How We Build a Client Prompt Set With Volume and Difficulty Data
The process we use to turn a new client's keywords, sales calls, and Search Console data into a tracked prompt set for ChatGPT, Gemini, Claude, and Perplexity, and how volume and difficulty decide what makes the cut.
The prompt set is the most important deliverable in the first month of a GEO engagement, and it's the one clients see least. Every report we send afterwards, every content brief, every share of voice chart, depends on whether those prompts reflect how the client's buyers actually talk to AI assistants.
We've settled on a process that takes about two weeks and produces a set we can defend line by line. Here's how it works.
Week one: collect more candidates than we need
We start wide. The goal in the first week is a long list of candidate questions, roughly three to five times what we'll end up tracking.
Sales and support language. We ask for a sample of support tickets, chat transcripts, and notes from discovery calls. We're looking for repeated questions and the constraints buyers attach to them: team size, budget, existing tools, location. Those constraints are what make a prompt realistic.
The client's Search Console. We connect Google Search Console to Promptwatch, which pulls queries in as candidates in Prompt Explorer. We treat them as topics, not prompts. "Payroll software netherlands" becomes "what payroll software works for a Dutch company with fewer than 20 staff?"
People Also Ask and ranking keywords. Prompt Explorer also suggests prompts from the client's ranking keywords, People Also Ask, manual searches, and pages on their site. The page-based suggestions are useful in a kickoff: they show the client which questions their existing content claims to answer.
Public discussion and reviews. We read the subreddits and review pages for the category. Phrases like "alternative to" and "better than" tend to show up in recommendation prompts.
AI autocomplete. For the core category questions, we use the autocomplete in Promptwatch's Prompt Volumes view to find the higher-demand phrasings of a question we already know matters.
Week two: cut with data
This is where volume and difficulty earn their place. Each candidate gets checked against three things.
Volume band. Promptwatch estimates volume from the keywords attached to a prompt, weighting AI search demand five times, Bing 1.5 times, and Google 0.5 times, and shows it as one to five bars. We drop candidates with no volume, with two exceptions we'll come to.
Difficulty. The score reflects the past 30 days of keyword competition plus the authority of sources AI already cites for that prompt. We don't drop high-difficulty prompts automatically. We flag them, because they usually need offsite work rather than content, and the client needs to know that before we promise movement.
Intent and type. We tag every prompt as Organic, Brand Specific, or Competitor Comparison, and by intent. That tagging is what lets us report share of voice by segment later instead of one blended number.
The two exceptions to the volume rule are brand accuracy checks (prompts that exist to catch wrong pricing or missing features) and competitor comparisons the sales team hears weekly. Both stay in regardless of volume, because their value isn't reach.
How big the set ends up
Promptwatch's own guidance is to start with 15 to 30 prompts per monitor or market and expand from what moves the needle. That matches our experience. For a single-market client with one product, we usually launch with one or two monitors in that range. Multi-market clients get a monitor per market, because buyers in different countries and languages phrase the same need differently.
We'd rather launch small and add prompts in month two than launch with 200 and spend the first report explaining why half of them show nothing.
What we show the client at sign-off
Before tracking starts, the client sees the full set in a simple table: prompt, segment, volume band, difficulty, and our reason for including it. Anything kept despite low volume has its reason written next to it.
That table saves arguments later. When a stakeholder asks why their favorite question isn't tracked, the answer is on the page.
Why we run this in Promptwatch
We've built prompt sets in spreadsheets before. The problem is that the research and the tracking end up in different places, and the volume numbers go stale the moment the sheet is saved. Running the whole process in Promptwatch means every tracked prompt keeps its volume, difficulty, and tags, and Search Console stays connected for the next round of candidates. On agency plans, prompts and projects are unlimited, so the cap we work to is the response budget, which starts at 10,000 a month on Kick-off ($199/mo).
The prompt set gets revisited every quarter, in line with how often the underlying keyword volumes refresh. If you want us to build one for your brand, get in touch.