Turn your existing website into an AI agent's knowledge base. Learn how to prepare pages, choose what to import, test answers and keep content fresh.
Most businesses already wrote their best customer answers once: on the website. Pricing pages, shipping policies, FAQs and how-to guides are all there, yet the same questions still arrive in website chat, WhatsApp and Messenger. This article shows how to turn your site into a knowledge source for an AI support agent, how to prepare it, and how to check that the answers are right before customers see them.
Why your website is the natural first knowledge source
Your website is usually the most complete, most reviewed and most public description of your business. Someone has already approved the wording on the returns page. Someone has already decided how to describe your plans. Reusing that content means your agent starts from statements you have already stood behind.
There are three practical advantages:
- Speed. You do not need to retype policies into a new tool. You point to what exists.
- Consistency. The agent and the website say the same thing, so customers do not see one answer in chat and another on the page.
- A single place to fix things. When a policy changes, you correct the page, not a copy of the page.
The website is a starting point, not the whole answer. It describes what you want customers to know. It rarely contains what customers actually ask, which is often messier: "Can I change the delivery address after paying?" or "Do you ship to Singapore on weekends?" We will come back to filling those gaps.
What reading a website can and cannot give you
When a tool reads your website, it collects the text on your pages and makes it searchable for the agent. A good mental model: the agent can only answer from words that are actually on a page it read.
That has some consequences worth knowing before you start.
What tends to work well
- Plain text pages: FAQs, policies, service descriptions, about pages.
- Pages with clear headings, because headings tell the agent what each passage is about.
- Pages that state facts directly, such as opening hours, supported regions and fees.
What tends to work poorly
- Information locked in images. A price table saved as a picture, or a PDF-style banner with the return window written on it, has no readable text on the page.
- Content behind a login. Pages that require sign-in are generally not available to a crawler, and that is usually what you want for private account areas.
- Vague marketing copy. "Flexible plans for every team" does not help the agent answer "How much is it?" if the real numbers live somewhere else.
- Contradictory pages. If an old blog post says the return window is 14 days and the policy page says 30, the agent has two sources to choose from.
Prepare your site before you import it
A few hours of cleanup before importing saves a lot of correcting afterwards.
1. Make a list of the pages that matter
Start with the pages that answer the questions your team gets every week. For most businesses that list looks like this:
- Pricing and plans
- Shipping, delivery and returns
- Payment methods and invoicing
- Opening hours, locations and contact options
- Product or service descriptions
- Setup guides and troubleshooting
- Terms, privacy and warranty pages
2. Fix contradictions first
Search your site for the same topic across different pages. Old campaigns, archived blog posts and outdated landing pages are the usual culprits. Update them or remove them from the site before the agent reads them.
3. Turn images of text into real text
If a shipping table is an image, add the same information as a text table underneath it. This helps customers using screen readers as well, so it is worth doing regardless of AI.
4. Write headings as questions where it makes sense
"How long does delivery take?" is a better heading than "Logistics". The agent matches customer wording to page wording, and the closer they are, the better.
5. Check every language you serve
If your site has English, Japanese and Korean versions, each version is its own set of pages. Make sure the translated pages carry the same facts as the original, particularly prices and policy periods. A mismatch between language versions is one of the most common reasons an agent gives a confident but wrong answer.
Decide what should stay out
A crawl is not an all-or-nothing choice. Importing everything is tempting, but more content is not always better. Irrelevant pages dilute the passages the agent can find.
Consider leaving out:
- Old news posts and expired promotions
- Careers pages and press releases, unless customers actually ask about them
- Pages that are mostly navigation, such as tag archives and search results
- Duplicate pages produced by filters, sorting options or tracking parameters
- Anything you would not want quoted back to a stranger in a chat
That last point matters. Anything the agent reads can end up in an answer. Do not point it at pages containing internal prices for resellers, staff directories or draft policies that happen to be publicly reachable.
Also keep in mind that platforms put limits on crawls, such as how many pages one crawl will read. A focused set of 60 useful pages usually beats a sprawling set of 600 where the useful ones are buried.
Test with real questions before going live
Importing is half the job. The other half is checking what the agent actually says.
- 01Choose the pages that matter
- 02Import the site as a source
- 03Test with real customer questions
- 04Fix the page or add a Q&A
Build a test list from your own history. Open the last month of chats, emails and messages from WhatsApp or Messenger and pull out 20 real questions. Then ask each one and mark the result:
- Correct and complete. Nothing to do.
- Correct but vague. The page probably needs clearer wording.
- Wrong. Look for a contradicting or outdated page.
