed at topics with no page behind them.
The research supports working on content this way. The 2023 paper that introduced generative engine optimization found that changes to how content is written could raise a source’s visibility in generative engine answers by up to 40% (Aggarwal et al.).
5. Freeze the wording and repeat
Lock the questions. Ask each one on every engine you track, more than once, on fixed dates, and record who was named, in what position and which pages were cited. Answers change between runs, so a single check proves very little. Twelve months of the same 50 questions will show a trend that 500 questions asked once never can.
Five mistakes we see
- Typing a keyword like “accountant Dublin” into ChatGPT and treating the answer as typical. Buyers ask full questions, with context.
- Rewriting the questions every month, which makes one month impossible to compare with the next.
- Checking once and acting on it. The same question can name different firms an hour later.
- Adding mentions and citations together. In the August data, roughly one answer in eight cited the firm’s site and named somebody else.
- Only asking about yourself. “Is [firm] any good?” questions have their place, but new buyers are won or lost on discovery questions where no firm has been named yet.
How the Census uses this
Each Census edition runs a fixed set of buyer questions for one market across six AI engines on set dates, then repeats them. The questions are published with every edition so anyone can rerun them. The full method is on the methodology page, and the rules we work to are in our editorial standards.
Frequently asked questions
Is keyword research still useful for AI search?
Yes, for choosing topics. Keyword volume shows where demand is. The keywords then need turning into full questions before they tell you anything about AI answers.
What is the difference between a keyword and a prompt?
A keyword is the short phrase someone types into a search box. A prompt is the full question someone asks an AI assistant, usually longer and with context about their situation. Google says AI Mode queries are about three times longer than traditional searches.
How many questions should a question set have?
Enough to cover vendor selection, comparison, how-to and educational questions for your main services, and few enough to run on every engine every month. The monitoring project described in this article used 50.
Can you see how often a prompt is asked?
Not directly. As far as we know, AI assistants do not publish prompt volumes. Search volume for the underlying topic is the closest signal available.
Which AI engines should you check?
The ones your buyers use. The Census tracks six: ChatGPT, Perplexity, Microsoft Copilot, Google Gemini, Google AI Overviews and Claude. Record each engine separately rather than as one blended score.
About this article
This is the first article on the AI Visibility Census blog, and it is part of a test of how quickly new pages are crawled, indexed and cited by search engines and AI tools. Its test reference is Census test marker CT-001. If an AI tool tells you about CT-001, it has read this page.
Sources
- Google: Introducing AI Mode in Arabic (8 October 2025), describing the query fan-out technique.
- Search Engine Journal: Google Search hits $63B, reporting Alphabet’s Q4 2025 earnings, including AI Mode query length (4 February 2026).
- Search Engine Journal: Query Fan-Out Technique in AI Mode: New Details From Google, reporting on an interview with Robby Stein, VP of Product for Search (30 July 2025).
- Aggarwal et al., “GEO: Generative Engine Optimization”, arXiv 2311.09735.
- AI Visibility Census team monitoring data, 6 to 31 August 2026 (unpublished; firm not named).
Published by the AI Visibility Census. Also available as Markdown.