Most FAQ pages are written the wrong way around: someone in marketing sits down and imagines what customers might ask. The result is a page that answers “What makes your product different?” while the support queue drowns in “Why was I charged twice?”, a question the FAQ never mentions.

Your ticket queue already contains the real FAQ. Customers have been writing the questions for months, and your agents have been writing (and refining) the answers. To build FAQ from tickets, you need a counting exercise, a drafting pass, and a maintenance habit. This guide walks through all three, gives you copy-paste LLM prompts for the drafting step, and covers the part most posts skip: measuring deflection without lying to yourself.

Step 1: Mine the repeated questions from your queue

The goal of this step is a ranked list of underlying questions, not ticket subjects. “Where’s my order,” “shipping status??”, and “has this even been sent” are one question wearing three outfits.

The lightweight version, doable in an afternoon:

  1. Pull 4–8 weeks of tickets. Long enough to smooth out a one-off incident spike, short enough to reflect current product and policy.
  2. Skim and label each ticket with the question behind it, in your own words: “refund window,” “password reset fails,” “invoice for accounting,” “cancel subscription.” Ignore how the customer phrased it; label the intent.
  3. Merge labels that are the same question. Be aggressive: “change billing email” and “update invoice address” probably belong together as “billing details change.”
  4. Count. Sort descending. Done.

If your helpdesk supports tags, tag as you go for two weeks instead of doing a retro export; the counts are equally useful. Teams already running ticket summarization have an easier time here, because skimming six-line summaries goes several times faster than skimming raw threads.

What most teams find: a sharp head and a long tail. The top 10 questions commonly account for a third or more of the queue; the tail is hundreds of questions that appear once or twice. The tail is not FAQ material. Resist it.

Step 2: Pick the top 10 — and only the top 10

Ten is not arbitrary. A ten-item FAQ can be scanned in one screen, maintained by one person, and every item on it has demonstrated demand. Every item past ten costs scanability and maintenance while deflecting a shrinking sliver of volume.

Apply two filters to your ranked list before locking the ten:

  • Answerable in one screen? “How do refunds work” is FAQ material. “My integration returns a 403 in one specific configuration” is a docs page or a ticket, not an FAQ item. Skip questions that need branching diagnostics.
  • Stable for at least a quarter? Questions about a temporary outage or a promo that ends next month will be stale before customers find them. Handle those in a banner or an email, not the FAQ.

A question that fails either filter yields its slot to number eleven on the list.

Step 3: Turn ticket answers into reusable articles

You are not writing from scratch. Your agents already answered each of these questions dozens of times, and among those replies are a few genuinely excellent ones. The drafting job is to convert “great reply to one customer” into “great answer for every customer.”

The conversion checklist:

  • Strip the personal: names, order numbers, apologies, thread references.
  • Keep the structure of the best reply: direct answer first, then the how-to steps, then the edge cases.
  • Phrase the question the way customers phrase it, not the way your org chart does. “Why was I charged twice?” outperforms “Duplicate transaction inquiries” both for scanning humans and for search.
  • Write the first sentence as the complete answer. Many readers stop there; make stopping there safe.
  • State specifics with dates of truth. “Refunds within 30 days of purchase,” and note internally when that was last verified.

This is a mechanical transformation, which makes it a good AI task. In Replydesk, the FAQ-draft workflow takes a pasted batch of real tickets on one topic plus your best agent answer and returns a clean question-and-answer article draft: customer phrasing on the question, direct-answer-first structure, placeholders where account-specific details used to be. An agent then does the part AI cannot: verifying every policy claim against the current policy, not against what the tickets said three months ago.

Budget-wise this is a light lift. Drafting all ten articles fits comfortably inside the free tier’s 20 drafts a day, so the whole first FAQ can cost you nothing but the review time. The same pass is also a good moment to fix register: FAQ answers should sit in the neutral-professional voice described in our customer support tone guide, warm enough to be human, plain enough to be quoted.

Copy-paste LLM prompts for FAQ generation

If you are running the drafting step through a general LLM instead of a packaged workflow, the prompt is most of the quality. These three cover the pipeline; paste them as-is and fill the brackets.

