Prompt Research: How to Find AI Search Keywords

Prompt research is the AI search version of keyword research. You find the questions your customers ask AI assistants, you group them by what the person wants, and you test them to see who the AI recommends.

Semrush, which sells a prompt research tool, defines it in a similar way: identifying and tracking the questions that lead AI systems to compare options and recommend specific brands (Semrush, August 2026). Similarweb describes it as finding the full, conversational questions people type into AI systems, working out which matter to a brand, and spotting where the brand is missing (Similarweb, May 2026).

We use a slightly wider definition, because prompt research does three jobs at once.

It is a demand map. It shows what people actually ask, in their own words, including the details they add: budget, location, language, who it’s for, what they’ve already tried.

It is a visibility check. When you run a prompt, you see the answer. You see which brands are named, in what order, and which pages are linked as sources. A keyword report never told you that a competitor was being recommended by name.

It is a content brief. Every prompt where a competitor gets cited and you don’t is a gap. Often the fix is a page that answers the question more clearly, or a fact that’s missing from the page you already have.

Here’s the running example we’ll use in this guide. Imagine a dental clinic in Bangkok that treats Thai patients and many international visitors. For years its keyword list looked like “veneers bangkok”, “dental implant price” and “ทำฟันที่ไหนดี”. Its prompt list looks very different, as you’ll see in the steps below.

Prompts differ from keywords in length, structure, context and intent. The easiest difference to measure is length, and the gap is large in every dataset we found.

Semrush’s clickstream study found that ChatGPT prompts averaged 23 words, while searches averaged 4.2 words (Search Engine Land, 2025). Similarweb, citing its own 2025 generative AI report, puts the average ChatGPT prompt at about 60 words against 3.4 words for an average Google search (Similarweb).

The two studies disagree on the exact number, which is normal. They use different panels, different time windows and probably different ways of counting words. Read the chart as a direction, not a precise ratio: prompts are somewhere between five and fifteen times longer than searches.

Google’s own number points the same way. It says the average AI Mode search is about three times longer than a traditional search (Search Engine Journal, May 2026). AI Mode sits inside Google, where people still carry keyword habits, so it makes sense that it lands between a search and a ChatGPT prompt.

Three differences that change how you research

Context. A prompt carries the details a keyword strips out. “Veneers bangkok” becomes a question that includes a five-day trip, a language need and a budget concern. Those details decide which clinic the AI recommends, because the AI looks for pages that address them.

Conversation. A keyword is one shot. A prompt is often a turn in a conversation, and the follow-up (“which of those has the best reviews from foreigners?”) can matter more than the opener. Google said follow-up queries in AI Mode in the US grew by more than 40% per month in its first year (Search Engine Journal).

Personalization. Two people can type the same prompt and get different answers, because of location, chat history and settings. OpenAI says ChatGPT search shares general location inferred from your IP address to improve results (OpenAI Help Center). That’s one reason prompt research has to run prompts several times and from the right country.

Most people use AI tools to ask questions and get practical guidance, which is exactly the kind of demand prompt research tries to capture.

The best public data comes from OpenAI itself. In September 2025, a team of OpenAI and academic researchers published “How People Use ChatGPT” as an NBER working paper. They classified a sample of about 1.5 million conversations with automated tools, without humans reading the chats (NBER; BGR).

Three findings matter for prompt research.

Asking is the biggest use. About 49% of messages were “Asking”, meaning the person wanted information or advice to make a decision. Another 40% were “Doing” (a task like drafting an email) and 11% were “Expressing” (FlowingData). Asking prompts are where brands get recommended, so that’s where most of your research should go.

Three topics dominate. The paper says “Practical Guidance”, “Seeking Information” and “Writing” together account for nearly 80% of conversations (NBER). Practical guidance is advice tailored to the person (“how should I prepare for implant surgery if I’m flying home a week later?”), which is exactly the kind of question a service business can answer better than a generic article.

