What Is a Prompt Corpus?
A prompt corpus is the curated, versioned set of prompts a brand runs against AI engines to measure its visibility. It is the sampling frame for every downstream metric — mention rate, citation rate, share of voice all mean "across the corpus" — so the corpus defines what your visibility program can and cannot see. A biased corpus produces confidently wrong numbers.
How do you build a corpus from personas and funnel stages?
The reliable method is a matrix, not a brainstorm. One axis: buyer personas (the freelancer, the ops manager, the enterprise buyer), because engine answers shift with user framing. The other axis: journey stages, each with characteristic prompt families:
| Stage | Prompt family | Example |
|---|---|---|
| Problem discovery | how/why questions | "why is our churn rate so high" |
| Solution education | what-is / definitional | "what is revenue operations software" |
| Vendor research | best-of / category | "best RevOps tools for B2B SaaS" |
| Comparison | X vs Y | "Clari vs Gong for forecasting" |
| Switching | alternatives / migration | "alternatives to Salesforce for a 20-person team" |
Fill each cell with phrasings mined from real sources — sales calls, tickets, community threads — a process prompt research tools accelerate. Weight commercial and switching prompts deliberately: they are lower volume but carry the pipeline.
Why must the corpus be versioned and maintained?
Buyer language drifts. New competitors enter, category vocabulary shifts, and last year's phrasing stops matching how users actually talk to assistants. Mature programs review the corpus quarterly, adding mined phrasings and retiring dead ones — while keeping a stable core subset untouched so long-term trends stay comparable. Every corpus change should be logged, because a visibility jump caused by adding easier prompts is an artifact, not a win.
What makes a corpus honest?
Three properties: it is sourced from observed buyer language rather than invented; it is unbranded-weighted, since unbranded category prompts are where recommendations are actually contested; and it is distributionally realistic — long conversational phrasings, not keyword stubs, because that is how people query assistants.
Example
A cybersecurity vendor's initial 80-prompt corpus was 70% branded queries and showed a flattering 85% visibility. Rebuilt from sales-call mining to 260 prompts weighted toward unbranded and switching intent, measured visibility fell to 14% — the honest baseline the program actually needed.
Frequently asked questions
- How large should a prompt corpus be?
- Large enough to cover personas, funnel stages, and prompt families without becoming unmaintainable — typically 100-500 prompts for a single product category. Depth of sampling per prompt matters more than raw corpus size.
- Where do good prompts come from?
- Real buyer language: sales-call transcripts, support tickets, site-search logs, Reddit and community threads, and People Also Ask data. Prompts written by the marketing team from intuition consistently miss how buyers actually phrase questions.
Keep exploring
See how AI engines talk about your brand — track mentions across ChatGPT, Perplexity, Claude, Gemini and 5 more. Start with Menra