How do I research which prompts to track for generative engine optimization?
A step-by-step method to research which prompts to track for generative engine optimization, from customer questions to fan-outs, volumes and personas.
Prompt research is to generative engine optimization what keyword research is to SEO: the list you choose decides everything you measure afterward. Pick prompts nobody asks and your visibility score is meaningless. Pick only brand prompts and you will feel great about a metric that says little. This guide walks through a method that produces a prompt set you can defend.
Why prompts are not keywords
Keywords are short and stable. Prompts are conversational, longer and vary by person, and AI systems usually break each one into several narrower searches, called query fan-outs, before answering. So the question "which prompts should I track?" has two layers: the prompts people type, and the sub-queries the model runs behind them. Good research covers both.
Step 1: start from customer questions
Collect questions from the places buyers already speak:
- Sales call notes and demo objections.
- Support tickets and onboarding questions.
- Reviews of you and your competitors.
- Forum and community threads where your category comes up.
- The "People Also Ask" boxes for your main topics.
Write them as a person would type them into ChatGPT, not as you would phrase a keyword. "Best accounting software for a 10-person agency that bills in euros" is a prompt. "Accounting software agency" is a keyword.
Step 2: sort by type
Group prompts so you can read results in meaningful slices.
- Category prompts: "what are the best tools for X". Highest value, most competitive.
- Comparison prompts: "X vs Y", "alternatives to Y". Close to purchase.
- Problem prompts: "how do I fix X". Earlier in the journey, often informational.
- Brand prompts: your name, reviews, pricing. Useful for accuracy and sentiment checks, but do not let them dominate the set.
Add intent (informational or commercial) and topic tags. Without tags, a flat list of 100 prompts cannot be analyzed.
Step 3: add personas and places
The same question gets different answers for different people. A founder, a procurement lead and an agency owner phrase things differently and may get different recommendations. Add the personas you sell to, and the countries or cities you sell in, because answers vary by location. Tracking one generic version of each prompt hides these differences.
Step 4: expand with fan-outs
Look at what the model actually searches. The fan-out queries show which sub-questions matter, which sources the model reaches for and which words it uses. Fan-outs frequently reveal prompts you would not have written, and they show whether specific domains, years or qualifiers get added. Feed those back into the set.
Step 5: check demand and difficulty
Not every prompt deserves tracking. Where available, use volume and difficulty estimates to prioritize. Treat any volume figure as a modeled estimate rather than a count, and use it to rank prompts against each other, not as a forecast.
Step 6: right-size the set
Quality beats quantity. A focused set of prompts tied to real buying questions, tagged by topic, persona and intent, beats a thousand loosely related ones. Start with a core of the prompts closest to revenue, then widen once you see which areas move.
Tools that help with prompt research
Promptwatch has the most complete prompt research workflow among the tools we reviewed. Prompt Explorer suggests prompts from Search Console, ranking keywords, People Also Ask or any page you point it at. Prompt Volumes gives search volumes, difficulty scores and citation rates across AI search, Google and Bing. Query fan-outs are shown per model, and prompts can be organized with topics, tags, types (organic, brand, competitor comparison) and intent. Personas support persona stacking, so you can track how answers differ by buyer type. The free GEO Inspector Chrome extension shows ChatGPT fan-outs and fetched versus cited sources. Plans start with a free Explore plan (10 prompts, ChatGPT only), then Essential at $95 a month with 50 prompts, Professional at $245 with 150, and Business at $579 with 350. Agency plans include unlimited prompts.
Profound offers a Prompt Volumes dataset of real query demand, which the catalog calls genuinely unique, though it sits inside an Enterprise product around $40k a year. AthenaHQ supports unlimited topics and prompts on its Starter plan at $295 a month, with persona targeting gated to Enterprise. Mentionable captures fan-out queries and drafts the page you are missing, from 79 EUR a month. Aiso builds prompts and follow-ups from a consent panel of real human-AI conversations rather than a keyword list, from $19 a month. Scrunch AI includes personas but the catalog notes personas and prompts take significant setup.
Who should pick which
Promptwatch for a team that wants suggestion, volume, fan-out and persona work in one workflow, then wants to track and act on the same set.
Profound if you are an enterprise buyer and the volume dataset is the reason you are buying.
Aiso if where the prompts come from matters most to you, and you accept a lighter tracker.
Mentionable if fan-out capture plus page drafting at a modest price fits your workflow.
Keep the set alive
Review the prompt set monthly. Retire prompts that never trigger a mention in any answer and no longer match what customers ask. Add new ones from fresh sales conversations and from the fan-outs that appear in new answers. Keep a change log, so a score change after you edit the set is not mistaken for a real movement.
No prompt set can guarantee a recommendation. A good one gives you a stable, honest picture of where you stand in the conversations your buyers are already having.