--- title: "How to Research Prompts for GEO: A Step-by-Step Method" slug: "how-to-research-prompts-for-geo" category: "ai-visibility" canonical_path: "/articles/ai-visibility/how-to-research-prompts-for-geo" meta_title: "How to research prompts for GEO: a 7-step method — Prime AI Visibility" meta_description: "How to research prompts for GEO: collect real buyer questions, expand them by role and stage, score them, test them in the assistants and freeze a versioned set." author: "Bob Generale" date: "2026-10-25" last_updated: "2026-10-25" read_time: "12 min" keywords: - how to research prompts for geo - GEO prompt research - buyer question research - prompt set for AI visibility - how to choose prompts to track featured_image: "/brand/articles/ai-visibility/how-to-research-prompts-for-geo.png" featured_image_alt: "A pale slab with a loose scatter of small blank cubes behind and a tidy front row of larger cubes, one of them amber" og_image: "/brand/social/articles/ai-visibility/how-to-research-prompts-for-geo.jpg" cta_mid_headline: "Test your first five prompts free" cta_mid_body: "Run five buyer questions across ChatGPT, Claude and Perplexity free and keep each saved answer with the links it displayed. Use what you see to cut or rewrite weak prompts." cta_mid_button: "Run my free check" cta_bottom_headline: "Freeze a prompt set you can repeat" cta_bottom_body: "Create a workspace, bring the questions that survive your scoring, and compare the same saved answers next month. A fixed set is what makes changes readable." cta_bottom_button: "Create a workspace" --- # How to Research Prompts for GEO: A Step-by-Step Method How to research prompts for GEO comes down to this: collect the questions your buyers really ask, expand them by role and decision stage, remove the ones that do not match a real purchase, score what remains, test it in the assistants and freeze a versioned set of 20 to 40 prompts. Then repeat the same set on a schedule. No public source gives you the true frequency of any prompt, so the goal is realistic coverage of buyer decisions, not a ranked list of volumes. In this guide a "prompt" means a question or request a buyer types into an assistant. GEO, short for generative engine optimization, is the work of making your brand and pages available to answers like these. The method below is professional judgement built on first-party sources such as sales notes and Search Console. It is not a rule from any search engine or AI provider. ## How to research prompts for GEO: the seven steps 1. **Collect real wording.** Pull questions from sales calls, support tickets, lost-deal notes and search data. 2. **Expand by role and stage.** Ask who is buying and what they need to decide. 3. **Clean and cluster.** Merge near-duplicates and drop anything no buyer would say. 4. **Score each prompt.** Rate realism, decision impact, ownership and testability. 5. **Test in the assistants.** Run each prompt and note what was displayed. 6. **Freeze and version.** Fix the set so month-to-month changes are comparable. 7. **Review on a schedule.** Add or retire a few prompts each quarter and keep the history. ## What is a good prompt for GEO? A good prompt passes three checks. A real buyer in your segment would plausibly ask it during a decision. The answer could influence whether they shortlist, choose or reject you. And it is phrased the way a person writes it, in full sentences with the constraints that matter, such as team size, budget or location. Compare two versions. "AI visibility" is a topic, not a prompt. "Which tool should a five-person marketing team use to see whether ChatGPT names our brand?" is a prompt, because it has a buyer, a constraint and a decision. Prompts are usually longer and more specific than search queries, so a keyword list is a starting point and not a finished set. ## Where do you find real prompts? Start from evidence you already own. Each source below shows what people actually say. | Source | What it gives you | Watch out for | | --- | --- | --- | | Sales calls and demo notes | Exact buyer wording and objections | Small samples from one rep | | Support tickets and chat logs | Post-purchase and how-to questions | Mostly existing customers | | Lost-deal reasons | Comparison and "why not you" questions | Notes are often one line | | Search Console queries | Queries that already reach your site | Not prompts, and shows only your own visibility | | Site search logs | What visitors fail to find | Low volume on small sites | | Community threads and reviews | Plain language and complaints | Public posts may not represent buyers | | Competitor and comparison pages | The categories buyers weigh | Do not copy their framing as fact | Google documents the Search Console [performance report](https://support.google.com/webmasters/answer/7576553), which lists queries and pages that earned impressions and clicks for your site. It is a useful seed, but it shows how people searched, not what they typed into an assistant. Treat a long query as a hint to be rewritten as a full question. Do not use a vendor's "prompt volume" as proof of demand. Providers do not publish assistant query logs, so any volume figure is modelled or sampled from a panel. The guide to [what prompt volume means in AI visibility products](https://primeaivisibility.com/articles/ai-visibility/what-is-prompt-volume-in-ai-visibility-products) explains the data sources and what to ask a vendor. ## How do you expand a seed list into a full set? A seed list from one team is usually narrow. Expand it along four axes. - **Role.