What to Do When Your Chatbot Says "I Don't Know"

"I don't have information about that. Please contact our team for details."

Your chatbot has said this. Probably today. Probably to someone who was ready to buy.

The reply is polite, it is honest, and it is a dead end. The visitor came with a question, your website did not answer it, and now the only record of that failure is one line buried in a transcript nobody will read.

The problem is not that your chatbot said "I don't know." Given the content it had, that was the correct thing to say. The problem is that nobody found out.

TL;DR#

An "I don't know" is not a chatbot bug. It is a content gap your chatbot just discovered for you.
Bad answers come in five distinct types, and each one needs a different fix. Treating them all as "the bot got it wrong" is why most teams never fix any of them.
One missed question is an anecdote. Fourteen missed questions on the same topic is a content decision.
Reading transcripts does not scale. Grading answers automatically does.
Denser's Answer Quality grades every answer, labels why it fell short, and groups the gaps by topic.

Why "I Don't Know" Is Actually Good News#

A chatbot that says "I don't know" is behaving correctly. A grounded assistant should refuse to answer when its sources do not cover the question, because the alternative is inventing something.

The failure mode you should fear is the opposite one: a confident, fluent, completely fabricated answer about your refund policy. Compared to that, "I don't know" is a well-behaved system telling you the truth about its own limits.

So the honest framing is this: every "I don't know" is a free, unsolicited piece of market research. A real person, at a real moment of intent, told you exactly what your website is missing. Most companies pay for that information in survey tools. Your chatbot generates it every day and throws it away.

Compare it to your other research inputs:

Source	Intent	Sample	Their words or yours?
Keyword tools	Mixed, mostly strangers	Huge	Neither, it's aggregate
Surveys	Low, you interrupted them	Small, self-selected	Yours, you wrote the questions
Sales calls	High	Tiny, and filtered	Theirs, but only for people who booked
Chatbot questions	High, mid-evaluation	Everyone who visits	Theirs, unprompted

Chatbot questions are the only source that is simultaneously high-intent, unprompted, and captured at scale. They are also the only one you are already paying for and not using.

The question is not how to stop your chatbot saying "I don't know." It is how to make sure someone reads it.

The Five Ways an Answer Falls Short#

"The chatbot got it wrong" is not something you can act on. Before you can fix anything, you need to know which kind of wrong it was, because each one points at a different fix.

1. Missing content#

The knowledge base simply has nothing on the subject. The bot admits it, or redirects the visitor to "contact the team."

What it means: your content does not cover this. The fix: write or upload the page. This is the most common flag and the most actionable one.

2. Incomplete#

The answer gave real information, then explicitly left another material part of the question open. A visitor asks about pricing and delivery times, gets pricing, and gets nothing on delivery.

What it means: your source covers half the subject. The fix: revise the page it cited rather than adding a new one.

3. Contradicts its own sources#

The answer makes claims that conflict with each other, or with the passages it cited. Two different prices for the same plan. Two different warranty periods.

What it means: you have conflicting content, and the chatbot is not the problem. The fix: find the source conflict and fix your source of truth. This is what a knowledge base audit is for.

4. Out of scope#

The question was not this chatbot's job, and it handled that badly: rambling, pretending to knowledge it lacks, or leaving the visitor stranded.

What it means: your bot's instructions need tightening. The fix: teach it to decline cleanly. Note that a clean decline is not a failure. If someone asks your e-commerce bot for legal advice and it politely says it cannot help there, nothing is broken and there is nothing to fix.

5. A visitor told you#

Someone clicked thumbs-down. No automated judgment beats a human one, so this always overrides a passing grade.

What it means: the answer looked fine and was not. The fix: open the conversation and see what actually went wrong.

Five causes, five different fixes. This is why "improve the chatbot" never gets done. It is not one task, and until the flags are sorted by cause it looks like a wall of undifferentiated failure.

Why Reading Transcripts Does Not Work#

The standard advice is to review your chat logs. Here is what actually happens when you try.

A chatbot doing modest volume produces a few hundred conversations a week. Most of them are fine. The bad ones are not marked, so finding them means reading everything. Nobody has that time, so it gets scheduled for "next quarter," and next quarter it is thousands of conversations instead.

Thumbs-down feedback seems like the shortcut, and it is the standard answer from most chatbot tools. But the vast majority of dissatisfied visitors never click anything. They read "I don't have information about that," close the tab, and go to a competitor. That silence is your largest and least visible category of failure, and a feedback button cannot capture it.