Most use isn’t work. Non-work messages grew from 53% to over 70% of usage by mid-2025 (NBER). People ask AI about health, travel, money, shopping and family decisions. If you sell to consumers, your customers are already asking.

For our dental clinic, this changes the brief. “Asking” prompts like “is it safe to get veneers abroad?” and “how many visits do implants take?” are worth more research time than “Doing” prompts like “write me a message to cancel my dentist appointment”, which no clinic page will ever win.

No. Prompt research adds a layer on top of keyword research; it doesn’t replace it.

The fan-out data in the last section is the reason. ChatGPT, Gemini, Perplexity and Google AI Mode all search the web with shorter queries before they answer. Those queries are still keywords. If your page can’t be found for them, it can’t be cited.

Google is explicit that there are no additional requirements or special optimizations needed to appear in AI Overviews or AI Mode, and that normal SEO fundamentals still apply (Google Search Central). Keyword research is part of those fundamentals.

What prompt research adds is the missing context and the missing scoreboard. Here is how we think about the split.

If you’re new to the wider discipline, our explainer on optimizing content for large language models covers how the two fit together at the page level.

To find AI search keywords, you combine what your customers already ask with the query data Google and Microsoft give you, then turn those seeds into a tested set of prompts. Here is the process we use, in order.

You don’t need a paid tool for any of it. Paid prompt databases help in step 4 and with volume estimates, which we cover in their own section, but the highest-quality inputs are free: your customers, Search Console and Bing Webmaster Tools.

You find AI-style queries in Search Console by filtering the Performance report for long and question-shaped queries with a regular expression. It’s the closest thing Google gives you to a prompt list.

First, know what Search Console does and doesn’t show. Google’s documentation says traffic from AI features is included in the overall search traffic in the Performance report, under the “Web” search type (Google Search Central). So AI Mode and AI Overviews queries are mixed in with everything else.

In June 2026 Google added a dedicated view. The Search Generative AI performance reports show how often your URLs appeared in generative AI features in Search and Discover, broken down by page, country, device and date (Google Search Central Blog, 3 June 2026). Google says they reached all websites worldwide as of 31 August 2026.

The catch for prompt research: the report excludes click and query data (Search Engine Roundtable, September 2026). It tells you which pages appear in AI features. It doesn’t tell you which questions triggered them.

The regex filter

Search Console’s Performance report lets you add a query filter, choose “Custom (regex)”, and use RE2 syntax. Matching is partial by default and not case-sensitive (Search Console Help).

Two filters do most of the work. The first keeps queries that start with a question word:

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

The second keeps long queries of roughly eight words or more, whatever their shape:

^(\S+\s+){7,}\S+$

Run each over the last three to six months, sort by impressions, and export. The first filter finds questions. The second finds conversational searches that don’t start with a question word, such as “dentist bangkok veneers in 5 days english speaking price”. Both are the shape of query that comes from AI Mode or from people who’ve picked up prompt-style habits.

Look for three things in the export. Constraints that keep repeating (time, language, budget). Questions where you have impressions but a low click-through rate, which usually means the page doesn’t answer the question in the snippet. And questions with impressions but no page that really answers them.

Each of those rows becomes a seed for your prompt set. Rewrite it as the full question a person would ask an assistant, and keep the original query next to it so you can track both.

Bing Webmaster Tools is the only free tool that shows you the actual search phrases an AI system used before citing your pages, which makes it the most direct source of AI search keywords available today.

Microsoft launched the AI Performance report in public preview on 10 February 2026. It covers citations in Microsoft Copilot, AI-generated summaries in Bing and select partner integrations, and shows total citations, average cited pages per day, page-level citation counts, and grounding queries (Bing Webmaster Blog).

Microsoft’s Fabrice Canel, one of the product managers named on that announcement, shared the launch on X the same day.

That definition is why grounding queries are so useful. A prompt is what the person typed. A grounding query is what the AI searched for. If you want to know which AI search keywords to cover, this is the list.

In March 2026, Bing linked the grounding queries and pages views, so you can select a query to see which of your pages were cited for it, or select a page to see which queries led to it (Search Engine Roundtable). Fabrice Canel replied on X that more was on the way.

In June 2026 the report added intent buckets (such as informational, commercial and navigational), topic clusters, a citation share metric that shows your percentage of all citations for a grounding query across sites, and period comparisons (Search Engine Roundtable).

How we use it

Export the grounding queries for the last few months. Sort by citations. Then do three passes.

Pass 1: queries you already win. These show what Microsoft’s AI already trusts you for. Protect those pages: keep facts current and don’t remove the passages that get cited.

Pass 2: queries with low citation share. You’re cited sometimes, but others win more often. Compare your page with the sites that dominate and look for missing facts, prices or comparisons.

Pass 3: patterns. Group the queries by the constraint they contain (price, time, location, language, comparison). Each pattern becomes a family of prompts in step 5.

If you haven’t verified your site in Bing Webmaster Tools, do it today. It takes minutes, and the grounding query data starts building as soon as Copilot or Bing cite you.

You turn seed questions into prompts by writing them the way a person would speak to an assistant, then adding the details that change which brands get recommended.

Those details are modifiers. A seed like “veneers in Bangkok” says almost nothing. Add who is asking, what limits them and what they’re comparing, and you get a prompt that behaves like a real one.

Semrush makes the same point: the constraints that matter are the ones that push the AI from explaining to recommending (Semrush). “What are veneers?” gets an explanation. “Which clinic should I choose for veneers on a five-day trip?” gets brand names.

A few rules keep the set honest.

Write like a customer, not a marketer. Customers don’t say “premium aesthetic dentistry solutions”. They say “good dentist for veneers that won’t rip me off”. If a prompt sounds like your homepage, rewrite it.

Don’t put your brand in discovery prompts. “Is Clinic X good for veneers?” only tells you what AI says when someone already knows you. Keep a separate, small group of branded prompts for reputation checks.

Keep prompts the length real prompts are. The research says prompts are long, so a set full of three-word phrases tests the wrong thing. Most of ours land between 12 and 40 words, with a few short ones to mirror quick questions.

Vary one thing at a time. If you want to know whether language changes the answer, write the same prompt in English and Thai and change nothing else. Otherwise you can’t tell which detail made the difference.

Here’s how one seed from step 1 becomes a small family of prompts:

1. “How many days do I need in Bangkok to get porcelain veneers?”
2. “I’m in Bangkok for 5 days. Is that enough for veneers, and which clinics can do it?”
3. “Composite or porcelain veneers if I only have a week in Thailand?”
4. “ทำวีเนียร์ใช้เวลากี่วัน คลินิกไหนในกรุงเทพทำได้เร็ว”

You estimate prompt volume with third-party models, because no AI platform publishes how often a given prompt is asked. That makes volume the weakest part of prompt research, and it’s worth understanding why.

With Google keywords, tools at least start from a shared reference point: Google Ads data plus clickstream panels. With AI prompts there’s no equivalent. OpenAI, Google and Microsoft don’t release prompt counts, and Google’s new AI report in Search Console doesn’t include queries at all.

So vendors build volumes from what they can see. Two examples show the range of methods.

Semrush says its Prompt Research report draws on a prompt database of more than 317 million AI queries, refreshed daily, and describes its volumes as estimates (Semrush Knowledge Base).

Profound builds its Prompt Volumes from conversations it licenses from AI answer engines, cleaned and run through probabilistic modeling. Its documentation says coverage is strongest in the US, the UK and select European markets, with US history starting in January 2025 and other regions in January 2026 (Profound Docs). When Profound described the feature in August 2025, it covered only US ChatGPT conversations and published no accuracy metrics for its projections (Profound).

Good use: comparing topics. If one tool shows “veneers abroad” as a much bigger topic than “gum contouring”, that relative gap is probably real, even if neither number is exact.

Good use: discovering phrasing. Prompt databases show you real wording you can adapt, which is often more valuable than the number next to it.

Bad use: forecasting traffic. A prompt volume can’t tell you how many visits you’ll get, because AI answers often don’t produce clicks at all. Don’t put prompt volumes into a revenue model.

Bad use: small markets. For Thai-language prompts or niche B2B topics, panels are thin. Your own Search Console, Bing and customer data are usually a better guide there than a modeled number.

You test prompts by running each one several times on each platform that matters, from the right location and language, and recording mentions, citations and competitors in the same format every time.

Several runs matter because AI answers change. The same prompt can name different brands from one run to the next, and personal context changes results too. One screenshot is an anecdote. Three runs start to show a pattern.

That’s why our own tool, GrixTrack, runs each prompt three times in ChatGPT, Gemini and Google AI Mode and records whether the brand is mentioned and whether its pages are cited, alongside Google rankings by country, city and language. You can do the same by hand for a small set: use a logged-out or clean session, set the country you care about, and repeat each prompt three times. Setting it up for the first time? The GrixTrack step-by-step guide walks through every menu with pictures and short videos.

For each run, record the same fields:

• Date, platform, country and language
• Was your brand mentioned, and in what position?
• Was one of your pages cited, and which one?
• Which competitors were mentioned?
• Which sources were cited (your site, competitors, review sites, forums, media)?

The sources column is the one people forget, and it’s often the most useful. If an answer keeps citing a forum thread or a review platform, that’s where you need a presence, not just your own site. Our guide on measuring AI visibility and tracking citations goes deeper into turning these logs into metrics.

Then turn results into actions. For each prompt where you’re missing, look at what the cited pages have that yours doesn’t. When we compare pages, the gap is most often one of four things: a direct answer in the first sentence under a heading, a specific fact (a price, a timeline, a number), a clear comparison, or a mention on a third-party site the AI trusts.

Rerun the core set at least monthly. After you publish or update a page, rerun its prompts after the page has been recrawled, so you can see whether the change made a difference.

Most businesses should start with 20 to 50 prompts and grow from there. A set that’s too big becomes a dashboard nobody reads.

The vendors’ numbers give a sense of range. Semrush says that in its internal testing, about 10 well-chosen prompts per product are usually enough to see whether AI consistently recommends a brand, and that small teams can track up to 25 prompts in its AI Visibility Toolkit, while enterprise setups range from 500 to 15,000 (Semrush). GrixTrack’s free plan covers 5 prompts, 10 keywords, 1 project and 2 competitors, which is enough to test the method on your most important questions before you scale.

How we size a set: one group of prompts for each main service or product, about five to ten prompts per group across the four intent stages, a handful of branded checks, and a duplicate in each language your customers use. For our clinic, that’s veneers, implants, whitening and general dentistry, in English and Thai, which lands at around 50 to 60 prompts.

The mistakes we see most often

Invented prompts. Lists generated by asking an AI to imagine what customers ask. They’re tidy, generic and often wrong. Start from real data, then use AI to rephrase.

Branded prompts only. “What is Clinic X known for?” tells you about reputation, not discovery. If every prompt contains your name, you’ll look visible while competitors win every unbranded question.

One run per prompt. Answers vary. A single run can show you as the top recommendation one day and missing the next. Always use several runs and look at the trend.

No link to pages. A prompt with no page assigned to it never gets fixed. Every prompt should map to the page that should win it, or to a page you plan to create.

English only. In bilingual markets, the Thai and English answers can name different brands and cite different sources. We cover that next.

In bilingual markets, prompt research has to happen in each language separately, because people ask different questions in each language and AI tools can answer them from different sources.

Thailand is a good example. A Thai business can see Thai, English and transliterated queries side by side in its Search Console export, sometimes mixed in the same search. Check your own data to see which topics your customers ask about in which language, because the split is rarely the same across services.

The fan-out data shows the language effect too. Peec AI’s study of more than 20 million ChatGPT fan-outs included Thailand alongside Germany, the UK, Singapore and the US, and found the same lengthening trend across countries, with Thailand showing the lowest word counts throughout (Peec AI). Thai doesn’t put spaces between words, so word counts aren’t directly comparable, but it’s a reminder that you can’t assume English patterns hold in Thai.

Here’s what we do differently for bilingual sets.

Collect each language from its own sources. Thai prompts come from LINE chats, Thai reviews, Pantip threads and Thai People Also Ask results. English prompts come from expat forums, international reviews and English queries in Search Console. Translating your English prompts into Thai produces questions nobody actually asks.

Run both and compare. For each important prompt, run the English and Thai versions and compare the brands and sources. If the Thai answer leans on Thai-language sources and the English answer on international ones, your Thai pages and your English pages have separate jobs.

Set the location. A prompt asked from Bangkok and the same prompt asked from Sydney can produce different answers, because ChatGPT uses general location to improve search results (OpenAI Help Center). Test from the countries your customers are in.

Keep facts identical across languages. If the Thai page says a treatment takes three visits and the English page says two, AI tools see conflicting facts. Make prices, timelines and opening hours match everywhere, including Google Business Profile.

If you work in Thailand specifically, our SEO service in Bangkok covers the keyword side of bilingual search in more depth.

Prompt research isn’t a one-off project. Your customers’ questions change, AI platforms change how they search, and Google and Microsoft keep adding data to their reports. Start with 20 real questions this week, test them three times each, and you’ll know more about your AI visibility than most of your competitors do. If you want a head start, our explainer on AI visibility and AI share of voice shows how to turn the results into numbers you can report. You’ll find more of our AEO and SEO guides in the Knowledge Hub.

1. Semrush via Search Engine Land: ChatGPT growing as a traffic referrer, reshaping search behavior (80 million clickstream records, second half of 2024)
2. Chatterji et al., NBER Working Paper 34255: How People Use ChatGPT (September 2025)
3. BGR: What do people use ChatGPT for? OpenAI studied 1.5 million chats to find out (16 Sep 2025)
4. FlowingData: What people use ChatGPT for (16 Sep 2025)
5. Search Engine Journal: Google shares first AI Mode usage data after one year (20 May 2026)
6. Google Search Central: AI features and your website (updated 10 Dec 2025)
7. Google Search Central Blog: Introducing Search Generative AI performance reports in Search Console (3 Jun 2026)
8. Search Engine Roundtable: Google Search Console AI performance report and AI control go global (1 Sep 2026)
9. Google Search Console Help: Performance report advanced filtering and comparison (regex filters)
10. Bing Webmaster Blog: Introducing AI Performance in Bing Webmaster Tools Public Preview (10 Feb 2026)
11. Search Engine Roundtable: Bing Webmaster Tools rolls out AI Performance report (11 Feb 2026)
12. Search Engine Roundtable: Bing AI Performance reports add grounding queries to pages (24 Mar 2026)
13. Search Engine Roundtable: Bing adds intents, topics, citation share and compare to AI reporting (17 Jun 2026)
14. Peec AI: ChatGPT fan-outs have doubled in length in 4 months (20 million fan-outs, 12 Feb 2026)
15. Peec AI Research: Fan-outs per prompt by engine (about 5 million fan-outs, April 2026)
16. Ahrefs: AI Overview citations and the top 10 (863K SERPs, 2 Mar 2026)
17. Similarweb: How to do prompt research for AI SEO (May 2026)
18. Semrush: Prompt research for AI SEO (17 Aug 2026)
19. Semrush Knowledge Base: Prompt Research report
20. Profound: Prompt Volumes keyword volume projection (1 Aug 2025)
21. Profound Docs: About Prompt Volumes
22. OpenAI Help Center: ChatGPT search
23. Google for Developers on YouTube: Building a frontier AI search experience, with Robby Stein (23 Jul 2025)

Keep reading

• Content & technicalHow to Write Answer-First Content for AI Search
• Content & technicalLLM SEO: How to Optimize Content for LLMs
• Measurement & localHow to Measure AI Visibility and Track AI Citations
• Content & technicalHow to Optimize FAQs for AI (With FAQ Schema Examples)