** A founder, a marketing lead and an IT reviewer ask about the same product differently. - **Stage.** Early buyers ask "what is X". Mid-stage buyers ask "best X for Y" and "X vs Z". Late-stage buyers ask "is [brand] any good" and "[brand] alternatives". - **Constraint.** Add size, budget, region, compliance or integration needs. - **Brand.** Write some prompts that name no brand and some that name you or a rival. Keep every prompt in plain words. Remove names of clients or individuals, and anything confidential, before the set is shared. An example in another field shows the shape. [Pulse Terminal](https://pulseterminal.ai/) is a public investing research app, and its site describes grading stocks and ETFs. It is used as an illustration only. It is not a Prime client, this is not an endorsement or a claim about any AI result, and the prompts are invented. Its [pricing page](https://pulseterminal.ai/pricing) is linked as the place a buyer would check plans. A team in that field might write the following. | Stage | Role | Invented prompt | | --- | --- | --- | | Early | Beginner investor | What should a new investor look for in a stock research app? | | Mid | Active investor | Which stock research app explains how a stock is graded? | | Mid | Active investor | How does a grading app compare with a screener? | | Late | Buyer checking one brand | Does a stock research app show sources under its answers? | ## How do you clean and score the list? Merge prompts that ask the same thing in different words, then read each one aloud. If no buyer would say it, drop it. Then score what remains from one to three on four criteria. | Criterion | Question to ask | Score 1 | Score 3 | | --- | --- | --- | --- | | Realism | Would a buyer type this? | Unlikely | Heard in calls | | Decision impact | Could the answer change a shortlist? | Trivia | Directly affects choice | | Ownership | Could a page of yours answer it? | No page possible | A page exists or could | | Testability | Can you repeat it exactly? | Needs context | Self-contained | Keep prompts with a total of ten or more, and look at the rest by hand before dropping them. The scores are a judgement aid, not a measurement. Two people will score some prompts differently, so have a second person rate a sample of ten and discuss the gaps. ## How do you test prompts before you commit? Run every candidate in the assistants your buyers use and save the full answer with the links shown. You are checking the prompt, not yet your brand. - Does the answer address the question you meant, or a different one? - Does the same prompt give a similar kind of answer on a second run? - Does it name brands at all? A prompt that returns a general explanation and no brand cannot show visibility. - Does it show links, so you can see which sources were displayed? Rewrite prompts that miss the point, and drop those that never name or link anything. Use an unpersonalised temporary chat where you can, because OpenAI documents that memory and custom instructions can change responses. The guide to [monitoring brand mentions in ChatGPT](https://primeaivisibility.com/articles/ai-visibility/monitoring-brand-mentions-in-chatgpt) shows that setup. ## How many prompts do you need? For a first baseline, 20 to 40 prompts across the stages is enough for most teams, and ten is enough to learn. More prompts add cost and reading time without adding clarity until you have a process to act on what you find. Split the set so about a third are early-stage, a third mid-stage and a third late-stage, and keep a few brand-named prompts in the late group. ## How do you freeze and version the set? Give the set a version number and a date, and store it where the whole team can read it. Do not change wording within a month. A reworded prompt is a different test, and a change in the answer could come from the wording and not from anything you did. When you review each quarter, change only a few prompts, note why, and bump the version. Keep the old version's answers. The history is what lets you say what moved. ## What should you track after you pick prompts? Keep five signals separate: brand named, site cited, recommendation, referral and conversion. A prompt that names you with no link behaves differently from one that links your page. Prompt research feeds all five, but it does not measure any of them on its own. The guide to [running an AI visibility audit](https://primeaivisibility.com/articles/ai-visibility/how-to-run-an-ai-visibility-audit) shows how to read the saved answers, and the [AI visibility