Chatbot Knowledge Base Starter Kit: 50 Questions Your Site Must Answer

An AI support agent can only answer from what you have written down, and customers look for written answers before they contact anyone. In the 2017 Harvard Business Review article Kick-Ass Customer Service, the authors reported that 81% of customers try to resolve an issue themselves before reaching out to a live representative. When the answer is missing, that attempt fails: in a Gartner survey of 5,728 customers published in August 2024, customers could not find content relevant to their issue in 43% of failed self-service attempts, as our AI customer service statistics page records.

The checklist is written for whoever sets up a support chatbot on a small business site: an online store, a SaaS product, a service business such as a clinic, salon or trade, or an agency. A knowledge base is the private collection of pages, documents and written answers that an AI support agent searches before every reply. Below are the 50 questions it should answer, where each answer usually lives, what the bot should route elsewhere, and how to load and maintain the answers.

The list is an editorial checklist built from common contact reasons and the research linked below, not a measured ranking of how often each question is asked. The product steps use PepoChat, the AI support agent we build. PepoChat's limits and every linked source were checked on 19 October 2026; the two Gartner figures were checked on 16 October 2026 for our statistics page. Store questions that need order data have their own guide, ecommerce chatbot use cases, and the public page itself is covered in how to write an FAQ page.

What should a chatbot knowledge base include?

A chatbot knowledge base should include a current, written answer to every question customers ask before and after they buy, phrased the way they ask it, with the numbers in the text. The 50 questions fall into eight groups, and most of the answers already exist on a handful of pages.

Treat the list as a starting point, not a script. Nielsen Norman Group's Susan Farrell wrote that FAQ questions "should echo the real concerns people contact you about", in your visitors' own vocabulary (FAQs Still Deliver Great Value, December 2014). Keep the questions your customers ask, reword them the way they say them, and delete the rest.

In the group tables, "Store" means ecommerce, "SaaS" means software sold by subscription, "Services" covers appointment businesses such as clinics, salons and trades, and "Agency" covers studios and consultancies that sell projects. Notes for course creators appear where access and payment plans differ.

Which questions about your business should the chatbot answer?

The first seven questions cover who you are, what you offer, where you work and when you are open. They are cheap to fill, since the content usually sits on your home, About and contact pages.

Visitors look for this context deliberately. Nielsen Norman Group's research on About pages, which observed more than 70 users, found that people want to learn an organisation's purpose, location, history and how to contact it, and worry about whether a company will ship orders, accept returns or sell personal data (About Us information on websites, May 2019).

Two questions in this group are easy to get wrong. Question 2 needs a written "no" as much as a "yes": if your services page lists what you do but never what you do not, the agent can only report that it found nothing. Question 6 needs a time zone and the holiday exceptions, because "Monday to Friday, 9 to 5" is ambiguous for every visitor abroad and wrong on every public holiday.

What pricing and payment questions do customers ask?

Customers ask what it costs, what costs extra, how they can pay and how they get a receipt. These eight answers decide purchases, so write them with amounts, currencies, what is included and the conditions, never with "competitive pricing".

Extra costs are where precise answers matter most. In Baymard Institute's 2026 survey of 1,083 US adults who had shopped online in the previous three months, the most common reason for abandoning a checkout, once shoppers who were just browsing are set aside, was extra costs such as shipping, tax and fees being too high, at 40%; 12% could not see or calculate the total cost up front, and 9% found too few payment methods (Baymard cart abandonment statistics). An agent that can state the full price before checkout answers the question behind many abandoned carts.

Service businesses and agencies often keep prices off the site, which leaves the chatbot nothing true to say. Nielsen Norman Group's B2B research found that business customers rank pricing as the information they most need online, and watched participants leave for competitors' sites when prices were missing (State the Price, December 2013). If a fixed price is impossible, publish typical prices for common jobs or a range, as the same article suggests.

What do customers ask about ordering and booking?

Ordering and booking questions are about the mechanics: how to buy or book, whether something is available, how to change it and what happens next. Most of the answers belong in short how-to articles with numbered steps, in the order the customer does them.