What you need is for every answer to be checked, not just the ones a visitor bothered to rate.

Grade Every Answer Instead#

The practical approach is to have a language model read every question and answer pair against a fixed rubric, decide whether the answer actually resolved the question, and label the cause when it did not.

Three properties make this work in practice:

It runs on everything. Every answer gets checked, not the 2% that get rated.

It runs after the fact. Grading happens on a schedule once conversations have gone idle, so it is never in the path of a waiting visitor. Your chat speed does not change.

It is careful about what counts as failure. A grounded factual negative such as "we do not offer refunds after 30 days" is a good answer. An "I don't know" is not. An off-topic question declined cleanly is fine. The distinction matters, because a grader that flags everything is as useless as one that flags nothing.

One Missed Question Is an Anecdote. Fourteen Is a Decision.#

Sorting by cause is only half of it. The other half is grouping by subject.

Label each question with a short topic, group the flagged ones, and count. Now the output is not two hundred individual failures but a ranked list:

pricing: 14
integrations: 9
security review: 6
shipping times: 4

That list is a content roadmap, written by your customers, ranked by demand. Nobody had to guess which page to write next. Fourteen people already voted.

This is where the exercise stops being quality assurance and starts being marketing intelligence. Those fourteen pricing questions are also telling you something about your pricing page.

Read the Wording, Not Just the Count#

The second thing hiding in these questions is vocabulary.

Your pricing page says "seats." Your visitors ask about "users." Your docs say "authentication," they ask about "logging in with Google." You call it an "implementation partner," they ask "do you do the setup for us?"

That mismatch costs you twice. It hurts retrieval, because the chatbot has to bridge a vocabulary gap to find the right passage. And it hurts search, because you are optimising for language your buyers do not use.

Failed questions hand you their phrasing verbatim. Use it. The heading on the page you write should sound like the question that prompted it.

A Practical Workflow#

Let it run for two weeks. You need enough conversations for topic counts to mean something.
Start with the biggest topic cluster, not the newest flag. Volume tells you where the money is.
Sort by cause within the cluster. Missing content means write the page. Contradicts sources means fix the conflict. Incomplete means revise what is already there.
Fix it at the source, not in the bot. Adding a one-off Q&A pair patches the chatbot; updating the underlying page fixes the website too, for the visitors who never open the chat.
Dismiss what does not matter. Out-of-scope questions and one-off curiosities are noise. Clear them so the real signal stays visible.
Check back next month. The topics that vanish are the ones you fixed. The topics that persist are the ones your content still is not landing.

How Denser Handles This#

Denser grades every answer your chatbot gives, on every plan including Free, so no history is lost while you get set up.

Each flagged answer arrives with its cause attached, a one-line rationale, and the fix within reach: upload the missing content, add or edit the Q&A pair prefilled with the question that failed, jump to the cited source to revise it, or open the full conversation for context. Nothing changes in your knowledge base until you approve it.

Flagged questions are clustered by topic with counts, so the content roadmap builds itself. Browsing them is included on the Standard plan and above, and every paid plan starts with a 7-day free trial, so you can see your own gaps before you pay.

Create a free Denser account and let it grade a week of conversations. The first list is usually uncomfortable, and it is the most useful content brief you will get this quarter.

FAQs About Chatbots That Cannot Answer#

Why does my chatbot say it doesn't know?#

Because its knowledge base does not contain the answer. A grounded chatbot only answers from the content you gave it, so "I don't know" means the content is missing, the wrong page was indexed, or the answer exists but is worded so differently that retrieval missed it.

Is it bad if my chatbot says "I don't know"?#

It is far better than the alternative, which is inventing an answer. The failure is not the honest refusal, it is nobody finding out that it happened.

How do I find the questions my chatbot cannot answer?#

Either read every transcript, which does not scale past a few hundred conversations, or grade answers automatically. Denser's Answer Quality checks every answer, flags the ones that fall short, and groups them by topic.

Can I just rely on thumbs-down feedback?#

No. Most unhappy visitors never rate anything, they simply leave. Feedback buttons capture a small and biased sample of your real failures.

What is the difference between a knowledge gap and a retrieval problem?#

A knowledge gap means the content does not exist and you need to write it. A retrieval problem means the content exists but the search did not surface it, usually because the visitor's wording differs from your page's. Checking whether the page exists tells you which one you have.