strategy](https://primeaivisibility.com/articles/ai-visibility/ai-visibility-strategy) guide shows how to pick what to measure first. ## Where do tools fit? A tool saves the answers on a schedule so you do not copy them by hand, and some suggest prompts. Check where a suggestion comes from before you trust it. Percepture, a company connected to [Prime AI Visibility](https://primeaivisibility.com/), compares the [AI visibility tools](https://percepture.com/geo-insights/ai-visibility-tools/) market and covers the measurement side, and its [generative engine optimization services](https://percepture.com/services/geo-services/) cover the page and structure work that follows from the answers. Because of the connection, compare it as you would any provider. [Prime AI Visibility](https://primeaivisibility.com/) works as an [AI Search Visibility Platform](https://primeaivisibility.com/) for the saved-answer step. The free check runs up to five buyer questions across ChatGPT, Claude and Perplexity, keeps up to 15 answers with the links each displayed and needs no card. It is one check a month. Starter is $69 a month and adds Gemini and an AI Overview-style preview, so all five assistants are covered. The [features page](https://primeaivisibility.com/features) describes captured answers, brand named, website cited, recommendation, multi-model tracking and Content Studio, and these are separate measures. It has no sentiment, crawler or traffic reporting, no managed service, no auto-publishing and no prompt volumes, so it does not suggest or rank prompts for you. As an [AI Visibility Platform](https://primeaivisibility.com/) it suits a team that brings its own question set. Use it as an [AI visibility checker](https://primeaivisibility.com/) to test the first five prompts, then as an [AI visibility tracker](https://primeaivisibility.com/) to repeat the frozen set. As an [AI Brand Visibility Tool](https://primeaivisibility.com/) it counts brand named, site cited and recommendation. For the buying decision, read the guide to [choosing a tool](https://primeaivisibility.com/articles/ai-visibility/how-to-choose-an-ai-visibility-tool). ## What mistakes should you avoid? - **Copying a keyword list.** Keywords are short. Rewrite them as full questions. - **Only asking about yourself.** Most buyers start without a brand in mind. - **Trusting a volume number.** Treat any volume figure as modelled until the source is shown. - **Changing wording mid-month.** It breaks comparison. - **Writing prompts that lead the answer.** "Why is our product the best" tests flattery, not visibility. - **Judging from one run.** Answers vary, so repeat before you act. > **Test your first five prompts free** > > Run five buyer questions across ChatGPT, Claude and Perplexity free and keep each saved answer with the links it displayed. Use what you see to cut or rewrite weak prompts. > > **[Run my free check](https://app.primeaivisibility.com/sign-up)** ## Frequently asked questions **How do I research prompts for GEO?** Collect real buyer wording from sales, support and lost deals, expand it by role and stage, score the candidates, test them in the assistants and freeze a set of 20 to 40 prompts. **Can I use keyword research for GEO prompts?** As a seed, yes. Keywords are short and prompts are full questions with constraints, so rewrite them and add the buyer's situation. **Is there a tool that gives true prompt volume?** No public source publishes assistant query logs. Vendor volumes are modelled or panel-based, so treat them as relative signals. **How many prompts should I track?** Ten is enough to learn from and 20 to 40 is enough for a first baseline. Add more only when you can act on what you find. **How often should I update the set?** Review quarterly, change a few prompts, and keep the earlier version's answers. ## References 1. Google Search Central, "SEO Starter Guide". 2. Google Search Console Help, "Performance report (Search)". 3. OpenAI Help Center, "Memory in ChatGPT". 4. OpenAI Help Center, "Temporary chat in ChatGPT". 5. Percepture, "AI visibility tools". 6. Percepture, "Generative engine optimization services". 7. Pulse Terminal, "Homepage" (illustrative example only). 8. Pulse Terminal, "Pricing" (illustrative example only). ## Next steps 1. **[Run the audit process](https://primeaivisibility.com/articles/ai-visibility/how-to-run-an-ai-visibility-audit)** on your first five prompts and read what the answers link. 2. **[Read the strategy guide](https://primeaivisibility.com/articles/ai-visibility/ai-visibility-strategy)** to decide which signals to track first. 3. When you are ready, **[create a workspace](https://app.primeaivisibility.com/sign-up)** in **[Prime AI Visibility](https://primeaivisibility.com/)** and bring your frozen question set. > **Freeze a prompt set you can repeat** > > Create a workspace, bring the questions that survive your scoring, and compare the same saved answers next month. A fixed set is what makes changes readable. > > **[Create a workspace](https://app.primeaivisibility.com/sign-up)**