Knowledge gaps: how to find the questions you cannot answer

Anyone running customer service at an online store with a few hundred tickets a week usually knows which questions are most common. Fewer know which questions the company has no good answer to, because those questions do not appear in any report. They appear as long response times, different answers from different agents, and customers who email a second time.

What is a knowledge gap?

A knowledge gap is a question customers ask that lacks a published, reviewed answer every channel can use. It is a broader concept than "article missing", and there are three kinds:

1. The question has no answer at all. No article, no template, no page. Every agent writes the answer from scratch, and the chatbot cannot reply.
2. Two sources say different things. The help centre says 30-day returns, the email template says 14, and a campaign page says 60. All three are published, and the customer gets whichever answer comes up. That is exactly what happened in Moffatt v. Air Canada in 2024: the airline's chatbot said one thing about bereavement fares, another page on the same website said the opposite, and the tribunal held the company to the bot's statement.
3. The answer exists only in one employee's head. Anna knows how to handle partial deliveries from one of the warehouses. When Anna is on holiday, nobody knows. This kind is hardest to see, because the tickets still get resolved, just by a single person.

The cost of gaps is time. According to the McKinsey Global Institute, office workers spend on average around 1.8 hours a day searching for and gathering information. The figure is from 2012 and covers knowledge workers in general, but the pattern is familiar from every customer service inbox.

Where do you find the knowledge gaps?

The gaps are in four data sources you already have: the ticket history, the help centre search logs, the chatbot's unanswered questions and the agents' free-text replies.

Some practical tips per source:

• The ticket history gives the most if you tag tickets by question type, even roughly; twenty categories are enough.
• The search logs: zero hits is the clearest signal, but searches that return a hit and are still followed by a ticket are at least as important: the article exists but does not answer the question.
• The chatbot's handovers are the cleanest source, because a grounded bot by definition hands over when the material is missing. How such a bot works is described in the article on grounded AI in customer service.
• The free-text replies you find by pulling the last hundred outgoing emails and reading them.

A Gartner survey from 2024 showed that only 14 percent of customer service issues are fully resolved in self-service, and only 36 percent of the issues customers themselves describe as very simple. The gap in between is largely a list of knowledge gaps.

How do you prioritise which gaps to close first?

Prioritise by volume times consequence: how often the question is asked, multiplied by what it costs when it is answered wrongly or late. A question asked ten times a day that only costs a longer email is less urgent than a question asked once a week that ends in a complaint, a lost customer or a broken price promise.

A simple scale is enough:

• Volume: 1 for a few times a month, 2 for a few times a week, 3 for daily.
• Consequence: 1 for irritation and extra back-and-forth, 2 for a return, compensation or lost sale, 3 for legal or financial liability, for example terms, prices, warranties and personal data.

Multiply and sort. Questions scoring 6 or 9 get written this week. Questions scoring 1 or 2 can wait. Contradictory answers almost always score high, because the customer stops trusting both channels.

How does a gap become a reviewed article?

A gap becomes an article when someone owns the question, writes the answer from a real ticket, has it reviewed by whoever knows the subject and publishes it where every channel gets its answers. The Knowledge-Centered Service method, KCS, described in the Consortium for Service Innovation's practices guide, is built on exactly that: knowledge is captured in the workflow as the ticket is resolved, not in a separate documentation project afterwards, and an article only needs to be sufficient to solve the problem before demand decides whether it should be improved.

This is what the flow looks like in practice:

1. Capture the question in the customer's words. The heading should be what the customer searches for, not your internal term. "Can I change the delivery address after the order has shipped?" is better than "Address change, outbound order".
2. Appoint an owner. One person, by name, responsible for the article being correct. If the answer is Anna's, Anna is the owner, and writing it down is how the answer leaves her head.
3. Write the draft from a real ticket. The best free-text reply you have already sent is often a finished first draft. AI can turn three emails into a draft, but the review is a person's job; see the article on AI drafts with review.
4. Have a subject expert review it. Whoever knows the warehouse, the finances or the terms reads and approves. With contradictory answers, this is where you decide which one applies and remove the others.
5. Publish in one place every channel uses. Help centre, chatbot and inbox should fetch the same article; otherwise you have created a fourth version instead of removing three.
6. Link the article from the ticket that triggered it. Then you can see afterwards which tickets the article saved.

