Knowledge base content health: the checks that matter

What does a healthy knowledge base look like?

Content health is about the articles you already have. A gap analysis asks whether customers' questions have answers. A health check asks whether the answers you have are any good. Both matter, and they fail in different ways.

A healthy article passes five plain tests:

• It is true. It matches current policy, pricing and process.
• It is readable. An agent mid-call, or a customer on a phone, can find the answer in seconds.
• It sounds like you. It uses your terms, your tone and your rules about what can and can't be said.
• It does one job. It answers one question or covers one process, so it can be found, trusted and kept current.
• It agrees with everything else. No other article overlaps it or gives a different answer.

A healthy knowledge base is one where those tests hold across the whole collection and keep holding as products and policies change. That second part is the hard one. Content decays quietly: a price changes in one article and not another, edits pile up in different styles, a short article grows into a long one.

What are the six content health checks?

The first three look at one article at a time and can propose a fix in the same pass. The next two compare articles with each other. Complexity looks at scope. They are designed to work together: a complexity split often clears readability findings that kept coming back, and a merge from similarity removes a future contradiction.

How do brand alignment, readability and validity work?

Brand alignment checks an article against a brand profile you define: the narrative, language and country, tone descriptors and approved terminology. It finds drift, which happens even with one brand as terms change and different authors edit. The fix is a rewrite in your voice that keeps the article's headings and sections where they are. If you run several brands, pick the right profile for each run.

Readability checks whether an article suits the people who read it. You set the audience and target level in plain words, for example "general customer service agents" at a Grade 8 to 10 reading level. Standard measures such as Flesch Reading Ease, Flesch-Kincaid Grade Level and the Gunning Fog Index help show where text is dense. The fix shortens and clarifies the wording inside the existing sections without dropping meaning, and you can compare the change line by line before accepting it.

Validity is the check that stops agents, and AI answers, repeating wrong ones; a source that is faithfully repeated but out of date is still a wrong answer. It compares the claims in an article with your authoritative source documents (policies, product disclosure statements, procedures) and flags any that conflict, citing the source. It is only as good as those sources. With none loaded, it finds nothing; with stale ones, it raises false alarms. Treat each finding as a question: the article and the source disagree, which is right? Sometimes it is the source that needs updating.

How do similarity, contradictions and complexity work?

Similarity compares each article with every other live article in its knowledge area, including drafts and articles in review, and flags near-duplicates. For a true duplicate, the fix is to merge them into one article. For overlap that is deliberate, such as separate articles for different regions, dismiss the finding with that reason.

Contradictions looks for articles that answer the same question differently. It extracts the key claims from each article, screens pairs for possible conflict, and runs a full comparison on the pairs that might disagree. Each finding shows the two statements and the reasoning, so a person can decide which is correct. It compares articles with each other; validity compares articles with sources. Both are needed. Our guide to duplicate and contradicting articles covers these two checks and how to resolve what they find.

Complexity asks whether one article is trying to resolve several distinct issues. If it is, it proposes a split plan: a list of focused articles, each with a title and purpose. You can edit the plan, merge sections that are too granular, or dismiss it. Accepting creates new draft articles and marks the original superseded. Long documents such as policies and contracts are meant to be long, so they are only split when their sections cover unrelated subjects.

What is template drift?

Most knowledge teams have templates: a standard set of sections such as summary, eligibility, steps and exceptions. Template drift is the gap between those templates and the articles that are meant to follow them. It builds up three ways:

• The template changes and older articles keep the old shape.
• Articles were written before the template existed, or imported from documents with their own layout.
• Sections go missing as articles are edited under pressure.

Drift matters because structure is what makes an article scannable for agents and predictable for AI answers, which retrieve passages rather than whole pages; see why AI search won't fix a legacy knowledge base. If "exceptions" is always the fourth section, an agent learns where to look.

Knowledge Quest works out each article's alignment from its content and marks it as aligned, partly aligned, not aligned, or behind an updated template. Restructure to template then moves the article's existing content into the template's sections in the right order. It doesn't invent anything. A required section with nothing to put in it becomes a "to be completed" placeholder for the owner, content that fits no section is kept under "Additional information", and a guard holds back automatic changes if any content looks like it might be lost. Where sections are empty, fill gaps can draft them from your authoritative sources, with numbered citations.

Can fixes be one click and still safe?

Yes, if the one click belongs to a person who has seen the change. Each finding that carries a rewrite shows the proposed article next to the current one. The reviewer can accept it, edit it and then accept, or dismiss it with a reason. That is the one click. It replaces the slow part, which is drafting the corrected text, while keeping the judgement.

A few safeguards make this hold up in a busy team:

• