# Building a Prompt Set — AI Visibility and LLM Brand Discovery

Source: https://www.geekswithgeeks.com/en/ai-visibility/meas-prompt-set

> Collect the real questions customers ask across funnel stages, languages and personas.

## Ask what customers ask

Build a list of 30 to 100 prompts that reflect real questions. Cover **discovery** ("best project tracker for small teams"), **comparison** ("Acme Tasks vs Trello"), **evaluation** ("is Acme Tasks good for agencies?", "Acme Tasks pricing"), **problem-led** ("how to manage tasks without email chaos") and **branded** questions ("what is Acme Tasks?"). Include different personas and also questions in **Hindi, Hinglish** and other languages your customers use. Source ideas from support tickets, sales calls, search queries and community threads, and keep the list fixed so results are comparable over time.

## Ask, record, count, compare

Answers vary, so measure with a fixed set of prompts, repeated runs and ranges.

![Four steps: prompts, runs, metrics, trends.](assets/figures/ai-visibility/section-4-map.svg) — Figure 4.1 — Prompts, runs, metrics and trends.

## A prompt set as data

Tags let you report results by intent, language and persona.

```json
[
  {"id": 1, "text": "best lightweight project tracker for a 5-person team", "intent": "discovery", "lang": "en"},
  {"id": 2, "text": "Acme Tasks vs Trello for a small agency", "intent": "comparison", "lang": "en"},
  {"id": 3, "text": "छोटी टीम के लिए सबसे आसान task manager कौन-सा है?", "intent": "discovery", "lang": "hi"},
  {"id": 4, "text": "what is Acme Tasks and how much does it cost?", "intent": "branded", "lang": "en"}
]
```

## Write prompts like users, not marketers

Real people write messy, specific, conversational questions. Test those, not only polished keyword phrases.

**Quiz:** Why keep the prompt set fixed over time?

- [ ] Prompts expire if edited
- [x] So changes in results reflect real change, not different questions
- [ ] It is required by assistants
- [ ] Fixed sets are cheaper to buy

*Answer:* So changes in results reflect real change, not different questions. A constant benchmark lets you compare before and after a change.