How to Build a Self-Service Help Center Customers Actually Use

Most help centers are built the wrong way round. Someone exports the product manual, splits it into pages, files them under the names of internal teams and calls it self-service. A customer searches for "refund", finds an article called "Billing overview", and emails you anyway.

This guide is for small teams building a help center for the first time, or rebuilding one nobody uses. It covers what self-service can and cannot do, how to choose the first articles, templates for four article types, writing rules, navigation and search, an AI agent on top, what to measure, and how often to update. Where PepoChat's built-in Help Center has a feature or a limit, the guide says so, including what it does not do.

Two definitions first. Customer self-service is any way a customer can resolve an issue without contacting a person: a help article, an account page, an order lookup or an AI support agent. A help center is the public, searchable collection of articles at the heart of it, grouped by topic; combined with account tools, it is often called a self-service portal. Some teams call theirs a knowledge base; in this guide, that term means everything an AI agent can search, which usually includes the help center and more.

How much can self-service actually resolve?

Less than most teams expect. In a Gartner survey of 5,728 customers, published in August 2024, only 14% of customer service issues were fully resolved in self-service, and only 36% of the issues customers described as "very simple".

Customers do try it first. The same Gartner press release says 73% of customers use self-service at some point in their journey, and an earlier Harvard Business Review article, Kick-Ass Customer Service (January 2017), reported that 81% of all customers attempt to take care of matters themselves before reaching out to a live representative. The gap between trying and succeeding is mostly a findability problem: in 43% of cases, the most common reason for failure, customers could not find content relevant to their issue. The guide to reducing support tickets puts these numbers in the context of a wider ticket-reduction plan.

Gartner's recommendations from the same release shape the rest of this guide. It suggests letting support reps create knowledge as part of resolving issues, putting a single digital concierge, such as a GenAI chatbot, at the most prominent entry point, and letting both customers and reps flag content that does not work. In plain terms: write from real tickets, make answers easy to reach, keep improving what is used, and never make self-service the only door.

How do you decide which help articles to write first?

Write the articles behind your most frequent contact reasons first. Tag a few weeks of tickets and chats with the reason in the customer's words, rank the reasons by count, and write one article per reason from the top of the list down.

If you already run an AI support agent, its conversations are a ready-made list. In PepoChat, Analytics topics label each conversation from the visitor's first message and rank the most asked topics over the last six weeks, and the documentation suggests treating a topic that grows week over week as an article worth publishing.

Before you write, sort each reason into one of three kinds, because only one of them is solved by an article:

• The customer can finish alone with instructions. "How do I change my delivery address?", "How do I add a teammate?". Write a help article.
• The answer depends on the customer's own account. "Where is my order?", "When does my plan renew?". An article can explain where to look, but the real fix is an account page, a notification or an integration that looks up the record.
• It needs a judgment call. Refund exceptions, damaged goods, complaints. Write a policy article that explains the rule and how to ask, then route the request to a person.

A first version with 15 to 30 articles is enough for most small businesses, as long as each one maps to a real contact reason. Resist the urge to document every feature on day one. An article nobody needs still has to be maintained.

Which help article templates should you use?

Four templates cover almost every support article: how-to guides for tasks, troubleshooting articles for problems, policy articles for rules, and reference articles for facts customers look up. A fixed template per type makes articles faster to write and easier to scan.

Documentation teams have sorted content by purpose for years. The Diátaxis framework, used for many software documentation sites, identifies four forms: tutorials, how-to guides, technical reference and explanation. A support help center leans on how-to and reference, adds troubleshooting and policy, and rarely needs long tutorials or explanations.

A how-to article built from that template looks like this:

# How do I change my delivery address?

You can change the delivery address until your order ships, from Orders in your account.

You need: the email address you ordered with.

1. Sign in and open **Orders**.
2. Select the order, then **Change address**.
3. Enter the new address and select **Save**.

You will get a confirmation email within a few minutes. If the order has already shipped,
contact the carrier using the tracking link in your shipping email.

