Knowledge Base
Evaluating search quality with real customer questions
Build a customer-question set, check the right search surfaces and diagnose poor results before changing knowledge base content.
Test knowledge base search with the questions customers ask, in their own words. A useful result lets the intended reader find an approved answer and take the right next step. Returning an article alone is not enough.
Build a question set
Collect examples from recent enquiries and, where available, help-centre search terms. Remove names, order numbers and other identifying details. Keep customer wording: an article titled “Change delivery address” may need to answer “sent it to my old place”. Include alternate terms, common misspellings, a broad question, a question with an important condition and one with no approved answer.
For each question, record the expected article and the condition it must explain. If the correct outcome is a handoff or an honest statement that no approved answer exists, record that instead. An article created solely to make a search return something would hide the gap.
| Record | Why it matters |
|---|---|
| Customer wording and entry point | Search surfaces may behave differently. |
| Audience and expected article | Restricted content must reach only the right reader. |
| Expected answer or handoff | A plausible but incomplete result should not pass. |
| Result position and title | Readers need to recognise the useful result. |
| Next action after opening | The article must resolve the question or direct the reader appropriately. |
Search where readers encounter the content
Check the public help centre, embedded widget and agent view separately where each is used. For customer routes, compare signed-out and relevant signed-in views. Record visible results, open the likely article and assess whether its instructions answer the question without losing a necessary condition.
Zendesk's instant Help Center suggestions match article titles, not body text or labels. After a user presses Enter, Zendesk's native search includes titles, body text and labels. A body edit alone may leave a weak instant-search title unresolved.
Help Scout Docs lets editors add keywords for alternate terms and misspellings. Its documented relevance order puts exact keyword and title matches ahead of body matches.
Search behaviour across platforms: Zendesk vs Help Scout
- Zendesk Instant Suggestions
- Matches article titles only
- Zendesk Native Search (after Enter)
- Includes titles, body text and labels
- Help Scout Keywords
- Editors can add alternate terms and misspellings
- Help Scout Relevance Order
- Exact keyword and title matches > body matches
Diagnose the failure
A no-result search can mean the answer is missing, restricted or described in different words. A wrong-result search may point to unclear titles or similar articles that need distinction. A found-but-unhelpful article may be outdated, omit a condition or fail to explain the next action. Identify the cause before adding keywords or rewriting content.
Reports help choose questions to inspect, but clicks and result counts do not prove a customer completed the task. Zendesk's Search dashboard includes queries, searches with no results and click-through measures, with role and locale filters. Intercom's Articles report and Help Scout's Docs report identify searches with no results. Review the actual answer and reader journey before deciding on a fix.
Key metrics from search reports
- Zendesk Search Dashboard: Queries with no results
- Available with role and locale filters
- Intercom Articles Report: Searches with no results
- Identifies missing answers
- Help Scout Docs Report: Searches with no results
- Tracks unmet queries
- Click-through rates
- Do not confirm task completion
Diagnosing poor search results
- No-result searchAnswer missing, restricted, or described in different words
- Wrong-result searchUnclear titles or similar articles need distinction
- Found-but-unhelpful articleOutdated, omits condition, or fails to explain next action
- Review actual answer and reader journeyBefore deciding on a fix
Fix and retest
Give each failed question to an article owner. Improve the title or keywords when the right article exists but is hard to find. Revise an incomplete answer. Add an article only when there is an approved answer worth publishing; otherwise, make the handoff clear.
Retest the original wording and a related variation in the same search surface and audience. Keep the question set for later content changes.
