Keyword research for GEO (Generative Engine Optimization) means finding the questions people ask AI assistants, the hidden sub-searches those assistants run to answer them, and the sources they trust, then building pages that cover all three. You still start from keywords, because keywords tell you where demand and money are. But you finish with a set of prompts, a map of fan-out queries and a short list of pages that can be cited in ChatGPT, Google AI Overviews, AI Mode, Gemini, Perplexity and Copilot.
The shift is practical, not philosophical. A person types “crm small business” into Google. The same person asks ChatGPT, “Which CRM should a 12-person agency use if we already pay for Google Workspace and hate long contracts?” That second question carries a role, a constraint, an existing tool stack and an objection. Classic keyword tools rarely show it, and an assistant answers it by searching for several narrower things you also need to cover.
This guide gives you an eight-step process you can run in a spreadsheet, the tools that help at each step, a scoring model for prioritization and the mistakes we see most often. Everything below works alongside normal SEO. If you want the bigger picture first, read our comparison of GEO vs SEO.
How is keyword research for GEO different from classic keyword research?
Classic research ranks phrases by search volume and difficulty. GEO research ranks situations by business value and by how likely an AI engine is to need your page to answer them. The unit of work changes from a keyword to a prompt cluster: one buyer situation, the different ways people phrase it, and the sub-questions behind it.
Three things change in practice:
- Length and context. Prompts are longer and include details such as budget, location, company size or a product the person already owns.
- Invisible demand. No public tool reports the volume of private chatbot prompts. You estimate demand from proxies: long-tail search queries, community threads, sales calls and support tickets.
- Sources over positions. The goal is to be one of the few sources an engine reads and cites, or the brand it names. That depends on entities, third-party coverage and how clearly a page answers each sub-question.
| Dimension | Classic keyword research | GEO prompt research |
|---|---|---|
| Unit of research | Keyword or phrase | Prompt cluster built around one buyer situation |
| Typical input | Two to four words | A full sentence with constraints and context |
| Demand signal | Monthly search volume | Proxies: long-tail queries, community threads, sales and support questions |
| Main competitor | Pages ranking in the top ten | Sources and brands the engine cites or names |
| What you map | Keyword to URL | Prompt to fan-out sub-queries to sections and pages |
| Success metric | Rank, clicks | Mentions, citations, recommendations, AI referral sessions and leads |
Why do prompts need their own research?
Because people write very differently to an assistant than to a search box, and engines respond to longer questions with AI answers more often. The data points below come from the platforms or from large studies, with dates.
- Prompts are much longer. A Semrush clickstream study of 80 million records from the second half of 2024 (published 3 February 2025) found ChatGPT prompts sent without search averaged 23 words, while prompts that triggered ChatGPT search averaged 4.2 words. Semrush sells SEO software, so treat this as directional, but the gap is large.
- Google sees the same pattern. At Google I/O in May 2026, Google said AI Mode had passed one billion monthly users and that an average AI Mode query is about three times longer than a traditional Search query, as reported by Talk Android (20 May 2026).
- Long questions trigger AI answers. Pew Research Center tracked 68,879 real Google searches by 900 US adults in March 2025 (published 22 July 2025). An AI summary appeared on 8% of one- or two-word searches, 53% of searches with ten or more words and 60% of searches phrased as a question.
- Clicks are scarcer. In the same Pew data, users clicked a traditional result on 8% of visits with an AI summary versus 15% without one, and clicked a link inside the summary on just 1% of visits.
- Usage is broad. OpenAI reported 900 million weekly ChatGPT users on 27 February 2026, as covered by TechCrunch. OpenAI’s own usage study (15 September 2025) found that practical guidance, seeking information and writing together make up about three quarters of conversations.
The conclusion for research: the question your buyer actually asks is long, specific and often never shows up in a keyword tool. If you only research the short version, you build pages that answer a question nobody is asking the assistant.
What are fan-out queries and why do they matter for research?
Fan-out queries are the extra searches an AI engine runs behind the scenes to answer one prompt. Google’s Search Central documentation on AI features says AI Overviews and AI Mode may issue multiple related searches across subtopics and data sources to build a response, which lets them show a wider set of supporting links. ChatGPT search, Perplexity and Copilot work in a similar way: they rewrite a long prompt into several shorter retrieval queries.
Take the CRM prompt from the introduction. An engine might search for “CRM Google Workspace integration”, “CRM month-to-month pricing”, “best CRM for agencies” and “CRM for small teams reviews”. Your page can be cited for any one of those. That is why fan-out mapping is the center of GEO research: each sub-query is a section, a table row or a page you can own.
Google’s same documentation states there are no special technical requirements for AI Overviews or AI Mode beyond being indexed and eligible to show with a snippet. In other words, fan-out coverage is won with content and normal SEO hygiene, not with a secret file or markup.
How do you do keyword and prompt research for GEO, step by step?
Run these eight steps in order for one product line or service at a time. A first pass for one topic area usually fits in a few working days.
Step 1: Start from revenue topics and seed keywords
List the products, services or use cases that make money, then pull seed keywords for each from Google Search Console, Ahrefs or Semrush. Keep volume and difficulty: they still show where demand concentrates. Cut topics you cannot sell into, however popular they are.
Step 2: Mine the questions people actually ask
Collect real wording from five places:
- Search Console long queries. In the Performance report, filter queries with a regex such as
^(\S+\s){7,}\S+$to see searches of eight or more words. Google reports AI Overviews and AI Mode traffic inside the Web search type, so these queries are a useful proxy for conversational demand. - Sales calls and demos. Ask your sales team for the five questions prospects ask before they buy. These often match prompts word for word.
- Support tickets and site search. Post-purchase questions become pre-purchase prompts for the next buyer.
- Communities. Reddit, niche forums, YouTube comments and review sites show how people describe problems in their own words.
- Google features. People Also Ask, autocomplete and AI Overview follow-up suggestions show related questions.
Step 3: Rewrite keywords into prompts by persona and stage
For each seed keyword, write three to six prompts the way a real buyer would type them. Add a persona (owner, marketing manager, engineer), a constraint (budget, location, compliance) and a stage: learning, comparing or choosing. Keep the wording natural; do not stuff the keyword.
Step 4: Map the fan-out for each prompt
Break every prompt into the sub-questions an engine would have to answer. Ask an assistant directly: “What would you need to search to answer this question?” Then check the answer against the sources it cites and against People Also Ask. Write each sub-query in its own row.
Step 5: Run the prompts and record who gets cited
Test each prompt in ChatGPT, Google AI Mode or AI Overviews, Gemini, Perplexity and Copilot. Log whether your brand is mentioned, cited, recommended or missing, plus the domains and competitors that appear. This tells you whether you need your own page, coverage on a third-party site, or both.
Step 6: Identify the entities each answer depends on
Note the products, brands, standards, places and people that keep appearing in answers. If engines always mention a certification or an integration, your content needs to name it clearly. Our guide to entity SEO explains how to make these connections explicit.
Step 7: Cluster prompts into pages and sections
Group prompts that share the same buyer situation into one page. Turn the fan-out sub-queries into H2 sections, comparison tables and FAQ entries. Give genuinely different situations their own page so they do not compete with each other.
Step 8: Prioritize, publish and track a fixed prompt set
Score each cluster (see the model below), build the top ones first, and keep a fixed list of 30 to 100 tracking prompts that you re-run every month. Changing the list every month makes trends impossible to read.
What does a keyword-to-prompt map look like?
Here is one worked example for a fictional project management tool. Replace the rows with your own topic.
| Seed keyword | Real prompt | Likely fan-out queries | Content that covers it |
|---|---|---|---|
| project management software | “What is the best project tool for a 10-person design studio that works with outside clients?” | client portal project tool; project tool pricing per user; design agency workflow tools | Use-case page for agencies with a pricing table and client access section |
| asana alternative | “We are leaving Asana because of price. What is cheaper but still has timelines?” | Asana pricing; tools with Gantt timeline; cheap Asana alternatives | Honest comparison page with feature and price table |
| project tool gdpr | “Which project management tools store data in the EU?” | EU data residency project software; GDPR compliant project tool | Security page that states hosting regions and certifications |
| project template | “Give me a launch plan template for a website redesign” | website redesign project plan; launch checklist template | Downloadable template with a written checklist on the page |
Which tools help with GEO keyword research?
No single tool covers the whole process, so match each tool to a step. Our full list of SEO tools goes deeper on pricing.
- Google Search Console: long-tail and question queries, impressions for pages that appear in AI features.
- Bing Webmaster Tools: an AI Performance report, launched in public preview in February 2026 according to industry coverage, shows how often Microsoft Copilot cites your pages and the grounding queries behind those citations. It is the closest thing to official fan-out data available today.
- Ahrefs and Semrush: seed keywords, volume, SERP features and People Also Ask, plus AI visibility modules that track brand mentions in AI answers.
- Dedicated prompt trackers such as Profound, Otterly.AI or Peec AI: run a fixed prompt list on a schedule and record mentions and cited sources.
- AlsoAsked and AnswerThePublic: question trees that are a good starting point for fan-out mapping.
- Google Analytics 4: referral sessions from chatgpt.com, perplexity.ai, gemini.google.com and copilot.microsoft.com. OpenAI’s publisher FAQ confirms that ChatGPT adds utm_source=chatgpt.com to referral links.
- The assistants themselves: manual testing is still the most honest check of what a buyer sees.
Treat AI visibility scores from any vendor as samples. Answers change between sessions, locations and accounts, so look at trends across many prompts rather than one result.
How should you prioritize prompts and topics?
Score each prompt cluster from 1 to 5 on four factors, weight them and build from the top. The weights below are a starting point we use; adjust them for your business.
| Factor | Weight | Score 5 when | Score 1 when |
|---|---|---|---|
| Business value | 35% | The prompt comes from people close to buying a core product | Curiosity question with no link to what you sell |
| AI answer likelihood | 25% | Long, comparative or advice-seeking prompts that already show AI answers | Navigational or one-word queries |
| Citation gap | 20% | Engines cite weak, outdated or generic sources | Strong, authoritative sources already dominate |
| Ability to win | 20% | You have unique data, expertise or product facts to add | You can only repeat what others say |
A cluster that scores high on business value but low on ability to win often needs a third-party strategy, such as reviews, partner pages or industry media, rather than another blog post.
What mistakes make GEO keyword research fail?
- Chasing volume only. High-volume short keywords are where AI answers satisfy people without a click.
- Inventing prompts at a desk. Prompts written without sales, support or community input sound like keywords with a question mark.
- One page per prompt. Dozens of thin pages for small wording differences compete with each other and help nobody.
- Ignoring third-party sources. If engines cite review sites and forums for a topic, your own page alone will not change the answer.
- No fixed tracking set. Without the same prompts each month, you cannot tell progress from noise.
- Skipping basics. Pages blocked from crawlers such as OAI-SearchBot or not indexed by Google cannot be used, however good the research is.
Our guide to tracking SEO results covers how to set up the reporting side.
What results has this approach produced?
Our evidence comes from clients in several industries, and we present it as proof of method, not as a promise for your market.
- Planika, a design fireplace brand. Seven US-localized articles took a site Google had not indexed to 1,052 monthly US visits. Ahrefs recorded 401 Google AI Overview mentions, 136 Perplexity mentions and 20 ChatGPT mentions. Read the Planika case study.
- Landingi, a SaaS company. Roughly 5,000 sessions a month referred by AI assistants, tracked in GA4. Read the Landingi case study.
- Galeria Zapachu 24, a fragrance ecommerce store. Starting from DR 1.4 with no link building, 750 product descriptions, more than 60 category descriptions and 30 articles led in 12 months to more than 9,100 ranking keywords, an estimated 21,200+ monthly visits (Ahrefs), a Top 10 position for “perfumy” in Poland and a 35% higher average order value. Read the Galeria Zapachu case study.
The common thread: a small number of pages that fully answer the questions buyers ask beat a large library of thin content.
Who should work with Marketer Coffee on GEO research?
We fit companies that sell in several markets or languages and want prompt research tied to revenue, not to vanity scores. We work in English, Polish, German, French, Spanish, Italian, Swedish and Danish, which matters because prompts and cited sources differ by language.
Who is not a fit: a business that only needs a one-off keyword list, or a single local shop that would get more from a basic local SEO setup.
Start with our free AI visibility audit to see which prompts already mention your brand and which do not. Details on the ongoing work are on our AI SEO service page.
Frequently asked questions about keyword research for GEO
Do I still need keyword tools for GEO?
Yes. Keyword tools show where demand and commercial intent are, and the fan-out queries engines run are often short, searchable phrases. Use them as the starting point, then add prompts, fan-out mapping and citation checks on top.
How many prompts should I track?
Most companies start with 30 to 100 prompts that cover core products, comparison questions and brand questions. Keep the set fixed for at least a quarter so trends are readable, and add new prompts in batches.
Can I see the real search volume of ChatGPT prompts?
No public source reports it. You estimate demand from proxies: long Search Console queries, community threads, sales and support questions. Bing Webmaster Tools shows grounding queries for Copilot citations, which is the closest official signal.
What is a fan-out query?
It is one of the extra searches an AI engine runs to answer a longer prompt. Google documents that AI Overviews and AI Mode may issue multiple related searches across subtopics to build a response.
How long until GEO research shows results?
New or improved pages can start appearing in AI answers within weeks of being indexed, and steady citation growth usually takes a few months. These are typical timelines, not guarantees.
What should you remember about keyword research for GEO?
Keywords show you where the money is. Prompts show you how buyers ask. Fan-out queries show you what an engine must find to answer. Research all three, build fewer and better pages, and track a fixed prompt set every month alongside revenue.
If you want a second pair of eyes on your prompt map, or want us to build it with you across several markets, send us your site and your top products. Book a free consultation.
Sources: Google Search Central, AI features and your website (documentation, accessed November 2026); Google I/O 2026 AI Mode figures, as reported by Talk Android (20 May 2026); Pew Research Center, Google users are less likely to click on links when an AI summary appears in the results (22 July 2025); Semrush, ChatGPT search trends study of 80 million clickstream records (3 February 2025, vendor data); OpenAI, How people are using ChatGPT (15 September 2025); OpenAI 900 million weekly users, as reported by TechCrunch (27 February 2026); OpenAI Help Center, Publishers and developers FAQ; Bing Webmaster Tools AI Performance report, as reported by Search Influence (2026); Marketer Coffee case studies: Planika, Landingi, Galeria Zapachu 24 (other industries).
