AI SEARCH (GEO)

Turn real questions into a quotable FAQ

Use when you have a list of real customer questions and want an FAQ assistants can quote.

quotable-faq-for-assistants.md

Download .md

You are turning real user questions into an FAQ block designed to be quoted by AI assistants.

Raw questions from support, sales or search: {{RAW_QUESTIONS}}
What we can truthfully answer: {{SOURCE_MATERIAL}}
Product or topic: {{TOPIC}}

Steps:
1. Cluster the raw questions into distinct intents. Show the clusters with the raw questions grouped under each.
2. For each cluster, write one canonical question in the phrasing a user would actually type.
3. Write an answer of 40 to 80 words. First sentence answers directly. Second sentence gives the condition or exception. Third gives a concrete example or number if one exists in {{SOURCE_MATERIAL}}.
4. Rank the finished questions by how often the cluster appeared in {{RAW_QUESTIONS}}, most first.

Output as:
| Rank | Question | Answer | Source fact used | Confidence (high/medium/low) |

Constraints:
- Never answer from general knowledge. If {{SOURCE_MATERIAL}} does not cover a cluster, output the question with the answer "NOT ANSWERABLE - needs input from {{TOPIC}} owner".
- Do not merge two genuinely different intents to shorten the list.
- No em dashes, no "great question", no sales language in answers.

Fill in before running

Replace each placeholder with your own detail. The more specific you are, the less the model invents.

{{RAW_QUESTIONS}}
{{SOURCE_MATERIAL}}
{{TOPIC}}

Getting a better result

Keep the user phrasing in the question even when it is grammatically awkward.
Answers that lead with a condition instead of the answer rarely get quoted.
Drop any question your sales team invented; only real ones earn a slot.

Questions about this prompt

When should I use this rather than the topic coverage map?

When you already hold real questions from support tickets, sales calls or site search. The coverage map generates a question set when you have none. Use this when the list exists and the job is turning several hundred untidy versions of the same question into a handful you answer once, properly.

What do I need in front of me?

The raw questions in the wording people used, duplicates included, because ranking is by cluster size and deduplicating first destroys that signal. You also need source material to answer from. The prompt is barred from answering out of general knowledge, so anything {{SOURCE_MATERIAL}} misses comes back unanswered.

What comes back, and which rows matter most?

Clusters, one canonical question each, answers of 40 to 80 words, and a ranked table with the source fact used and a confidence level. The NOT ANSWERABLE rows are the most valuable output. Those are questions asked often enough to form a cluster that nobody at your company has answered.

What is the mistake that costs me?

Rewriting the canonical questions into tidy marketing English before publishing. The phrasing people type is the phrasing that matches, awkward grammar included. The second habit worth resisting is padding the list with questions sales invented, which dilutes the ranking and pushes genuinely common questions down it.