The work is not free, but it is work customer service already does today, just without saving the result. Gartner reports in a survey from 2026 that 58 percent of customer service leaders plan to upskill agents into knowledge specialists, because AI and self-service need correct, current content.

How do you follow up that the gap is closed?

A gap is closed when the question stops arriving as free text, and you see that in the same four sources that revealed it. Follow up four weeks after publication:

• The ticket history: the number of tickets in the category should fall, and those remaining should have the article linked in the reply.
• The search logs: searches for the question should return a hit, and the share followed by a new ticket should fall.
• The chatbot: the question should disappear from the handover list, and the bot should cite the article as its source.
• The free-text replies: agents should insert the article instead of rewriting it.

If none of this falls, the article is either hard to find, hard to understand or wrong. Also set a review date on every article; terms, prices and delivery times change, and an article that was right in September is a new gap in December. Which metrics capture the effect is covered in the article on customer service metrics.

What to do

One week is enough for a first round:

1. Monday: pull the last 200 tickets and tag each with a question type. Count per type how many were answered with a linked article and how many with free text.
2. Tuesday: go through the help centre searches for the last 30 days. List search terms with zero hits and check whether the ten most common terms get an answer in the first result.
3. Wednesday: read the last 100 outgoing emails. Mark every reply longer than five sentences that does not link to an article. Also mark questions where two agents answered differently.
4. Thursday: score. Volume times consequence for every gap. Pick the five with the highest score.
5. Friday: write the first one. Appoint an owner, draft from the best existing reply, have an expert review, publish, link from the tickets. Schedule the other four for next week.

If you want a ready-made basis, you can let us read your ticket history and website and get a list of questions without answers; see the page on knowledge analysis. The method above works just as well without it.

Common questions

How many knowledge gaps are normal?

There is no documented normal value, and the number depends on how broad the range and how many policies the company has. The gap between what customers try to solve themselves and what they succeed with is wide in most companies, and the first review almost always finds more gaps than the team expected. Plan to prioritise hard rather than close everything.

What do we do when two departments give different answers?

Decide who owns the question and let that person decide which answer applies. Often both departments are right but about different situations, for example returns in store versus returns online, and then the article should describe both. Then remove or rewrite every other place the answer appears. Contradictory published answers are the gap that damages trust most, because the customer can show you both.

Do we need a new tool to find the gaps?

No. The ticket history is in your ticketing system or inbox, search logs are in most help centre tools and in web analytics, and the free-text replies you read by hand. A tool makes the work faster and repeatable, but the first review is best done manually, because that is when the team learns to recognise the patterns. Start with one week and a list in a spreadsheet.

How often should we look for new gaps?

Continuously for the chatbot's handovers and searches with no hit, which should be reviewed weekly by whoever owns the knowledge base. A larger review of ticket history and free-text replies is enough quarterly, and always ahead of high-volume periods, such as the Christmas season or a product launch, when new questions appear in bulk and response times are already under pressure.

Sources

• Moffatt v. Air Canada, 2024 BCCRT 149 — Civil Resolution Tribunal of British Columbia via CanLII, 2024
• The social economy: Unlocking value and productivity through social technologies — McKinsey Global Institute, 2012
• Gartner Survey Finds Only 14% of Customer Service Issues Are Fully Resolved in Self-Service — Gartner, 2024
• KCS v6 Practices Guide — Consortium for Service Innovation, 2016
• Gartner Survey Finds 91% of Customer Service Leaders Under Pressure to Implement AI in 2026 — Gartner, 2026