Train a Chatbot on Website Content: 6 Common Mistakes

When you train a chatbot on website content and your catalog, the result depends less on the model and more on what you feed it. Outdated pages, scanned PDFs and contradictory text make the bot answer wrongly, with complete confidence. Here are the six most common mistakes and how to fix each, including a routine to keep the knowledge base alive.

Mistake 1: feeding the bot outdated content

Old pages, prices from past campaigns and discontinued products linger on the site, and the bot treats them as truth. The result: it promises lead times and prices that no longer exist.

Fix: before importing, take inventory. Mark what is current, what to archive and what to delete. Every document should carry a review date and an owner.

Mistake 2: using poor PDFs without cleanup

Scanned catalogs, tables as images and PDFs with tangled columns turn into scrambled text. The result: the AI reads a table row as if it belonged to another product and mixes up dimensions and part codes.

Fix: prefer the original source, such as a spreadsheet or editable file. If only a PDF exists, convert tables to structured text and spot-check several products by hand.

Mistake 3: contradictory information across sources

The site states one warranty, the catalog another and the returns policy a third. The result: the bot answers differently depending on how it is asked, and the customer quotes whatever it said.

Fix: name one official source per topic (warranty, lead time, shipping, terms) and remove or correct the rest.

Mistake 4: not defining what the bot should not answer

With no limits, the bot invents rather than stay silent. The result: prices, discounts or specifications made up on the spot.

Fix: instruct it to say it lacks the information and hand off to a person on topics like firm quotes, special lead times and technical certificates.

Mistake 5: testing only with easy questions

Whoever builds the bot asks what it knows. Real customers write with typos, abbreviations and truncated part numbers. The result: the test passes and production fails.

Fix: assemble 50 real questions from emails, chat messages and the sales team, and rerun them after every update.

Mistake 6: training once and forgetting

The site changes, the catalog gains items, policies get revised. The result: quality drops slowly and nobody notices until the first customer problem.

A simple routine to keep the knowledge base current

• Every price, product or policy change: update the base the same day, by whoever published the change.
• Every week: read the conversations where the bot could not answer and turn the gaps into new content.
• Every month: rerun the 50 test questions and compare with the previous result.
• Every quarter: review the inventory and archive what is no longer valid.

If you want help structuring the base and the routine, take a look at our AI and automation work. The rule holds for any tool: the bot is only as good as the content you maintain.

Frequently asked questions

Can I use my website as the chatbot’s only source?

It can be a good start, as long as the content is current and consistent. Add FAQs and catalog data for technical questions.

How often should I update the chatbot’s knowledge base?

Whenever prices, products or policies change, plus a weekly review of unanswered questions.

What should I do when the chatbot answers wrongly?

Fix the source of the error, whether stale, contradictory or badly parsed content, and add the question to your test set.