Support knowledge gap analysis: where your support team runs out of answers

Short answer

To see your support team's knowledge base gaps, look for three signs in conversations: a question left unanswered (it is not in the knowledge base), an answer given with outdated information (it is there but out of date), and different answers to the same question (the knowledge base contradicts itself or is unclear). Each gap gets a content owner, and a fix cycle runs once a month.

A knowledge base checked once before launch and then forgotten goes stale fast. Conversations show every day where it fails — collecting that signal is the cheapest way to keep it alive.

This differs from a pre-launch audit

Checking a knowledge base before launch is a separate job (chatbot knowledge base audit). This article is about the stage after launch: the knowledge base is live, agents and the bot use it, and customers test its edges with questions every day. The aim here is not a one-off audit but a continuous maintenance cycle.

Gap type 1: no answer

The customer asks, the agent says "I'll check and get back to you", the bot replies "I don't have information on that" or hands the conversation over. Signs: the same question repeating, followed by a late answer or none. This gap is the easiest to find and the fastest to fix: the answer needs to be written and approved.

Gap type 2: an outdated answer

An answer exists but is out of date: a price, opening hours, delivery time or rule has changed. Sign: after the answer the customer says "the website says something else" or "that was last month", or comes back with the same problem in a later conversation. This gap is the most dangerous, because the bot and agents confidently give wrong information.

Gap type 3: contradiction

The same question gets different answers: one agent says "returns are 14 days", another "30 days". The cause is either two different documents in the knowledge base or a vaguely written rule. Sign: different answer sentences on the same topic. To find it, read the answers from 15–20 conversations on the same topic side by side.

How to collect gaps from conversations

1. 1Field"Knowledge gap type": no answer / outdated / contradictory / none. Give example sentences in the instruction.
2. 2TopicWhich question it is: a closed list (price, delivery, returns, technical, documents, etc.).
3. 3EvidenceThe customer's question and the answer given — in one place.
4. 4FrequencyHow often the gap appears in each topic — for prioritising.

The content owner

The weakest point of a knowledge base is ownerless items: nobody knows who wrote them or who should update them. Every topic needs an owner: price — sales or finance; delivery — logistics; technical questions — the product team; rules — operations or legal. The support lead finds the gap; the owner writes and approves the answer. The support team should not write another team's rules itself.

A monthly maintenance cycle

1. 1ListAt month end, gaps are counted by topic and type, and the 10 most repeated are selected.
2. 2Distribute to ownersEach gap goes to its owner with the evidence conversations.
3. 3Write and approveThe owner writes or corrects the answer; the date and version are recorded.
4. 4PublishThe knowledge base, the bot and agent templates are updated on the same day.
5. 5CheckNext month, the gap share in those topics is compared.

One indicator: the gap share

One simple indicator is enough to track knowledge base health: in what percentage of analysed conversations any knowledge gap appears. Track the share overall and by topic every month. If the overall share falls, the maintenance cycle is working. If it suddenly rises in one topic, something has changed there — a new product, rule or campaign — and the knowledge base has fallen behind. Do not use the indicator for individual agent evaluation: it measures the quality of the material, not people's work.

Priority: which gap first

• An outdated answer about price, money or rules — first, because there is a risk of a wrong promise.
• A contradiction — second, because customers feel treated unfairly.
• The most repeated "no answer" — third, because it takes the most time.
• Rare, complex questions — not into the knowledge base but into a rule for handing over to a specialist.

Illustrative example

This is an illustrative example. A mobile operator's support team collects knowledge gaps for a month. The most common "no answer" case is the roaming terms of a new tariff package — the package launched last week and was never added to the knowledge base. The contradiction is in the rule for transferring a number to another person: two documents list different paperwork.

The product team writes the roaming answer within two days, and the number transfer rule is merged into one document with the legal department. Next month, the gap share on both topics is compared.

Typical mistakes

• Checking the knowledge base only before launch.
• The support team writing other departments' rules itself.
• Updating the bot and agent templates at different times.
• Recording a gap as an agent's error — the problem is usually the material.

Limits

• To recognise an outdated answer AI needs current information; without it, it only infers from the customer's objection.
• Finding contradictions needs enough conversations on the same topic.
• Some questions should never go into the knowledge base — individual cases, legal advice, medical questions.

Knowledge gaps in Vexvon

In Vexvon the knowledge pieces each bot reply was built from are stored — which makes it easier to find an outdated or contradictory item. The company sets up the "knowledge gap type" and "topic" fields itself, and the panel's AI assistant gives their split and trend. Question-and-answer pairs from conversations can be exported to Excel and approved answers added to the knowledge base. More: the knowledge base and analytics.

Next step

Find 30 conversations from the last two weeks with an "I'll check and get back to you" answer and sort them by topic. The three most repeated topics are the list for your first maintenance cycle. For the lost-sales side, see unanswered customer questions; we can set up the cycle together in a demo.

Further reading on this topic: FAQ from customer questions, customer escalation analysis.