How do I research which prompts to track for generative engine optimization?
How our team researches which prompts to track for generative engine optimization: customer questions, personas, fan-outs and a set small enough to act on.
The prompt list is the most consequential decision in a GEO program, and it is usually made in a hurry. We have learned to slow down at this step, because everything measured afterward inherits its flaws. A set full of questions nobody asks produces a flattering, useless score. A set dominated by brand prompts produces a different flattering, useless score.
Here is how we research the list for a client.
Prompts are not keywords
Keywords are short and stable. Prompts are conversational, longer and phrased differently by different people. On top of that, AI systems typically split a prompt into several narrower searches, called query fan-outs, before writing an answer. So we research two layers: what people type, and what the model searches behind it.
Our process
1. Mine real customer language
We start from conversations, not tools. Sales call notes, demo objections, support tickets, reviews of the client and its competitors, and community threads where the category comes up. We write each question the way a person would type it into ChatGPT. "Best project software for a 15-person design agency that invoices in euros" is a prompt. "Project software agency" is a keyword.
Then we add Search Console queries and People Also Ask boxes to catch questions the client has not heard directly. In Promptwatch, Prompt Explorer can suggest prompts from Search Console, ranking keywords, People Also Ask or any page we point it at, which saves a day of manual collection.
2. Classify before tracking
A flat list is hard to read. We tag every prompt by type (category, comparison, problem, brand), intent (informational or commercial) and topic. Promptwatch supports topics, tags and prompt types such as organic, brand and competitor comparison. Tagging up front is what lets us tell a client "you win on problem prompts and lose on comparisons" instead of reporting one blended score.
We cap brand prompts. They are useful for checking accuracy and sentiment, but they inflate the average if left unchecked.
3. Add personas and locations
The same question can produce different answers for different buyers and in different markets. We list the personas the client sells to, and the countries or cities that matter, and track the important prompts for each. Promptwatch personas support persona stacking, and tracking is available by country on every paid plan, with state and city from the Business plan and on agency plans.
4. Read the fan-outs
Once tracking runs, the fan-outs tell us what the model actually searched. They surface prompts we would not have written and show which sources, qualifiers and years the model adds. Promptwatch shows fan-outs per model, and the free GEO Inspector extension shows ChatGPT fan-outs and fetched versus cited sources. We feed what we find back into the prompt set.
5. Prioritize with demand and difficulty
We do not track everything. Prompt Volumes in Promptwatch gives search volumes, difficulty scores and citation rates across AI search, Google and Bing. We treat those as modeled estimates, useful for ranking prompts against each other and not as forecasts we would promise a client.
How big should the set be?
Smaller than people expect. A focused set tied to real buying questions beats hundreds of loosely related prompts. We begin with the prompts closest to revenue and widen once we see which areas move. Plan limits are a practical constraint too: Essential includes 50 prompts, Professional 150 and Business 350, while agency plans include unlimited prompts, which suits programs covering several clients.
Maintaining the set
- Monthly review. Remove prompts that never trigger any mention and no longer reflect buyers' questions.
- New additions. Pull fresh questions from recent sales calls and from new fan-outs.
- Change log. Whenever we edit the set, we record it. Otherwise a score movement after an edit gets mistaken for a real change in visibility.
What we tell clients up front
A prompt set is a sample. It is not the full space of conversations, and it cannot promise any engine will recommend the client. What it provides is a stable, honest baseline that we can compare month to month and tie to content and citation work.
We build and maintain these sets in Promptwatch because the suggestion, tagging, persona, volume and fan-out steps sit in one place, and the same set then drives tracking, citation analysis and reporting. When research and measurement live in separate tools, the prompt list drifts, and a drifting list is how a program ends up proudly reporting progress on questions nobody asks.