What Is a Knowledge Base? Types, Examples and How to Build One

Some people meet the word "knowledge base" on a pricing page and want a definition. Others have been asked to build one and want to know what goes in it.

This guide covers both: the definition, where the term came from, how a knowledge base works, the five types, what goes in one, how it differs from an FAQ page, a wiki or a database, six public examples, a seven-step build plan, how to measure it, and what changes when an AI agent answers from it. Every source linked below was opened on 12 November 2026.

What is a knowledge base?

A knowledge base is a collection of written answers, organised by topic and searchable, that lets someone solve a problem or finish a task without asking another person. It can be public, for customers, or private, for staff, and the same collection can also be read by software such as an AI chatbot.

What turns a pile of documents into a knowledge base is findability and upkeep. Each piece answers something specific, has a title a reader would search for, and has someone responsible for keeping it true. A shared drive of old PDFs holds knowledge, but nobody can find anything in it.

Is it "knowledge base" or "knowledgebase"?

"Knowledge base" and "knowledgebase" mean exactly the same thing. The Oxford English Dictionary lists the term as two words; the one-word form is a common variant in product and site names, such as the University of Wisconsin–Madison "KnowledgeBase". Both are shortened to KB, which is why support articles are called KB articles.

Where did the term come from?

The Oxford English Dictionary's entry dates its earliest evidence for "knowledge base" to 1953, in the journal Accounting Review. Its technical sense came from artificial intelligence research, and the OED's earliest evidence for "knowledge engineering" is from 1977. In the expert systems of that era, a knowledge base was the store of facts and rules a program reasoned with, kept apart from the inference engine that applied them. The 1984 book on Stanford's MYCIN medical expert system devotes a whole part to "Building a Knowledge Base".

Customer support borrowed the word for its articles, and the abbreviation stuck. Microsoft's Windows 11 release information page still lists every update with a "KB article" number: the 10 November 2026 update for Windows 11 version 26H2 is KB5132061.

How does a knowledge base work?

A knowledge base works as a loop: someone asks a question, search matches it to an article, the article answers it, and every miss is recorded so a person can write or fix the answer. Here is one question passing through an imaginary online shoe shop.

1. A question arrives. At 11 p.m. a customer types "swap trainers bigger size" into the help center search. It could equally come from a support agent answering an email, or from an AI agent in a chat.
2. Search finds candidates. Keyword search matches words in titles and text; semantic search matches meaning, so "swap" can find "exchange". The article "Can I exchange shoes for a different size?" comes up first.
3. The article answers in its first line. "Yes. You can exchange unworn shoes for another size within 30 days, and UK exchanges are free." The steps follow, with the exact button names.
4. The result is recorded. The customer taps "Was this helpful? Yes". Had the search found nothing, the query would be logged as a search with no results.
5. A miss goes to a person. If the article did not help, the customer contacts support, and the agent replies and fixes the article while the context is fresh.
6. Someone reviews the gaps. Each month the owner reads the no-result searches; if "swap" keeps failing, the word goes into the article.

An AI agent runs the same loop faster: it retrieves the passages closest in meaning to the question and writes a reply from them instead of handing over a link. PepoChat's agent works this way, and when nothing relevant is found it says so and offers to bring in a person, which is step 5 in automated form.

What are the types of knowledge base?

There are five common types of knowledge base: an external help center for customers, an internal knowledge base for employees, an IT knowledge base, product documentation, and an AI knowledge base that a chatbot or AI agent reads.

An AI knowledge base is usually built from the others: the help center, a few policy PDFs and some written answers, loaded into an AI agent. Most small businesses need two types at most, a customer help center and a short internal one, and the first doubles as the AI agent's source.

What goes in a knowledge base?

A knowledge base holds five kinds of article: how-to guides, troubleshooting articles, FAQs, reference pages and policies, each answering one question or one task.

• A how-to article walks through a task in numbered steps, such as "How do I add a teammate?".
• A troubleshooting article starts from a symptom in the customer's words, such as "Why didn't I get my reset email?", and gives the quickest fix first.
• An FAQ is a short question with a two- or three-sentence answer, for facts people want before they buy.
• A reference article is a table or list people look up, such as shipping times by country.
• A policy article states a rule about money, time or eligibility, with the numbers, the exceptions and the date it took effect.

A fixed outline per type saves writers from a blank page and shows readers where to look. A troubleshooting outline:

Title: [Symptom in the customer's words, as a question]
First sentence: [The fix that solves most cases]
Applies to: [Product, plan, version or region]

Quick fix
1. [One action per step, with the exact button names]
2. [How to tell it worked]

If that didn't work
- [Second most likely cause]: [its fix]
- [Third most likely cause]: [its fix]

Still stuck?
[What to send support: order number, screenshot of the error, device]

Related: [Link to the matching how-to article]
Last reviewed: [Date] by [Owner]

To decide which articles you need, start from the questions customers already ask; the knowledge base starter kit lists 50 that most small business sites must answer.