Last updated: 6 October 2026

Troubleshooting articles deserve the most care, because the reader is already frustrated. Put the fix that solves most cases first, such as "check your spam folder", even if it feels obvious, and end with the exact details support will need, such as the order number and a screenshot of the error, so the follow-up contact takes one message instead of three.

How should you write help articles?

Write each article for a customer in the middle of a problem: title it with their question, answer it in the first two sentences, then give numbered steps with one action each. Keep each article to one task and put the facts, numbers and dates in the text. The same answer-first rule makes a good FAQ page.

Nielsen Norman Group's tenth usability heuristic sums up the standard: help and documentation "should be easy to search, focused on the user's task, list concrete steps to be carried out, and not be too large". Its article on help and documentation adds that support pages should be easy to scan, with clear headings and numbered lists, and that text instructions should accompany any video, because not everyone can or wants to watch one. Eight rules put that into practice:

1. Title the article with the customer's question or task. "Can I get a refund after 30 days?" beats "Refund policy", because it is what people type into search.
2. Answer in the first two sentences. Many readers stop there, and so do search snippets and AI summaries.
3. Number the steps, one action per step. Use the exact labels from your interface in bold, such as Settings and Save, so readers can match them on screen.
4. One task per article. If an article covers changing an address and cancelling an order, split it and link the two.
5. Facts, not adjectives. "Refunds reach your card in 5 to 10 business days" instead of "quickly".
6. Use screenshots sparingly. Each one goes stale with the next interface change, and search and AI agents read the text, not the image.
7. Write the instructions instead of "contact support". If a person really is needed, say why and how to reach them.
8. Date every article. A visible "last updated" line tells readers, and your team, how much to trust it.

How should you organise help center navigation and search?

Organise the help center around what customers are trying to do, not around your internal teams. Put the most used articles on the front page, write titles that match what customers search for, and keep a route to a person visible from every article.

Name your categories after customer tasks: "Orders and delivery", "Returns and refunds", "Billing", "Your account", "Using the product". A handful of clear categories works better than a deep tree, and a category called "Miscellaneous" is a sign that a few articles need better homes. NN/g's documentation guidance makes the same three points: group help topics into relevant categories, optimise for search so the right content comes up fast, and highlight the content people view most.

Many visitors go straight to the search box, and search can only match what you wrote. The fix for "customers couldn't find it" is usually a title in the customer's words rather than a cleverer search engine. When a search returns nothing, the words the customer typed are your next article title, or a better title for an existing one.

In PepoChat, the Help Center lives inside the chat widget. You group articles into collections, up to 30 per workspace. The widget's Help tab lists the collections and lets visitors search article titles, which makes titles in the customer's words even more important. The home screen's help card shows a search box and your top four articles, and the Send us a message card is always on the home screen, so starting a conversation is never more than one tap away.

Should you put an AI agent on top of your help center?

Yes, if it answers only from your content and hands over to a person when it cannot. An AI agent lets customers ask in their own words instead of guessing your article titles, which tackles the most common self-service failure. A bad one costs trust quickly.

An AI support agent is software that answers customer questions in a chat window by searching your own content, and hands the conversation to a person when it cannot help. On top of a help center, it can do what title search cannot: treat "my parcel never came" as a question about lost deliveries, combine two articles into one answer, and ask a follow-up question when the first message is vague.

The risk is real, though. In a Gartner survey of 3,566 customers conducted in February and March 2026 and published in September 2026, only 27% said they would try a chatbot again after a negative experience. While 49% said they would have been willing to use a chatbot, only 7% had used one in their most recent service interaction. Gartner's advice was to "prioritize reliability over reach", and it found that 87% of customers say access to a human agent is essential when companies use GenAI in customer service.

Three conditions make an agent worth adding:

• It answers only from your content. An agent that searches your help center on every reply and says when the answer is not there will not invent a refund policy. The guide to stopping chatbot hallucinations lists what to check.
• It hands over cleanly. When it cannot help, or the customer asks for a person, the conversation should reach your team with the transcript attached. The human handoff guide covers how escalation should work.
• Its scope is stated. A greeting that says what it can help with steers customers to questions it answers well.

