A knowledge base is an operational resource, not an editorial project. Its job is not to describe your product, but to spare your team from answering the same questions for the hundredth time. That leaves two key questions: where do you find topics that address real needs, and how do you tell whether any of it made a difference?
Your topics are already there—in customer conversations
The most common mistake is holding a brainstorming session where the team imagines what customers might not understand. The result usually looks more like product documentation than a list of real problems. Better source material is already sitting in your inbox and conversation history.
- Export questions from the past three months.Pull them from chat conversations, tickets, and call notes. Focus on the customer's opening sentence.
- Reduce them to intent.‘How long until my parcel arrives?’, ‘When will it be delivered?’, and ‘How long does shipping take?’ are all the same question.
- Count how often each one appears.A list sorted by frequency is your publishing order. Do not guess at priorities.
- Separate sales questions from support questions.The first group affects purchasing decisions; the second affects support workload. Each matters in a different way.
- Record unanswered questions too.The ones that prompt your team to say, ‘I'll check and get back to you,’ are the most valuable—they reveal gaps within the company itself.
Your ten most common questions from the last quarter are a ready-made one-month editorial calendar. See how Meetlead brings conversations and tickets together in one place.
The same list is useful in several places at once: on your offer page, in your FAQ, and as source material if you plan to use an assistant that answers questions. We explain what this kind of knowledge needs in our article on AI in customer service.
The same list helps your team in another way: recurring questions are natural candidates for response templates. We explain how to build them in our article on customer service response templates.
Unanswered questions are useful when improving an assistant, too. Learn how to find gaps in an AI knowledge base and turn them into a queue of specific content improvements.
FAQ or knowledge base—they are not the same thing
Separating the two will save you a lot of work. An FAQ should resolve a doubt at the point of decision, in a few sentences and on the same page. A knowledge base is for someone already looking for specific information and willing to read a longer explanation.
Scroll the table sideways to see all columns.
| Characteristic | FAQ | Knowledge base |
|---|---|---|
| When it is used | During a decision, alongside the offer | After purchase or when solving a specific problem |
| Length | Two or three sentences | A complete step-by-step guide |
| Location | On a product or pricing page | A separate, searchable section |
| Typical question | ‘Can I return it?’ | ‘How do I make a return, step by step?’ |
| Measure of success | Less hesitation before purchase | Fewer tickets about the same issue |
If you have just one day to work on this, start with an FAQ beside your offer. An answer given at the moment of hesitation works immediately; a knowledge base article only starts working once someone finds it.
Before giving content to AI, check which documents should be sources at all. The guide to choosing your own data for a chatbot knowledge base separates public information, internal content and data that should be excluded.
How to write answers that resolve the issue
A good knowledge base answer looks different from good marketing copy. It should be concise, unambiguous, and free of cautious wording that will only make the customer send a message anyway. A concise, unambiguous answer may also be easier for AI features in Search to draw on, although nobody can guarantee it will be cited; see our guide to AI search visibility.
- Start with the answer, not an introduction—the first sentence should settle the question.
- Give a concrete figure instead of a range: ‘within three business days,’ not ‘usually quickly.’
- Say when something is impossible, too. An unanswered question generates a support request.
- Use the customer's language, not internal terminology. Nobody searches for a ‘retention module.’
- Cover one subject per article—combining three issues makes each one harder to find.
- End with a clear next step, even when that step is contacting a person.
Give the text to someone outside the team and ask what they would do next. If they hesitate for more than a moment, the answer does not resolve the issue.
The final point on the list matters more than it might seem. A knowledge base is not meant to cut customers off from human contact—it should eliminate questions with straightforward answers. A visible path to a person when the issue is more difficult increases trust in the entire section.
If these answers will also power an assistant, learn how to prepare a chatbot knowledge base: separate conditions, exceptions and conflicting versions before publishing.
When customers cannot identify a guide by its title, use the guide: Help article titles: use the words your customers search for.
How to tell whether it works
This is the part most guides omit—and the reason knowledge bases grow for years without any evidence that they help. Measurement does not need to be complicated. One number tracked over time is enough.
Scroll the table sideways to see all columns.
| Metric | What it shows | How to interpret it |
|---|---|---|
| Number of questions about a topic | Whether the article reduced your team's workload | The most important figure—compare it month over month |
| Share of recurring questions | How much support work is repetitive | A decrease represents real time saved |
| Article visits | Whether anyone is finding the content | High traffic with no change in question volume means the answer is not sufficient |
| Questions asked after reading | Whether the article resolves the issue | The best place to identify them is in conversations immediately after a visit to the help page |
The third row describes the most common trap. Growing article traffic is often celebrated as a success, even though an unchanged number of tickets means exactly the opposite: people read the article and still have to ask. We explain how to measure support workload honestly in our article on first response time in live chat.
If the number of questions about a topic has not fallen after a month, the article is not working—regardless of its view count. Schedule an implementation session, and we will help you set up the measurement.
To measure this systematically, you need consistent topic categories. We explain how to design categories that will still work a year from now in our article on tagging conversations and tickets.
Maintenance: how to keep it from decaying
An outdated knowledge base is worse than having none at all: customers follow its instructions and encounter a different reality. Maintenance does not require a full-time role, however—just a few rules and a single owner.
- One ownerThe person who decides what is added and what is removed. Without an owner, content gets duplicated.
- An update triggerAny change to a price, deadline, or process should always prompt a review of the related articles.
- A review dateAny article that has not been reviewed for a year goes into the queue, even if nobody has reported a problem.
- Remove content without sentimentalityAn article that nobody reads and that does not reduce question volume is a cost.
If your knowledge base also supplies an assistant that answers customer questions, keeping it current stops being a matter of tidiness and becomes a matter of accuracy. An outdated page quoted with complete confidence is more costly than no answer at all—we discuss this further in our article on AI in customer service.
One final practical point: when you publish an article, link to it in your team's standard responses straight away. A knowledge base that your own support team does not use has no chance of working for customers.
An article your own team does not use will not work for customers. See how Meetlead keeps conversations and tickets together in one view.
International traffic raises another question: which content should you translate, and which content should you not entrust to a machine? We answer it in our article on multilingual customer service.
If an AI assistant uses this content, editing a document and activating new knowledge are separate steps. Plan a chatbot knowledge update with a change owner and checks on the old version.