How is a knowledge base different from an FAQ page, a wiki or a database?

A knowledge base is a maintained set of answers organised for finding; an FAQ page is one page of short answers, a wiki is a set of pages anyone on a team can edit, and a database stores structured records rather than explanations.

Google introduced its knowledge graph on 16 May 2012 with the slogan "things, not strings"; its launch post said it held more than 500 million objects and 3.5 billion facts. A knowledge graph stores facts as connections, such as "this person founded that company", while a knowledge base stores answers people read.

A help desk tracks each request; a knowledge base holds the answers that stop it coming back. An FAQ page is often the first step towards one.

What are examples of good knowledge bases?

Six public knowledge bases worth studying are Mozilla Support, Stripe Support, the Shopify Help Center, the University of Wisconsin–Madison KnowledgeBase, the GitLab Handbook and Microsoft's KB articles. Each does one thing especially well.

Mozilla Support: answers that match the reader's setup

Mozilla's Firefox articles, such as Refresh Firefox, let readers pick their Firefox version and operating system so the steps match their screen, and show a last-updated date and the share of readers who found the article helpful. Mozilla says volunteers power its support.

Stripe Support: titles in the customer's words

Stripe Support opens on popular articles titled like the customer's own question, such as "Trouble signing in?". Some show the share of users the article helped, and most are tagged with more than one topic, so they can be found from several directions.

Shopify Help Center: organised by what the merchant is doing

The Shopify Help Center groups articles under "Browse by topic" into tasks such as Orders and shipping, Payments and Your account, with three starting articles each. A "Chat with a human" link sits directly under the "How can we help?" heading, so self-service is never a dead end.

University of Wisconsin–Madison KnowledgeBase: one platform, many owners

The UW–Madison KnowledgeBase is a shared platform where more than 350 campus partner groups each manage their own section and can share articles with the others. It is a good model for an IT knowledge base: central search, local ownership.

GitLab Handbook: an internal knowledge base, published openly

The GitLab Handbook is the company's internal knowledge base, open to anyone. It started when GitLab was a company of just ten people, so later hires could read decisions made before they joined, and anyone can propose a change through a merge request.

Microsoft KB articles: a number people search by

Microsoft publishes a numbered KB article for each Windows update, and IT teams know updates by those numbers. A stable ID lets forums and patch tools refer to the same article for years.

Why does a business need a knowledge base?

A business needs a knowledge base because customers try to help themselves first, and when the answer is not written down they contact you or give up.

Gartner's survey of 5,728 customers, conducted in December 2023 and published on 19 August 2024, found that 73% of customers use self-service at some point, but only 14% of issues were fully resolved there. Even issues customers called "very simple" were fully resolved only 36% of the time. The most common reason self-service failed, in 43% of cases, was that customers could not find relevant content. The right articles under the right titles attack that directly.

Three other benefits matter as much. Every agent and every chat gives the same refund window instead of three versions of it. New hires read answers instead of interrupting colleagues, one of the advantages GitLab lists for its handbook. And an AI support agent can only be as good as the content it searches.

How do you build a knowledge base in 7 steps?

You build a knowledge base in seven steps: collect the top questions, pick a structure, write one answer per article, title each article as the customer's question, add search, publish, and review on a schedule.

1. Find the top questions in tickets and chat transcripts

Export a month of tickets, emails and chat transcripts, tag each with the reason in the customer's words, and rank the reasons. Fifteen to thirty articles covering the top reasons is a solid first version.

2. Pick a structure

Name a handful of categories after customer tasks, such as "Orders and delivery" and "Billing", not after internal departments.

3. Write one answer per article

One question or task per article. If an article covers changing an address and cancelling an order, split it and link the two.

4. Title each article as the customer's question

"Can I return a sale item?" beats "Returns policy", because it is what people type. Answer it in the first sentence.

5. Add search

Make sure search covers article text, not only titles, and reports searches that returned nothing.

6. Publish where people look

Link the knowledge base from your site's header or footer, your contact page, your order emails and your chat widget.

7. Review on a schedule

Update an article the day its fact changes, review the most-read articles monthly, read everything once a quarter, and give each category a named owner.

How do you write a knowledge base article an AI can also use?

Write every section so it makes sense on its own: put the answer in the first sentence, name the product, plan and region in each section, keep facts in text rather than images, and keep each fact on one page.

AI agents rarely read a whole article. Most split documents into short passages and retrieve the few that best match the question, so a section that says "as mentioned above", or shows the price only in a screenshot, loses its meaning once cut loose. Repeat the subject ("Starter plan refunds", not "this plan"), date anything that changes ("From 1 November 2026, returns are accepted within 14 days"), and link to one page for each fact instead of copying it. The guide to formatting docs for an AI chatbot has twelve rules with before-and-after examples.

