An unanswered-question report is useful only when it helps a team decide what to fix. The most common mistake is to treat every “I don’t know” as missing content. Sometimes the answer really is absent. Sometimes the knowledge exists but search did not find it. Sometimes the question asks for a decision that no public source should make.
Start by separating the three outcomes
Classify each repeated unanswered question before writing anything new:
- Genuine gap: no approved source covers the visitor’s need.
- Retrieval failure: a relevant knowledge item exists, but the visitor’s wording did not retrieve it or it was not used.
- Human-only question: the answer depends on an account, an exception, negotiation or a decision by a colleague.
This protects the knowledge base from duplicate entries and keeps people involved where judgment is required. It also produces better analytics: a report of “unanswered questions” should show the work type behind the count, not merely a pile of chat transcripts.
Prioritise impact without losing the evidence
Group similar questions and rank them by frequency, recency, business risk and visitor journey. Five requests for a shipping restriction on a checkout page may deserve attention before twenty vague questions that are off-topic. Preserve the original visitor wording even when an editor writes a cleaner internal title; the wording is evidence of what people actually searched for.
RobiFox’s missing-knowledge screen groups similar questions and shows how often they were asked. Its workflow preserves the original visitor question and supports improving the wording used to find relevant knowledge. That combination is useful: the team can write clear knowledge without erasing the phrase that revealed the gap.
| What you find | Next action |
|---|---|
| No relevant candidate | Write an approved answer if the business wants the assistant to answer it; otherwise mark it human-only or dismiss it. |
| Relevant candidate exists | Link the visitor wording to the existing knowledge and improve retrieval; do not create a duplicate fact. |
| Conflicting candidates | Resolve the source conflict before teaching either answer. |
| Account-specific or discretionary request | Keep it out of general knowledge and provide the correct human route. |
Teach a new answer deliberately
For a genuine gap, write the answer in the way a helpful colleague would: short, specific and conditional where needed. Include the source or decision owner behind it. Then make it searchable and wait for the system’s re-encoding step before assuming the assistant can retrieve it. RobiFox’s interface is clear on this sequence: saving sets the answer aside; teaching it creates an approved item; the assistant uses it after the next re-encoding.
Test the original question and variants after the update. Check the answer, cited material and a nearby edge case. If the new content causes a conflict with an older policy, route the conflict to the content owner instead of masking it with another explanation.
Fix failed retrieval without duplicating knowledge
When candidates already exist, inspect them with the original conversation. Is one actually the correct answer? Is its question too formal, its terminology too narrow, or its source outdated? A repair may mean adding the visitor’s wording to the searchable text, correcting the existing item or removing a conflict. It should not mean adding a second answer that says nearly the same thing.
RobiFox provides this exact branch in its missing-knowledge flow: an operator can mark an existing knowledge item as the answer, after which the original wording is added to searchable text and re-encoding starts. The interface warns that a duplicate could make the assistant answer from conflicting sources. That is a useful safeguard for any knowledge-base team.
Close the loop with product and content work
Review the top categories weekly with the people who can act: content owners, support, sales and product. Decide whether the right repair is a knowledge item, a clearer website page, an adjustment to routing or no automated answer at all. RobiFox’s conversation view preserves candidates, citations and timing data, while its public product description positions unanswered questions as a frequency-ranked work list. Those records make the review specific rather than speculative.
Track whether the same category falls after the repair, but do not treat a lower count as proof that the answer is good; seasonality and traffic change too. Pair volume with samples of the resulting conversations. For more on the controls behind this loop, see RobiFox features, How it works and How we measure.
For the next step, turn repeated questions into better FAQs.
Cover image: AI-generated editorial illustration, not a product screenshot.
