Ask most chatbots something they cannot know, and they answer anyway. The answer is fluent, confident and wrong. On a website that is worse than no chatbot at all, because the visitor believes it.
We built ReplyBash to do the opposite: answer from your own pages, show where the answer came from, and say so when it does not know. Here is why we think that is the right default, and what it gives you.
A wrong answer costs more than a missing one
A visitor who gets "I am not sure, here is how to reach the team" loses a minute. A visitor who is told your clinic is open on Sundays when it is not turns up at a locked door. A buyer who is quoted a price you do not charge writes to you angry, or does not write at all.
The wrong answer is also hard to catch. Nobody on your side sees it happen, and the visitor rarely tells you. The only sign is the one you never get: a customer who quietly leaves.
Why chatbots invent answers
A language model is built to continue text in a plausible way. Given a question and nothing relevant to read, plausible is exactly what it produces. It is not lying; it is doing what it was built for with nothing to stop it.
The fix is not a smarter model. It is to give the agent your content to answer from, and to treat "the content does not say" as a real outcome rather than a failure to avoid.
What an honest agent does instead
When ReplyBash cannot find the answer in your site or files, it says so and offers a next step: your contact details, a form, or a handover to someone on your team. It does not guess a number, a date or a policy.
When it can answer, it links to the page it used. That link is not decoration. It lets the visitor check, and it lets you check, which is how you find the page that needs fixing.
Every "I don't know" is a to-do item
This is the part most people miss. The questions your agent could not answer are a list of what your website does not say. They are what your visitors want to know and cannot find.
In ReplyBash they appear as unanswered questions. Answer one once, as a Q&A, and the agent uses your wording from then on. A daily email brings the new ones, so you do not have to go looking.
Over a few weeks the list gets shorter, and the pages you should write get obvious: the missing price, the unclear delivery time, the policy nobody can find.
What to do this week
- Try to break your own agent. In the Playground, ask five things your site does not cover and check that it admits it.
- Read the unanswered questions. They are in the dashboard Overview, and in Analytics under Insights.
- Answer the top three as Q&A pairs, using the exact wording you want visitors to see.
- Decide who catches the rest. Anyone on your team can take over a chat when a visitor needs a person.
The short version
An agent that sometimes says "I don't know" is one you can trust when it says something else. Start with training it on your website, and let the questions it cannot answer tell you what to write next.