Prompt 1: cluster and rank the questions (Step 1 at scale):

Here is an export of [N] support tickets (subject plus first customer message).
Group them by the underlying customer question, ignoring phrasing differences.
Return a table: question (phrased the way a customer would ask it), ticket
count, example ticket IDs. Sort by count, descending. Do not invent questions
that are not present in the data.

TICKETS:
[paste export]

Prompt 2: draft one FAQ article from real material (Step 3):

Below are [N] real support tickets that all ask the same question, followed by
our best agent reply. Write an FAQ article from them.
- Title: the question, phrased the way customers phrase it.
- First sentence: the complete direct answer. A reader who stops there must
  not be misled.
- Then numbered steps if there is a how-to, then edge cases.
- Replace all names, order numbers, and account details with [placeholders].
- Do not add any policy detail that is not stated in the tickets or the reply.

TICKETS:
[paste tickets]

BEST AGENT REPLY:
[paste reply]

Prompt 3: staleness check on an existing article (Step 4, monthly):

Here is a published FAQ article and three recent agent replies on the same
topic. List every factual difference between the article and the replies
(amounts, timeframes, steps, policy details). Quote both versions of each
difference. If there are no differences, say so.

ARTICLE:
[paste article]

RECENT REPLIES:
[paste replies]

The guardrail in prompts 1 and 2 (“do not invent / do not add”) is not decoration. Without it, models pad FAQ drafts with plausible-sounding policy that nobody approved, and that is exactly the failure mode the next section exists to prevent.

Step 4: Keep the FAQ in sync with policy

Here is the failure mode that turns an FAQ from an asset into a liability: policy changes, the FAQ does not, and now your own website contradicts your agents. Customers quote the stale answer back (“your FAQ says 60 days”) and every such ticket takes longer than a ticket with no FAQ at all.

Sync is a process problem, so solve it with process:

  • Give every FAQ item an owner and a last-verified date. Ten items, one owner each (or one owner for all ten on a small team). The date is visible internally, even if not on the page.
  • Add “update FAQ” to the checklist of any policy, pricing, or feature change. The person changing the refund window is the person best positioned to know the FAQ just went stale.
  • Do a monthly 30-minute review. Read all ten items against current reality. On a ten-item FAQ this is genuinely half an hour, and prompt 3 above automates the diffing.
  • Watch for the contradiction signal. Any ticket where a customer cites the FAQ against you goes to the FAQ owner the same day.

Step 5: Measure deflection honestly

Vendor marketing routinely promises FAQ or knowledge-base deflection of 30 percent and up. Across real teams, the honest picture is humbler: meaningful deflection on the exact topics the FAQ covers well, roughly nothing everywhere else, and an overall effect that most teams would describe in single digits to low teens, when they measure it at all.

Measure it like this:

  1. Baseline per topic, before launch. From your Step 1 counts: “refund window” was, say, 41 tickets/month.
  2. Compare the same topic after launch, over at least a month, adjusting for obvious volume shifts (seasonality, a launch, an incident).
  3. Count assisted deflection separately. Agents linking an FAQ answer instead of writing one is a real win (faster replies, consistent answers) even though the ticket still exists. Watch your time-to-resolution on covered topics; it usually moves before ticket counts do. That reply-speed effect compounds with the tactics in reducing first response time.
  4. Accept that some questions will not deflect. Anything requiring account access (“where is my refund”) gets asked regardless of documentation quality. The FAQ’s job there is to set expectations, not prevent contact.

A sober scorecard after three months looks like: two or three topics down noticeably, a few flat, resolution time improved on covered topics, and a handful of stale-answer incidents caught by the monthly review. That is a successful FAQ. If someone reports 40 percent overall deflection, check whether ticket volume actually dropped or whether a form redesign just made contacting support harder; those are very different outcomes wearing the same metric.

The rhythm after launch

Once the first ten items are live, the ongoing loop is small: monthly, re-rank questions from fresh tickets, promote any newcomer that out-volumes a current item, retire what stopped appearing, verify dates-of-truth. An hour a month keeps the page honest.

The queue keeps writing the questions. Your job is just to keep listening to it.