Research · Published October 27, 2026

Testing a Knowledge-Base Deflection Claim Before Repeating It

A protocol for checking whether self-service actually prevented tickets, rather than attributing every drop in volume to the library.

Key stats

deflection signals

comparison groups

Methodology and findings

Research question

Did a knowledge-base article prevent support contacts, or did contact volume fall for another reason? The unit is a request family over a defined period, with a comparison between a period before and after an article change and, where possible, a similar family that did not receive the change. The study tests a local claim; it does not produce a universal deflection rate.

Separate deflection from avoidance

A customer who reads an article and solves the problem is deflected. A customer who gives up, asks a colleague, or uses a different channel is not. A drop in ticket volume can therefore hide a worse experience. The study looks for positive signals: article views followed by no contact, a self-service completion event, or a customer statement that the article helped.

Pick the comparison groups

Use three groups where possible: the request family that received the new article, a similar family without the change, and the same family in the prior period. The second group controls for a general change in demand; the third controls for seasonality. If a comparison group is not available, say so and treat the before-and-after result as suggestive rather than conclusive.

Define the four signals

Record article views, unique users, self-service completions, and contacts that reference the article. A contact that references the article can mean the article helped someone get close, or that it failed. Read a sample of those contacts to classify them. Do not treat a view as a success; a view is exposure, not resolution.

Account for the launch effect

Announcements, search ranking changes, and a link from the portal can raise views without changing the underlying need. Record every distribution change during the period. If the article was promoted in a newsletter, the volume change may reflect the promotion rather than the article's quality. Keep a simple change log alongside the data.

Sample contacts for cause

For each period, sample contacts in the request family and classify whether the customer had seen the article, whether they tried the steps, and where they stopped. This qualitative layer explains the numbers. A high view count with many contacts that stopped at the same step points to a prerequisite or a confusing instruction, not to a demand problem.

Calculate with visible denominators

Report contacts per thousand eligible users or per thousand sessions, not only the raw count. Show the article views and the contact volume together. A decline in contacts alongside a decline in eligible users is not evidence of deflection. Include the observation window and any known outages or product changes.

Test one revision at a time

If the article changes, keep the old version and the date. Compare the four weeks before and after, and note any concurrent change. Do not rewrite the article, move it, and promote it in the same week and then attribute the result to the writing. One change per cycle gives a cleaner reading, even if it takes longer.

Report the claim with its limits

State what the evidence supports: for example, contacts in this family fell after the article was linked from the portal, while a comparable family was flat. Then state what it does not support: that the article would work for another product, another customer group, or another channel. Deflection claims are local and perishable.

Limitations

Customers may solve a problem without any tracked signal, and they may contact support without mentioning the article. Search and portal analytics are incomplete. A product change can dominate the result. The protocol is designed to make the claim honest, not to prove causation from a single before-and-after comparison.

Test one article family

Choose one request family with a stable volume, record a four-week baseline, then change one article element and compare with a similar family.

Plan a deflection test

Sources

U.S. GAO, Assessing Data Reliability - Guidance for documenting data sources and uncertainty. Checked October 27, 2026.
Atlassian, Jira Service Management Documentation - Request channel and reporting concepts. Checked October 27, 2026.
NIST SP 800-53 Rev. 5, Security and Privacy Controls - Access and record-keeping context for self-service content. Checked October 27, 2026.