Home / Guides / Escalations and support operations

Knowledge base gap analysis: find the questions your help center doesn't answer

By Draft First. Updated October 2026

Every repeat ticket is a question your help center didn't answer. Sometimes there's no article. Sometimes there is one, but it skips the step people get stuck on, or it uses words nobody searches for. A knowledge base gap analysis finds those questions, counts them and tells your article writer where to start.

This guide covers the manual method, an AI prompt that does the counting, and a check to run before anyone writes a word.

What you need

• A ticket export for the last 30 days. Include the subject and the customer's first message. That first message is where the real question lives.
• A list of your help articles. Titles and links are enough to start. An export or your help center sitemap works.
• An approved place to do the work. Tickets hold customer data. If your company hasn't approved an AI tool for it, take names and emails out first and spot-check the export yourself.

The four kinds of gap

Not every gap needs a new article. Sort each question into one of these, because each one has a different fix.

Step by step

1. Pull the export. Subject, first customer message, ticket ID, date. Leave out internal notes.
2. Group by question, in the customer's words. "How do I rotate an API key?" is a question. "API" is a category. Merge tickets that ask the same thing in different words, and keep the ticket IDs behind each group.
3. Count each group. Count each ticket once, even if it appears twice in the export.
4. Check each question against the article list. Search with the customer's words, not yours. Mark it covered only if an article clearly answers it.
5. Rank the gaps. Start with how often the question comes up. Then weigh how long those tickets take and how easy the fix is. A ten-minute title change can beat a new article.
6. Outline the top two. A title and four to six headings is enough for your article writer to start.
7. Check again next month. Once an article is live, see whether that question got less common. That's the only proof the work paid off.

An AI prompt that does the counting

Counting and matching is the slow part, and it's the part an AI agent is good at. Paste this with your export and article list. Swap the [brackets] for your own details.

Using the ticket export and article list below:
1. Group the tickets by the customer's actual question. Merge tickets
   that ask the same thing in different words. Keep the ticket IDs.
2. For each question, say whether an article clearly answers it,
   partly answers it, or doesn't. Name the article if there is one.
3. List the uncovered questions, most asked first, with the count
   and up to 3 example ticket IDs.
4. For "partly covered", say what the article is missing.
Start with one line: the date range, how many tickets and how many
articles you read. Use only what's in the export. Don't quote
customer names or details. Outlines only, don't write articles.

Tickets: [paste or attach the export]
Articles: [paste the list of titles and links]

What a good result looks like

Short, counted and checkable. This sample uses made-up data.

Sep 16 to Oct 15 | 388 tickets read | 142 articles read

No article
1. "How do I rotate an API key?" 17 tickets (#4831, #4795, #4702)
2. "Why does SSO send me back to the login page?" 12 tickets
   (#4820, #4823, #4719)

Partly covered
- "Export times out": "Exporting reports" doesn't mention the
  row limit.

Check before anyone writes

Agents miss articles with vague titles and sometimes merge questions that aren't the same. Run this in the same chat before you hand the list over.

For the top 5 questions you marked as having no article, search the
article list again using other words a customer might use. Show any
article that might answer it and why. Then list any group where the
merged tickets are really asking different things.

Mistakes that waste the effort

• Counting categories. "Login: 60 tickets" tells a writer nothing. The question inside the ticket does.
• Trusting titles. An article called "Getting started" might answer five of your gaps. Read before you write.
• Writing what nobody can find. Use the customer's words in the title and first line.
• Never measuring. If you don't recount next month, you won't know which articles worked.

The same ticket export feeds a voice of the customer report, and the ticket summary prompts help when one theme needs a closer look.

Questions

How often should we run a knowledge base gap analysis?

Once a month suits a steady queue. Run an extra one after a big release or a pricing change, because that's when new questions show up.

How many tickets do I need?

Enough for repeats to show. A month of tickets is a sensible start. If your volume is low, use a quarter instead.

Can AI write the missing articles?

It can write a first draft from an outline. A person who knows the product should check every step and setting, then publish it.

Is a ticket category report enough?

Usually not. Categories like Billing or Login hide the real question. Use the subject and the customer's first message.