Knowledge gaps: using unanswered questions to fix your site
Unanswered questions from chat and site search show exactly what your website is missing. How to collect them, sort them into gap types and fix each one.
ASDK TeamPublished 7 min read
On this page
TL;DR
- Every question your website cannot answer is a small, free piece of user research: a visitor told you, in their own words, what they needed and did not find.
- Collect unanswered questions from three places: the chat assistant's "I don't know" answers and handoffs, site searches with no results, and questions that reach sales and support.
- Sort each gap into one of five types, missing, buried, conflicting, outdated or out of scope, and fix it at the source. Then check that the same question gets answered next week.
Why unanswered questions are worth your time
Most content plans start from what a company wants to say. Unanswered questions start from what visitors need to know. Support teams formalized this idea long ago. The Knowledge-Centered Service (KCS) methodology, maintained by the Consortium for Service Innovation, calls it the demand-driven principle: "our focus should be driven by the needs (demand) of those we serve," and content work should be "just-in-time action, not just-in-case action" (KCS).
A website assistant makes that demand visible at scale. When it answers only from your content, which is what groundingGrounding means tying an AI model's answer to specific, verifiable sources, such as your own pages or a search index, so each claim can be checked. means, every question it cannot answer points to a hole in that content. A well-built assistant says "I don't know" and offers a way to reach you instead of guessing. An assistant that guesses hides the gap behind a hallucinationAn AI hallucination is a fluent, confident answer from a language model that is false or unsupported, such as an invented price, policy or citation., which is worse for the visitor and for you.
The same goes for site search. Nielsen Norman Group notes that a "no results" page carries "a high risk of site abandonment," and that in its research people often reach one "when the content exists" but does not match their query (NN/g). Both signals tell you something different from analytics: not where visitors went, but what they wanted.
Where to collect unanswered questions
| Source | What to pull | What it tells you |
|---|---|---|
| Chat assistant | Answers where it said it did not know, handoffs to a person, thumbs-down ratings | Questions your content cannot answer or answers badly |
| Site search | Searches with zero results, and searches followed by an exit | Missing content and vocabulary mismatches |
| Sales inbox and calls | Questions asked before or during first calls | Buying questions your pages leave open |
| Support tickets | Questions from customers that public pages could answer | Help content and policy gaps |
For site search, Google Analytics 4 can record searches automatically: its enhanced measurement sends a view_search_results event when a search results page loads, detected from a query parameter such as q, s or search in the URL, and fills a search term dimension (Google Analytics Help). It does not record how many results a search returned, so zero-result searches need your search tool's own report or a custom parameter. Our article on site search analytics covers that setup.
The five types of knowledge gap
Before writing anything, sort each question. The fix depends on the type.
| Gap type | What it looks like | Typical fix |
|---|---|---|
| Missing | The answer is nowhere on the site | Add it to the page where it belongs |
| Buried | The answer exists, but in words visitors don't use or deep in a PDF | Use the visitors' wording; move the answer to an HTML page |
| Conflicting | Two pages give different answers | Decide the true answer; fix or remove the other |
| Outdated | The page is wrong since a change in pricing, product or policy | Update it and anything that copies it |
| Out of scope | The question should not be answered on the site, such as individual legal advice | Write a clear boundary and a route to a person |
Two more categories come up and are not content problems:
- Undecided policy. "Do you offer non-profit discounts?" may have no answer because nobody has decided. Escalate to whoever owns the decision; do not let the assistant improvise one.
- Noise. Spam, tests and off-topic chatter. Discard, but watch for patterns such as bots probing the assistant.
A weekly loop that takes an hour
Collect
Export last week's unanswered chat questions, handoffs and zero-result searches into one list. Keep the visitor's exact words; they are the most useful part.
Cluster
Group questions that ask the same thing in different words: "Do you work with Xero?", "Xero integration?" and "can it sync invoices to our accounting software" belong together. Count each cluster.
Classify
Give each cluster a gap type from the table above.
Prioritize
Rank by how often it was asked and how much it matters. A question asked twice by buyers on the pricing page outranks one asked ten times by job seekers.
Fix at the source
Change the page, not the assistant. Your knowledge baseA knowledge base is the curated set of content an AI agent is allowed to answer from, such as web pages, help articles, PDFs and FAQs, kept indexed and current. is built from your content, so a fix to the page fixes the website, the assistant and what AI search tools read about you at once.
Verify
After the site is re-crawled, ask the original questions again in the visitors' words. Close the cluster only when the answer is correct and cites the right page.
KCS has a phrase for the moment a search fails: its Capture practice includes a technique called "Searching is Creating," because a failed search is where a gap is first noticed (KCS Practices Guide). The loop above applies that to your website.
How to write the fix
- Answer on the page where the question arises. A pricing question belongs on the pricing page, not in a new blog post.
- Use the visitors' words alongside your own. If visitors say "sync invoices" and your page says "accounting connector," include both. This is search intentSearch intent is the goal behind a query, such as learning something, comparing options, finding a site or taking an action, which shapes what a good result is. at the smallest scale.
- One fact, one home. State a fact once, in one place, and link to it elsewhere. Facts copied across pages go out of date one copy at a time.
