A customer types: "Do you ship to Canada, and can I split payment over three months?" Your bot knows your shipping zones cold. Nobody ever wrote a line about financing. So what does it say?
That single moment decides whether people trust the thing. A bot that confidently invents a payment plan you don't offer is worse than no bot at all. The good news is that handling unknown questions well is mostly a matter of setup, not luck.
Where the gaps actually come from
Every chatbot has a knowledge boundary. It knows what you fed it, plus whatever general reasoning the underlying model brings. Questions fall outside that boundary for a few ordinary reasons.
Sometimes the answer exists but lives somewhere the bot can't see, like a policy buried in a PDF you never uploaded. Sometimes the question is real but you've genuinely never decided the answer, like that installment plan. And sometimes people ask things that have nothing to do with your business, from the weather to their horoscope.
Each of these needs a different reaction. The mistake is treating them all the same, or worse, letting the model treat "I don't have this" as an invitation to improvise.
The dangerous default: filling the silence
Language models are built to produce a plausible next sentence. Left unguided, that instinct works against you. Ask an ungrounded bot about a warranty you don't offer and it may cheerfully describe one, complete with a return window it made up.
This is the behavior people mean when they worry about "hallucination." It isn't lying, exactly. The model is pattern-matching toward what a helpful answer usually looks like, with no check on whether the specifics are true for your business.
The fix has two parts. First, ground the bot in your own content so its answers are drawn from real source material, not general guesswork. Second, tell it plainly what to do when the source material comes up empty: admit it, and route the person somewhere useful. In SpideyChat you set both of these when you build the bot, so the fallback behavior is decided before a single customer ever sees it.
What a good "I don't know" looks like
A blunt "I cannot help with that" ends the conversation and annoys people. A better fallback keeps the door open. Compare these two replies to "Can I pay in installments?"
Weak: "I'm sorry, I don't have information on that."
Better: "I don't have a confirmed answer on installment payments, and I'd rather not guess about your money. Want me to grab your email and have our team confirm today, or connect you with someone now?"
The second version does three things. It's honest about the limit. It explains why it isn't guessing, which actually builds trust. And it offers a next step instead of a dead end. That last part is what turns a knowledge gap into a captured lead rather than a lost visitor.
Build the escalation ladder
Think of unknown questions as a ladder, not a wall. The bot climbs down one rung at a time.
- Try the knowledge base. Search your uploaded content and site crawl for a real match before anything else.
- Reframe once. If the phrasing is odd, the bot can rephrase and search again. "Refund" and "money back" should land in the same place.
- Offer a partial answer plus a caveat. If it knows 80 percent, it can share that and flag what it can't confirm.
- Capture and route. No confident match? Collect an email or open a human handoff, tagged with the exact question.
- Log it. Every unanswered question gets recorded so you can decide whether it deserves a permanent answer.
That last rung matters more than it looks. Your unanswered-questions log is a free list of exactly what your customers wish your site explained.
There's a subtle judgment call between rungs three and four. A bot that hands off too eagerly annoys people who wanted a quick answer, while one that stretches for a partial answer risks getting it wrong. The right balance depends on the stakes. For a low-risk question about, say, your company's history, a best-effort partial answer is fine. For anything touching a customer's money or account, err toward the handoff. You can tune that instinct, and you should revisit it after reading a couple weeks of real conversations rather than guessing at it up front.
A quick before-and-after
Take Northside Bikes, a small shop with a chatbot trained on its product pages and store hours. In week one, the bot fumbled a recurring question: "Do you service e-bikes from other brands?" The store did, but the answer lived only in the owner's head.
The bot didn't invent a policy. It said it wasn't sure, offered to take a phone number, and logged the question. By Friday the owner noticed the same query eleven times in the log. He added two sentences to the service page and re-synced the bot. The next week, that question answered itself, and the handoffs dropped to near zero.
Nothing about that required fancy AI. It required a bot honest enough to say "not sure" and a log worth reading.
Tune the confidence, not just the content
You can shape how cautious the bot is. A support bot for a clinic should lean conservative, handing off anything touching medical or billing specifics. A bot answering general questions on a blog can be more relaxed, since a slightly imperfect answer about, say, your company history carries low stakes.
Here's a rough way to think about where to set the dial:
| Question type | Right instinct |
|---|---|
| Pricing, refunds, contracts | Answer only from confirmed content, else hand off |
| Account or order specifics | Verify identity or route to a human |
| General how-to and product info | Answer confidently from your docs |
| Off-topic or personal | Politely decline and redirect to what you can help with |
The goal isn't a bot that answers everything. It's a bot that's right when it speaks and graceful when it can't be.
Turn gaps into a feedback loop
The real payoff comes over time. A chatbot that logs what it couldn't answer hands you a prioritized to-do list every week. You're not guessing at what content to write. You're responding to questions real visitors already asked.
A simple rhythm works well: skim the unanswered log once a week, pick the three most common gaps, write short answers, and re-sync. Do that for a month and the bot's "I don't know" rate quietly falls, because you keep feeding it the exact things people wanted.
Handling unknown questions isn't about having every answer on day one. It's about being honest in the moment and getting a little smarter each week. Set the fallback thoughtfully, watch the log, and your bot earns trust precisely because it knows the edge of what it knows. Want to see how the fallback and handoff feel in practice? Spin up a bot on the demo and try to stump it.