How to find and fix chatbot knowledge gaps
Turn real No Information questions into a prioritized, owned and repeatable knowledge-base improvement process.
Chatbot knowledge gaps are easiest to fix when you start with the exact unanswered question, inspect its context and verify the improved source with variants.
A useful chatbot says when its sources do not support an answer. WebChatAgent can store that question, the assistant’s response and a short context summary in Questions. Entries classified as “No Information” form a practical inbox for likely knowledge gaps.
Not every row should become new content. A question can be outside the bot’s intended scope, contain private data, ask for an unpublished policy or repeat an answer that already exists under different wording. Review the original context and the authoritative owner before changing the knowledge base.
The verified example asks: “What is the 2028 warranty policy for Northstar Services?” The fictional source contains no such policy, so the assistant safely says it has no information. Questions then stores the exact request, answer and context as No Information. This proves detection; it does not justify inventing a 2028 warranty policy.
Privacy-protected two-click player
How to Find Chatbot Knowledge Gaps (Step by Step)
Find chatbot knowledge gaps from real unanswered questions, prioritize recurring demand, add the right source and retest question variants safely.
The YouTube player stays blocked until you choose Play. Loading it connects your browser to YouTube and may transfer technical data to Google.
Open directly on YouTubeWhat you will have at the end
- A reproducible No Information filter for the intended chatbot and date range
- A triage rule based on frequency, business impact, risk and scope
- The original question, answer and context reviewed together
- A documented decision to update an authoritative source, add approved information or keep the safe non-answer
- A fixed before-and-after regression question with an assigned content owner
Before you start
- Questions collection enabled for the chatbot
- Access to Questions and the relevant source content
- Enough real traffic to identify repeated wording; one controlled question is enough to learn the interface
- A content owner who can approve the expected answer
- Permission to update and re-index sources when a real gap is confirmed
Detect, triage, verify, then change
Classification is a signal, not an automatic publishing instruction. Frequency estimates reach, business impact estimates value and the conversation context reveals what the visitor actually meant. Scope and safety can make a deliberate non-answer the correct outcome.
When a real gap exists, update the smallest authoritative source that the organization already maintains. Avoid a hidden duplicate inside the chatbot if the website, handbook or policy document should be the source of truth. Indexing and a repeated test are required before the gap can be called closed.
01–05
Set it up step by step
Filter the Questions inbox to “No Information”
Start with one chatbot and a reproducible scope.
Open Dashboard → Questions. Choose the intended chatbot, then set Classification to No Information. Add language, date-range or search filters only when you need to narrow a larger queue. Keep the original unfiltered total in your notes so later reports remain comparable.
The visible example contains one Tutorial Lab question. It asks for a 2028 warranty policy and shows the assistant’s safe non-answer beside the No Information label. This row is a candidate gap, not proof that a policy should exist.
- Save chatbot, filters and review date together.
- Do not treat Answerable as automatically correct; it only indicates that supporting information was found.
- Protect personal data before exporting or sharing the list.
Prioritize by frequency, impact, risk and scope
Repeated questions matter, but one high-risk miss can outrank them.
Read the question, assistant answer, classification and count together. Group natural variants that describe the same intent before calculating frequency. Otherwise “refund”, “money back” and “cancel my annual plan” can become three small rows instead of one important need.
Use a simple triage rule: reach, business impact, legal or safety risk, and whether the topic belongs to this assistant. A frequent purchase blocker is usually more valuable than a rare curiosity. A single compliance or safety gap should be escalated immediately. Out-of-scope questions may correctly remain unanswered.
- Give every accepted gap an owner and due date.
- Do not rank solely by count when severity differs.
- Merge wording variants before adding new content.
Open Question Details and verify the visitor’s intent
Review the exact question, answer, context and session together.
Open the row or its detail control. Question Details shows the chatbot, session ID, language, classification, creation time, exact question, stored answer and context summary. Read these fields before selecting Add data source. A short isolated question can mean something different after the preceding messages.
In the verified record, the exact question is “What is the 2028 warranty policy for Northstar Services?” The stored answer says no information is available, and the context confirms that the requested year-specific policy is absent. The correct review result is “needs policy-owner confirmation”, not “generate a warranty answer”.
- Use the session reference when a content owner needs the complete conversation.
- Redact personal data from copied examples.