- Write the "no" answers. "We don't offer on-premise installation" is a complete, useful answer. A missing "no" is one of the most common gaps.
- Prefer HTML to PDFs for anything visitors or AI tools need to read.
- Do not mass-produce pages for every phrasing of a question. Google's guide to its AI features warns that producing pages for every query variation mainly to manipulate results "violates Google's scaled content abuse spam policy" (Google Search Central).
An illustrative example
A fictional B2B software company reviews one week of unanswered questions:
| Cluster (visitors' words) | Count | Gap type | Fix |
|---|---|---|---|
| "Do you integrate with Xero?" | 9 | Buried | The integrations page listed "accounting systems"; add named tools |
| "Is there a discount for annual billing?" | 6 | Missing | Add the annual discount to the pricing page and plan table |
| "Where is our data stored?" | 4 | Conflicting | Security page said EU, an old FAQ said US; fix the FAQ |
| "Can I get a quote for 300 users?" | 3 | Missing | Add how pricing works above the largest plan, with a contact route |
| "Can you review our contract?" | 2 | Out of scope | Add a boundary line and a route to the legal team |
Five fixes, most of them one or two sentences, close the gaps behind 24 unanswered questions. Next week the list starts again, usually shorter.
Measure whether it works
- Unanswered rate: share of conversations with at least one question the assistant could not answer. It should fall over the first weeks, then settle.
- Repeat clusters: questions that come back after you closed them mean the fix did not land or the page is hard to find.
- Handoffs for routine questions: if people ask for a human to get answers that should be on the site, that is a gap too.
- Zero-result search rate, if you run site search.
Do not aim for zero unanswered questions. Some questions should go to a person, and new ones appear whenever your product or market changes. The goal is that the same gap never stays open for long.
Who should own the loop
Knowledge gaps cross team lines. A missing integration answer belongs to product marketing, a conflicting data-location statement to whoever owns security and legal pages, and an undecided discount policy to the person who sets pricing. If nobody owns the list as a whole, every gap becomes someone else's problem.
Give one person the weekly review and the authority to assign fixes, with a short service level: buried and conflicting gaps fixed within a week, missing answers within two, policy questions decided or explicitly parked within a month. On small teams this is often the marketer who runs the website. On larger ones, support operations already does something similar for help content and can extend it to the public site.
Common mistakes
- Fixing the assistant instead of the page. Adding a hand-written answer to the bot hides the gap on the website, where buyers and AI search tools still hit it.
- Reviewing only the chat. Sales and support questions are often the most valuable, because they come from people who were already close to buying.
- Treating every question as a content request. Out-of-scope and policy questions need a boundary or a decision, not another paragraph.
- Closing a cluster without re-testing it in the visitors' own words.
Related work
Unanswered questions overlap with two other content tasks. Turning common questions into new articles for search is covered in turning visitor questions into content ideas. Keeping the assistant correct as pages change is covered in keeping a website AI assistant accurate. And if you are setting up an assistant for the first time, start with training an AI chatbot on your website.
In ASDK, the Chat agent, in early access, says when the knowledge base has no answer and offers a way to contact you. The dashboard lists the questions it could not answer on every plan, and the "What visitors ask" insights report, with top questions and most-cited pages, is an early-access feature included from the Pro plan; see pricing for plan details.
Try the Chat agent in early access
Answers visitors from your own content, 24/7, and captures contacts.
Unanswered questions are the cheapest content research you will ever get. Collect them every week, sort them before you write, fix them where they belong and check the fix with the visitors' own words. Over a few months, your site will answer what buyers actually ask, for people, for your assistant and for every AI tool that reads it.
Frequently asked questions
What is a knowledge gap on a website?
A question visitors ask that your site cannot answer, answers in words they do not use, answers in two conflicting ways, or answers on a page nobody finds. Each type has a different fix, so sort gaps before you write anything.
Where do I find the questions my website does not answer?
In chat conversations where the assistant said it did not know or handed over to a person, in site searches with zero results, and in the questions that reach sales and support by email. Together they show what visitors looked for and did not find.
How often should I review unanswered questions?
Weekly for the first months after launching a chat assistant, when gaps are most visible, then every two weeks. A short regular review beats a large quarterly project, because fixes reach visitors sooner and the list stays manageable.
Should every unanswered question get a new page?
No. Most gaps are best fixed by adding a sentence or a section to an existing page, such as pricing, a product page or the FAQ. Create a new page only when a topic is big enough to deserve one, and never mass-produce thin pages.
Sources
Pages we read for this article, with the date we checked each one.
- Demand Driven — KCS Principles and Core Concepts, Consortium for Service Innovation — accessed October 21, 2026
- Practice 1: Capture — KCS v6 Practices Guide, Consortium for Service Innovation — accessed October 21, 2026
- 3 guidelines for search engine 'no results' pages — Nielsen Norman Group (January 2014) — accessed October 21, 2026
- Enhanced measurement events — Google Analytics Help — accessed October 21, 2026
- Optimizing your website for generative AI features on Google Search — Google Search Central (updated July 2026) — accessed October 21, 2026