Training a chatbot on documentation sounds like a technical project. In practice, most of the work is not technical at all. It is making sure your docs contain the answers people actually need, then checking the assistant finds them.
This guide walks through the setup for a SaaS team, from the first crawl to the weekly routine that keeps answers accurate.
Step 1: tidy the docs before you train anything
A docs chatbot is a mirror. If your documentation is out of date, the assistant will be too. Spend an hour before training on three checks:
- Remove or update pages about retired features. Old pages are the most common source of wrong answers.
- Check your most visited articles for accuracy. Screenshots and menu names drift after redesigns.
- Make sure setup, billing and account basics are covered. These drive a large share of questions.
This is the unglamorous part, and it pays off more than anything else in this guide.
Step 2: crawl your docs and help centre
With SpideyChat, you point the crawler at your docs site and it reads your pages, up to 150 of them, building a searchable index. On a typical documentation site this takes a few minutes.
If your docs live on a separate subdomain, include it. If some content is in PDFs or internal files, upload those directly. You can also add question and answer pairs for things that are not written anywhere yet. The training guide covers each method.
Step 3: add the context the docs leave out
Docs explain how features work. They often skip the things people ask before they get that far:
- Which plan includes which feature.
- What happens at usage limits.
- How to contact a person, and when support is available.
- What the product does not do.
A few short Q&A entries on these topics close the most common gaps. Pricing questions in particular come up constantly, as we covered in pricing page questions that delay a SaaS purchase.
Want a 24/7 AI Employee for Your Website?
See how SpideyChat can answer customer questions, capture leads and handle enquiries while your team is offline.
Step 4: test with real questions
This is the step people skip. Do not test with "what does your product do?" Test with the questions your support team actually receives, in the words users actually use.
Pull twenty recent tickets and ask each first message to the assistant word for word. Score each answer:
- Right: accurate and useful.
- Vague: technically correct but not helpful enough.
- Wrong or missing: the docs do not cover it, or cover it badly.
Vague and wrong answers are almost always content problems. Fix the docs, re-crawl, and ask again.
Step 5: install it where questions happen
Put the assistant on your docs site, your marketing site and, if it makes sense, your app's help menu. The install is one script tag, and the installation guide covers the common platforms.
Placement matters. Trial users get stuck in the docs and on the pricing page far more than on the homepage. As we explain in trial users who never ask for help rarely convert, those are the users who will not open a ticket.
Step 6: set the escalation rules
Decide what should always reach a person: bug reports, billing disputes, security concerns and anything about data loss. Configure the assistant to open a ticket or hand over for these, with the conversation attached.
This keeps the assistant from trying to handle things it should not, and it means your team receives better tickets than before, with context included.
Step 7: make the gaps list a weekly habit
After launch, the most valuable thing the assistant produces is a list of questions it could not answer. SpideyChat for SaaS companies logs these as knowledge gaps.
Treat that list as a docs backlog. Each week, pick the most frequent gaps and write or fix the relevant articles. Re-crawl. Watch the same questions start getting answered. This habit also changes how you think about your help centre, which is easy to see even in a different setting like the one in a returns policy nobody reads is still your busiest page: the questions tell you what the page should have said.
What good looks like after a month
A well-trained docs assistant answers most routine how-to questions instantly and accurately. It hands the rest to your team with context. And your documentation is noticeably better than when you started, because the gaps list forced you to fill the holes users kept falling into.
That last effect is easy to underestimate. Better docs help your SEO, your onboarding and your support team, whether or not anyone uses the chat.
Start with the crawl
Pick an afternoon, tidy your top twenty docs pages, and run the crawl. Then test with twenty real tickets. Within a couple of hours you will know how well your documentation answers your users, and exactly where to improve it.