How do you measure whether a knowledge base works?

Measure a knowledge base by whether people find answers: track searches with no results, article views against tickets on the same topic, and article ratings, and read all three every month.

For the second row, divide views by tickets. If the returns articles had 2,000 views last month and you received 100 returns tickets, that is 20 views per ticket; watch the trend after you rewrite the articles, not the number itself.

PepoChat's Help Center, for example, has no article ratings or no-result search report, so in PepoChat the gaps show up as escalated conversations and Analytics topics.

What is Knowledge-Centered Service (KCS)?

KCS, now called Knowledge-Centered Success and formerly Knowledge-Centered Service, is a method for building a knowledge base as part of solving problems, rather than as a separate writing project, developed and maintained by the Consortium for Service Innovation, a non-profit association.

According to the Consortium's KCS Practices Guide, the people who answer requests reuse, improve and, when nothing exists, create knowledge in their normal workflow, and take ownership of it. The method runs as two loops: the Solve Loop of daily work (reuse, improve, capture, structure) and the Evolve Loop that learns from it (content health, process integration, performance insight, change management). The Consortium kept the acronym when it changed the name, and says the guide draws on more than thirty years of member experience.

The most useful KCS idea for a small team is "Reuse is Review": check an article each time you use it and fix it then.

What kinds of knowledge base software are there?

Knowledge base software falls into six categories: help centers built into help desks, standalone knowledge base tools, internal wikis, documentation platforms, IT service management tools, and AI support agents with a built-in knowledge base.

Each maker's site was opened on 12 November 2026; this is a map, not a ranking. PepoChat is our product; we include it where it fits and say where other tools are the better choice. For a large public help site that ranks in Google, versioned developer documentation or an internal wiki with page permissions, one of the dedicated tools is the better fit.

What changes when an AI agent answers from your knowledge base?

With AI, the knowledge base becomes the source an AI agent answers from, so every gap and every stale page turns into a missing or wrong answer in a live conversation.

Before AI, a bad article failed quietly: the customer gave up and emailed you. Now the agent retrieves the passage and repeats it confidently. Two pages that disagree let it quote either one, and a page that says "contact us for pricing" makes it say the same. Gartner's 2024 release recommends three things that fit here: a single front door such as a generative AI chatbot, reps who create knowledge as they resolve issues, and a way for customers and reps to flag content that does not work.

The upside is speed. A wrong AI answer is usually a content problem you can fix in minutes by editing the source page or adding a question-and-answer pair. The knowledge base chatbot explainer walks through retrieval step by step, and the monthly transcript routine turns missed questions into articles.

What are the most common knowledge base mistakes?

The most common knowledge base mistake is writing what the company wants to say instead of what customers ask. The rest follow from it:

• Internal titles. "Fulfilment SLA" instead of "When will my order arrive?".
• One article for everything. A long "Billing overview" that answers ten questions badly.
• Answers only in screenshots or videos. Search and AI tools read text.
• No owner. Articles go stale the week a price changes.
• No way out. No visible contact route turns a missing article into an angry customer.
• More instead of better. Fifty accurate articles beat five hundred half-true ones, especially once an AI agent searches them.

How PepoChat uses a knowledge base

In PepoChat, the knowledge base is what the AI support agent answers from. Before each reply it searches your workspace's private knowledge base for the passages that match the question and writes from them; when nothing relevant comes back, or the visitor asks for a person, it says so and hands the conversation to the team inbox.

You fill it from several kinds of knowledge source: a website crawl of up to 200 pages, 3 levels deep, or a sitemap import; files (PDF including scanned pages, DOCX, Markdown, HTML, TXT, CSV, JSON and images); text snippets and Q&A pairs; and Notion and Google Drive, which sync daily, as the Notion and Google Drive guide shows. A whole crawl, sitemap or connected app counts as one source. PepoChat also has a built-in Help Center: articles you publish appear in the widget's Help tab and are copied into the same knowledge base, so one edit updates what customers read and what the agent says.

The limits: Free allows 5 knowledge sources, 20 MB of text and 50 Help Center articles; Starter, Growth and Scale allow unlimited sources with 200 MB, 1 GB and 5 GB, and 300 articles. Paid plans re-fetch web pages weekly; on Free you sync by hand. The Help Center lives inside the chat widget rather than as a standalone help site, its search matches titles, and it has no article ratings. Create a free workspace to see what the agent says from your own pages.

What to do next

• Write your first 15 to 30 articles with the four templates in the self-service help center guide.
• See how Help Center articles are published and mirrored into the agent's knowledge base.
• Compare knowledge base limits by plan on the pricing page.
• Start a free PepoChat workspace, crawl your site and ask your top ten customer questions: every chat feature is on the free plan, with 100 AI replies a month and no card required.
• Related explainers: What is a help desk?, What is RAG? and What is a chatbot?.