Research ·
Knowledge article usage and deflection claims in help desk reporting
A bounded review of what article views and self-service outcomes can prove, and which claims a help desk should avoid without a linked ticket record.

Key stats
article outcome groups
matching method per review
articles per pilot
Methodology and findings
Question and scope
This review asks what a knowledge article usage report can show about deflection in an outsourced help desk. The population is articles published before a declared window, and the unit is one article linked to its views, searches, and any resulting support contacts. The question matters because a high view count is often presented as proof that customers solved their own problem. The analysis looks at observable usage and ticket linkage, not at intent. A view is not the same as a resolved question.
The scope excludes internal articles that customers cannot reach and pages that are only linked from agent tools. It also excludes visits from staff testing the content during publication. Those boundaries keep the review on customer-facing self-service. A finding should help a desk decide which articles to improve, not justify a staffing cut. The measurement must separate access from outcome.
Why view counts mislead
A view can mean the customer found the answer, skimmed the page, or bounced because the title was wrong. The same count can rise after an article becomes harder to find because customers open several pages before leaving. Without a linked outcome, the number measures exposure rather than success. Reporting it as deflection overstates what the desk knows. The label matters because staffing and scope decisions may follow.
Deflection claims also ignore the customer who reads an article and then contacts support anyway. That behaviour may be healthy if the article prepared the customer for a faster conversation. It is not a failure of the article by itself. The review should look for the contact that follows a view and read the reason. A simple before-and-after count cannot distinguish preparation from frustration.
Defining usage and deflection
Define a usage event as a customer session that opens a public article from a search, a help centre link, or a support reply. Exclude internal previews, crawler traffic, and repeated reloads within the same session. Define deflection as a usage event with no related support contact within a stated period. Record the period and the matching rule. A broad definition will produce an impressive but meaningless number.
Match contacts by customer identifier, session, or a link parameter where the platform supports it. If the match is impossible, report the usage and the contact counts separately instead of claiming deflection. State the unmatched share and the reason. A transparent gap is more useful than a precise-looking estimate built on assumptions.
Counting article outcomes
For each article, record the views, the unique sessions, the search terms that led to it, and the related contacts. Add the contact reason where a link can be made. Group the reasons into answered, partially answered, and not answered. The middle group is often the most actionable because it shows where the article nearly worked. A binary success measure hides that useful detail.
Compare articles with similar topics rather than ranking every page together. A short password article and a long billing guide have different natural usage. Look for pages with many views and many follow-up contacts, because they may be creating work rather than reducing it. Also look for pages with low views and high contact volume, which may signal a discoverability problem. The pattern matters more than the absolute number.
Keep a short definition sheet beside the dashboard so every reader uses the same terms. A new analyst may otherwise treat a session, a view, and a visitor as the same event. Record the date of the last definition change and the person who approved it. When a metric is quoted in a management meeting, attach the matching period and the unmatched share. A shared definition prevents two reports from telling different stories about the same month.
Evidence available in the record
The help centre can show views, search terms, and the path through pages. The ticket system can show contacts that reference an article or arrive shortly after a session. It cannot show whether the customer read the page or understood it. The review should state that limit plainly. A view is a weak signal when it stands alone.
Check whether agents link articles in replies and whether those links are tracked. An untracked link breaks the connection between the article and the contact. Record the share of replies that include a link and the platform used. If links are pasted as plain text, the review may need a manual sample. The matching method is part of the result.
A bounded review method
Select a fixed window and a set of public articles with stable URLs. Pull the usage events and the support contacts for the same period. Match them by the strongest identifier available and record the method. Sample twenty sessions and read the related contacts where a match exists. Apply the answered, partially answered, and not answered groups consistently.
Have a second reviewer inspect a subset of the matches without seeing the first labels. Compare the outcome group and the reason for each contact. Discuss disagreements against the definitions and refine the wording. Record the disagreement rate and the unmatched share. A result that cannot be reproduced should not be used to justify a staffing change.
What the record can support
The record can support a claim that an article is widely opened but rarely followed by a contact, or that a popular page precedes many contacts. It can show which search terms lead customers to the page and which terms return no useful result. Those findings justify a rewrite, a new article, or a better title. They are practical because they connect content to a specific customer path.
The record cannot prove that the customer resolved the issue without help. It also cannot measure the value of an article that prevents a contact the customer never considered. Any deflection figure is an inference, not a direct observation. Report the definition and the matching limits beside the number. A cautious claim protects the desk from a decision it cannot defend.
Controls that improve the evidence
Add a short feedback control to every public article so customers can say whether it helped. Keep the control simple and report the response count beside the views. Track article links in agent replies with a consistent format. Review the search terms that return no results each week. These controls improve the record without requiring a new platform.
Use the findings to fix one article at a time and compare the contact pattern after the change. Look for a reduction in related contacts, not just an increase in views. Check that the article still answers the question after the rewrite. A page that ranks well but misleads customers will create more work. The measure should follow the customer outcome.
Limits of the finding
Usage data depends on analytics that may block, sample, or lose events. Customer privacy settings can hide sessions and break matching. The review should report the coverage and the excluded traffic. It should also avoid treating a missing match as proof of deflection. A gap in the data is not a success.
The finding is time-bound and sensitive to search changes, product releases, and campaign traffic. Repeat the review after a material change and keep the earlier reading for comparison. A current, bounded result is more useful than a permanent claim about self-service. The next review should use the same article set and definitions where possible.
Applying the review to one queue
Start with five public articles that support a common request type. Pull their usage and contacts for one month and classify twenty matched sessions. Read the contact reasons for the pages that precede the most work. The exercise should show whether the content answers the question or only introduces it. The result is a short list of rewrites and missing articles.
Share the result with the content owner and the queue manager. Choose one rewrite and one new article, then review the same request type after a month. Compare contacts, not only views. The aim is a useful help centre, not a higher deflection score. The next cycle should test whether customers reach the right answer with less effort.
Sources
- NIST SP 800-53 Rev. 5 security and privacy controls -
- ISO 18295-1:2017 customer contact centres -
- ISO 9001:2015 quality management systems -
- BLS Occupational Outlook Handbook: customer service representatives -