In PepoChat, every published Help Center article is mirrored into the agent's knowledge base automatically and removed when you unpublish it, so you edit one place and the answer shows up in two: the article and the chat. Drafts are private and the agent ignores them. The agent searches the knowledge base on every reply and escalates to the team inbox when it finds nothing relevant. If you turn the Help tab off, visitors no longer see the articles, but the agent still answers from the published ones.

How do you measure whether your help center works?

Measure whether customers find answers, not how many articles you have. Track searches that return nothing, article feedback and the contact rate on the topics your articles cover, and check them against your top contact reasons every month.

Contact rate is the number of support contacts per 100 orders or per 100 active customers in a period. It matters more than raw ticket counts, because counts rise whenever the business grows. If articles on a topic work, contacts on that topic should fall as a share of orders even while total volume rises.

Be clear about what your tool measures. PepoChat's Help Center documentation does not describe article votes or a search log. What PepoChat's Analytics does show is the share of positive and negative visitor votes on resolved conversations over 90 days, and the most asked topics over six weeks. It has no charts of conversation volume or resolution rate, so compute contact rate from your inbox counts or the conversation CSV export. The guide to AI chatbot metrics explains which numbers mislead.

How often should you update help center articles?

Update an article the day the product, price or policy it describes changes, review the most used articles every month, and audit the whole help center every quarter. Articles customers use every day need the most attention; articles nobody opens can wait.

A cadence that works for a small team:

• On change, the same day. Keep a list of the facts your articles depend on, such as prices, delivery times, plan limits and interface labels. When one changes, update every article that mentions it before the change goes live.
• Weekly, 30 minutes. Read the conversations your AI agent escalated and the tickets answered by hand. A missing answer becomes a new article; a wrong one gets fixed at the source.
• Monthly. Review your ten most used articles, the searches that returned nothing and any negative feedback.
• Quarterly. Read every article once. Delete what nobody needs, merge duplicates, and replace stale screenshots.

KCS, the knowledge management methodology maintained by the Consortium for Service Innovation, calls the core of this Reuse is Review: review knowledge in the process of using it, while the context exists to improve it, and spend no time reviewing articles that are never used. In practice, whoever answers a ticket with an article fixes the article if it was wrong. Give each category a named owner so that "someone should update this" becomes a person's job.

One PepoChat-specific warning: deleting an article is immediate and has no confirmation. If you are unsure, unpublish it instead. That removes it from the widget and the agent's knowledge base but keeps the article as a draft.

What does PepoChat's Help Center include?

PepoChat's Help Center is a small help center inside the chat widget. You write Markdown articles in collections, publish them to the widget's Help tab, and every published article is also available to the AI agent. It is on every plan, including Free.

The documentation lists these details:

• Articles have a title of up to 200 characters, an optional short description of up to 500 characters shown under the title, a collection, and a Markdown body of up to 100,000 characters.
• Import brings existing help content in as drafts: one help page by URL, up to 20 pages from a sitemap, or the first 20 entries of your existing knowledge base. Pages with less than 200 characters of text are skipped. Importing the same URL again updates the article in place and keeps its publish state, and imported articles give visitors a View original page link.
• Limits are 50 articles on Free and 300 on Starter and Growth, counting drafts and published articles, as the billing page shows. Articles do not count toward the knowledge-source limit; their text counts toward the plan's knowledge-base text cap.

The limits are worth stating plainly too. Help tab search matches article titles, so titles carry the weight of findability. The documentation describes the Help Center inside the widget and does not describe a separate public help site, so if you want your articles to rank in Google, keep them as pages on your own website as well and use Import to bring them in, since a re-import updates the article in place.

What to do next

Start with the list, not the tool. Tag the last few weeks of tickets and chats, rank the reasons, and write the ten articles behind the top ten using the templates above. Then add an AI agent on top with a clear route to a person, and run the weekly review.

To build it in one place, create a free PepoChat workspace. The Help Center, the AI agent, Analytics topics and the team inbox are all on the free plan, with 50 articles, 200 AI replies a month and no credit card.