Write a customer FAQ from real questions with ChatGPT

A useful FAQ answers the questions customers actually ask, in the order they ask them, with answers the company stands behind. ChatGPT gets you there in about 25 minutes if you give it two things: a month of real questions and the documents that contain the answers. The model groups, merges and writes. It must never be the source of a delivery time or a refund rule, and the prompt is built to stop it from becoming one.

Collect the real questions

Go through the last month of email, chat logs and call notes and copy out 30 to 50 customer questions as they were asked. Strip names, phone numbers and order numbers; only the question text goes into the service. Resist the urge to write questions you think customers have. The real list is usually more practical ("can I change the delivery address after paying?") and more repetitive, and the repetition tells you which topics deserve to be at the top.

Then gather the documents that hold the answers: the price list, delivery terms, return policy and terms of sale.

The prompt

Role: you are a customer support manager at a company that {{what it does}}. Context: below is a list of real customer questions and the company documents: {{list of documents}}. Tone with customers: {{informal or formal, warm or businesslike}}. Task: group the questions by topic, merge questions with the same meaning, and answer each in up to three sentences based only on the documents. If the documents have no answer, write [not in the documents] and state what needs to be clarified. Add to each topic one question customers have not asked yet but will. Format: sections by topic, each with the question in bold and the answer. Plain, friendly language, no bureaucratic phrasing and no long dashes. Questions: {{questions}}

Attach the documents as files or paste them above the questions. The tone field matters for the brand: an online shop and a B2B supplier answer the same return question very differently.

Working through the draft

The draft comes back grouped by topic with short answers and a few [not in the documents] markers. Those markers are the most useful output. Each one is a question customers ask that your written policies do not answer, which means your support team answers it differently every time. Decide the answer, write it down, and consider adding it to the policy itself.

Then check every answer that contains a number, a deadline or a price against the source documents. The model occasionally carries a figure from one section into another, such as a delivery time for one region applied to all of them. Finally, look at the "questions customers will ask" it added to each topic; keep the good ones and delete the speculative ones.

What to do

1. Collect 30 to 50 real customer questions from the last month and remove names and contacts.
2. Gather the price list, delivery terms, return policy and terms of sale.
3. Run the prompt in a new chat with the documents and questions.
4. Fill every [not in the documents] gap with an answer you decide and confirm.
5. Check each answer with a number or deadline against the documents.
6. Publish the FAQ and add new questions to it every month.

How to know it is working

Every answer rests on a company document or your confirmation, questions are grouped so a customer finds theirs in ten seconds, and repeat inquiries on those topics start to fall. The last one is the real measure; watch your inbox for a month after publishing.

Plans and settings

One FAQ fits the free plan's limits. Plus at 20 USD per month adds Projects, where the documents stay attached for the monthly update; Business at 25 USD per seat with annual billing keeps team data out of training by default. See ChatGPT pricing. On personal plans, turn off Improve the model for everyone in Data controls first. Marketers can use the saved version on the customer FAQ page for marketers, and the template itself is on the FAQ prompt page.