- Record the approved expected answer before editing sources.
Choose the smallest authoritative source change
Add approved information, not a plausible AI guess.
Select Add data source only after the owner confirms the answer. Use the text field for one stable, narrowly scoped fact. Upload a maintained TXT, PDF, DOC, DOCX, MD or XLSX file when the answer belongs to a broader policy or handbook. Give the source a descriptive name, category, owner and effective date.
If the official website or handbook should contain the fact, update that system first and re-index it instead of creating a hidden duplicate. Review any Optimize with AI proposal line by line; it can improve structure, but it cannot approve business facts. The English Light Mode screenshot proves the complete dialog state and supported file formats, not a completed source update.
- Never publish a policy that the owner has not approved.
- Avoid duplicate text when an authoritative source already exists.
- Budget indexing quota and wait for Completed before retesting.
Record the safe baseline, then retest after indexing
A closed gap needs before-and-after evidence with the same question.
Before changing content, preserve the exact visitor question and response. The real English Light Mode chat shown here asks for the 2028 warranty policy. The assistant replies that the information is unavailable. This is the verified baseline and the correct result while no approved policy exists.
After an approved source change reaches Completed, reset the chat and ask the exact original plus two natural variants. Confirm the answer matches the source, cites or clearly reflects the right evidence and does not break nearby questions. Then return to Questions and check whether the item becomes Answerable as documented by the product. That post-fix state was not captured in the current evidence set, so the tutorial must not claim the fictional policy gap is closed yet.
- Keep the original question unchanged for the primary regression test.
- Use variants to test robustness, not to replace the baseline.
- Monitor whether the same intent returns as No Information over the next review period.
Example & result
See the practical test and its result
Every tutorial includes a fixed input, the expected outcome and a transparent record of what was actually verified locally.
Practical example: How to find unanswered chatbot questions and knowledge gaps
This exact scenario was completed with the temporary tutorial account.
Exact test input
Ask: “What is the 2028 warranty policy for Northstar Services?”
Expected result
The bot abstains safely and the question becomes visible as a knowledge gap.
What was actually verified
The live chatbot said it had no information. Questions then stored the exact question, answer and context with the classification “No Information”.
Tips & tricks
Make the setup reliable
Test with realistic examples, record your baseline and change one setting at a time. That makes real improvements visible.
Fix the source of truth
Update the maintained website, handbook or policy first and re-index it. A chatbot-only duplicate will become stale and can conflict with the official version.
Keep a gap backlog owner
Every accepted gap needs an owner, expected answer, due date and verification question—not just a dashboard row.
Use a model change only for a model problem
A smarter model cannot retrieve a fact that is absent. Compare models only when the correct passage is consistently retrieved but handled poorly.
Review on a fixed schedule
A weekly review works for active support bots. Record open, accepted and closed gaps so the team can see whether coverage is improving.
When something does not work
Troubleshooting
Check status, permissions and test data systematically before changing the model or prompt.
Questions stays empty although visitors report failures
Confirm collection is enabled for the intended chatbot, remove restrictive filters and verify the date range. Reproduce one safe test question, then check the exact chatbot and session.
A known answer is classified as No Information
Search the indexed content for the exact fact. Check source status, extraction, category scope, duplicates and recent updates before adding more content.
Several rows describe the same need
Group them under one intent, keep representative wording and add one maintainable source answer. Do not create one source per phrasing.
The new source is saved but the answer has not changed
Wait for Completed indexing, confirm the fact appears in Content Search, reset the conversation and repeat the exact baseline. Clear conflicting outdated sources if necessary.
The bot invents details around the newly added fact
Rewrite the source with explicit scope, conditions and effective date. Add a safe evidence rule and test unsupported variants that should still receive a non-answer.
Ready for a production-style test
Review No Information on a fixed schedule. For each accepted gap, store the intent, frequency, impact, owner, authoritative source, approved expected answer and regression question. Mark it closed only after indexing, a real post-fix answer and the corresponding Questions status are verified.
Related resources
Understand Data Sources and indexing states
Choose the right source type, inspect status and maintain indexed content safely.
Test the corrected answer in Knowledge Optimizer
Use fixed questions, retrieved evidence and regression checks after the source update.
Questions product reference
Review the current filters, classifications and Add data source behavior.