- No answer. The information is missing from the site, or hidden in an image.
For example, if you test 20 questions and 14 come back correct, six need attention. That number is purely illustrative, but the exercise is real: each failure points to a specific fix, and most fixes improve your website for human visitors too.
Pay special attention to questions where being wrong has a cost: refunds, delivery promises, medical or legal topics, and prices. For those, read the answer against the source page word by word.
Keep it fresh, and combine it with other sources
A website changes, and an imported copy of it does not change by itself. Build a small habit: whenever someone updates a policy, pricing or opening hours page, the same person checks that the agent's knowledge reflects it.
Websites also have blind spots, which is why most good setups mix several sources.
Practical combinations:
- Website plus short question-and-answer pairs for the questions customers ask that no page answers cleanly.
- Website plus uploaded files for manuals, brochures and spec sheets that are not on the site.
- Website plus a help-centre or document connector if your documentation lives in a separate tool.
- Website plus corrections from real chats. When a human agent rewrites a weak reply, that rewritten answer is worth keeping as training.
Common mistakes to avoid
- Importing before cleaning. The agent will faithfully repeat your outdated page.
- Never re-checking. A policy changes in March, nobody updates the agent, and in June it is still quoting the old one.
- Trusting the first few good answers. Test edge cases, other languages and unusual phrasing.
- Using the site as the only source. Customers ask things your site never says.
- Skipping the handoff plan. Even a well-fed agent will meet questions it cannot answer. Decide in advance who takes over and how.
Doing this in AI SNS CHAT
In AI SNS CHAT, a website is one of the knowledge sources you can attach to an agent. Everything starts from the sidebar.
- Open Data sources. In the sidebar, choose Data sources. You will see the source types: Files, Text snippets, Website, Q&A, Notion, Social, Tickets and Suggestions.
- Add a Website source. Select Website and provide your site address so the pages can be read into the agent's knowledge. Start with the section you cleaned up, such as your help or FAQ area, rather than the whole domain.
- Mind the limits. The maximum pages per crawl, the number of sources per agent, the size per source and the total knowledge-base size are limits set by the platform admin, so a very large site may need to be imported in focused parts.
- Test in Playground. Open the agent in Playground and ask your 20 real questions. For a side-by-side view of two models answering the same question, use the Playground A/B model compare.
- Fill the gaps. Add a Q&A source for questions your pages do not answer well, Text snippets for short facts, or Files for manuals. If your documentation lives elsewhere, the connectors under Channels cover tools such as Confluence, GitBook, ReadMe, Zendesk Guide and Google Drive.
- Improve from real conversations. Conversations from website chat, WhatsApp, Messenger and other channels appear in Activity, under Chat logs, where you can review ratings. Email conversations live in Inbox. In both, you can take over a chat and edit a reply into the ideal answer to train the agent.
- Track changes. Under Workspace settings, the Activity log records every create, edit, delete or upload with name, email, feature and time, and each data source shows who added it and who last edited it. That makes it easy to see who changed the knowledge behind an answer.
Two related points. Answering from the live web when the knowledge base has no answer is currently a LINE WORKS behaviour (Behaviour, then Web search), and it is not a general replacement for importing your own site. If you want any agent to look things up in real time, the Tavily action provides web search as an integration, which needs Hobby or higher. Also, knowledge-base passages shown to strangers are scrubbed for personal information, but you should still keep private pages out of the import.
If you get stuck on where a setting lives, Synthia, the built-in AI fleet manager, can walk you to the right screen and highlight the control. Plans, limits and current prices are on the pricing page, and the full list of capabilities is under features.
Frequently asked questions
Will the agent know about changes I make to my website later?
Treat the imported content as a snapshot of the pages when they were read. When you change a policy, price or opening hours, update the source and test the affected questions again. Checking the data source in your dashboard, and who last edited it, helps you keep track.
Should I import my entire website or only some pages?
Usually only the pages that answer customer questions. Pricing, shipping, returns, FAQs and guides help the most. Old news, careers pages and duplicate listing pages add noise, and platform limits on pages per crawl and knowledge size mean a focused import is also easier to manage.
What if a customer asks something my website never mentions?
The agent cannot invent a reliable answer from nothing, so fill the gap with a Q&A entry, a text snippet or an uploaded file. You can also set up human handoff so a team member takes over when the agent has no good answer.
Can I use the website as knowledge for WhatsApp and Messenger, or only for website chat?
Knowledge belongs to the agent, not to a single channel. An agent that has learned from your site can use the same knowledge wherever you deploy it, including website chat, WhatsApp and Messenger, as long as those channels are connected to it.