Question 17 is the exception to writing everything down. Stock and free slots change by the minute and a crawled page is a snapshot, so the knowledge base should explain how availability works rather than state it. In PepoChat, booking runs through a connected PepoSmart calendar, and the agent offers only the slots the calendar returns.

Question 21 is easy to answer in advance: say who sends the confirmation, what its subject line says, how long it takes and to check the spam folder.

What do customers ask about delivery and access?

Delivery and access questions ask when, where and how the customer gets what they paid for: a parcel, a finished project, a login or a course. Each answer needs a number: processing days, transit days, countries, postcodes or the length of access.

Slow delivery was the second most common reason in Baymard's 2026 checkout survey, named by 20% of shoppers who abandoned a checkout. "Fast shipping" does not answer question 22; "orders placed before 2 pm ship the same day and arrive in 2 to 4 business days" does.

Question 24 splits in two. "Where is my order?" needs live order data that no page holds, and the Shopify and WooCommerce order status guide covers connecting a lookup. "How is my project going?" is better answered with the update schedule you promise, such as a written update every Friday.

What do customers ask about returns, refunds and cancellations?

Customers ask whether they can return or cancel, how, by when, what it costs and when the money comes back. These six answers need the conditions in full, because a vague returns answer can become a dispute later; Baymard found that 13% of shoppers who abandoned a checkout did so over an unsatisfactory returns policy.

If you sell to consumers in the EU, your policy cannot be less generous than the law. The EU's Your Europe guidance says that for most online purchases a consumer can cancel and return the order within 14 days without giving a reason, with exceptions such as personalised goods and opened sealed media, and that a minimum two-year guarantee always applies to faulty goods (Your Europe: shopping consumer rights). Write your window and your exceptions as plain statements the agent can quote.

For SaaS, question 31 matters most: where the cancel button is, whether the plan runs to the end of the paid period, and what happens to the data.

Which account and security questions should the knowledge base cover?

Account questions are how-to questions with a security edge: logging in, resetting a password, changing details, adding users, deleting an account and keeping card details safe. The knowledge base should explain how customers do each of these themselves, because an agent that answers from documents cannot reset a password or change an email address for them.

Two of these questions sit behind abandoned checkouts. Baymard found that 18% of shoppers who abandoned did so because the site wanted them to create an account, and 19% because they did not trust the site with their card details. If you offer guest checkout, say so; if a payment provider handles cards and you never see the full number, say that too.

Keep staff-only procedures out of this group. A page explaining how your team overrides a lock or transfers an account by hand could be quoted to anyone who asks.

What do customers ask about support and contact?

Support questions are about reaching a person: how, how fast and through which channel. The answers belong on the contact page, and they must match what your chat actually does, or the agent will promise a reply nobody sends.

In a Gartner survey of 3,566 customers run in February and March 2026, 87% said it is essential that companies using generative AI offer a way to reach a human agent (see our AI customer service statistics). In PepoChat, the Team availability card in Widget Customization tells the agent whether your team answers in the chat, only during business hours, or by email follow-up, and the agent adjusts what it promises; your contact page should say the same.

For question 44, the self-service help center guide covers which articles to write first.

Which policy questions does a support chatbot need to answer?

Policy questions ask what you do with data, what the terms say, whether you are licensed or insured, how you source and work, whether the site is accessible, and whether the chat is a bot. The answers usually sit in legal pages written for lawyers, so add a plain-language summary the agent can quote and link to the full text.

Be exact about credentials. List only the licences, insurance and certifications you hold, with the issuing body, because the agent repeats whatever the page says to anyone who asks. Question 50 also has a legal side: the EU AI Act's duty to tell people they are talking to an AI has applied since 2 August 2026, and the disclosure laws guide covers the US state rules. Answer it in one line in the knowledge base and again in the greeting.

Which questions should a chatbot not answer from the knowledge base?

Questions about one customer's own data, requests for a money decision, and requests for medical, legal or financial advice should not be answered from the knowledge base. The first kind needs a live lookup or a person; the others need a person every time, with the details collected first.

In PepoChat, refusals go in the Never do this field of the agent instructions, up to 2,000 characters, for example "Do not give legal or medical advice"; the agent declines those requests even when the knowledge base contains related text. Account lookups run through integration actions, which email verification does not gate: an action's "when to use" text can tell the agent to ask for verification first, an instruction rather than a lock. The verification guide explains when to require it.

How do you load the answers into an AI support agent?

Import your public pages with a website crawl or sitemap, upload the documents that hold the rest, and write anything that exists nowhere as a text snippet or Q&A pair. The table shows the PepoChat option for each kind of content and how it counts against the Free plan's five knowledge sources.

Limits read from PepoChat's documentation on 19 October 2026. Text from every kind of source counts toward the plan's knowledge-base text cap, which is 20 MB on Free.

A knowledge source is one item that counts toward a plan's source limit, such as a crawled website, an uploaded file or a connected app. A Q&A pair stores a question with its answer, so retrieval matches the way visitors phrase it; the question can run to 500 characters and the answer to 20,000. The text and Q&A docs recommend one fact per pair.

The crawler keeps each page's main content, honours robots.txt and reads the HTML the server sends without running scripts, so a page that builds its prices with JavaScript, or sits behind a login, may import empty; upload that content as a document instead. Website pages are re-fetched weekly, and Sync website pages refreshes them on demand. The website import docs and the training guide have the steps.

How do you find the gaps after launch?

Read the conversations the agent could not answer. In PepoChat, three places show them.

• Escalated conversations. Filter the team inbox by Escalated. The alert email "A visitor needs a person" states the reason, but a gap rarely shows there: every handover the agent makes, including one after it found nothing, gives the reason "The visitor asked for a person.", and "The agent could not find an answer or failed." appears only when a reply failed to generate. Open the transcript and look for the agent's "I don't have specific information about that" just before the handoff.
• Topics. The Analytics Topics card ranks the most asked topics over the last six weeks, labelled automatically from each visitor's first message. A growing topic with no page behind it is the next answer to write.
• Unanswered questions that never escalated. When the knowledge base has nothing relevant, the agent is instructed to say it has no specific information and to offer a person. A visitor who declines leaves no escalation behind, so export the conversations as CSV from the inbox and search the transcript column for "specific information about that" to find most of them.

Fix each gap the same day with a Q&A pair in the visitor's words, then correct the public page at your next review so the site and the chat agree. PepoChat's best practices recommend the same two habits: a Q&A pair for each wrong answer, and a page for each topic that keeps coming up. While a gap is open, the fallback message examples show what the agent should say instead of guessing.

How do you fill the gaps in 60 minutes before launch?

Import first, mark the checklist, write what is missing, then test. On a site with current pricing, shipping and returns pages, one focused hour is enough for a first pass.

1. Minutes 0 to 5: start the import. In Knowledge Base → Add New, start a crawl of your site or import its sitemap, and upload the two or three documents that hold prices or policies. Imports run in the background.
2. Minutes 5 to 20: mark the checklist. Go through the 50 questions and mark each one: answered on a page, answered in a document, written nowhere, or needs a lookup or a person. Strike out the questions your customers never ask.
3. Minutes 20 to 45: write the missing answers. Open Knowledge Base → Add New → Text & Q&A, choose Q&A, and write each answer in two or three sentences: the answer, then the condition, the number and the next step.
4. Minutes 45 to 55: test. Click Test the agent on the Knowledge Base page and ask fifteen of the questions in a customer's words, including three that should be handed to a person. Test chats are free, do not count toward your replies and stay out of Analytics.
5. Minutes 55 to 60: set the limits. On Widget Customization, add your refusals to Never do this and make your three most asked questions the widget's suggested questions.

At one or two minutes per answer, step 3 produces twelve to twenty-five Q&A pairs, and the Free plan holds 100. The launch checklist has the test questions and the handoff checks to run before the widget goes live.

What to do next

Copy the 50 questions into a spreadsheet, delete the ones your customers never ask, and write the gaps, starting with the money and delivery questions. If you sell software, the SaaS setup guide covers docs, billing and bug reports. To load the answers into an agent, create a free PepoChat workspace: you get every feature on the free plan except new-lead emails, with 200 AI replies a month, 5 knowledge sources and 100 Q&A pairs, and no credit card.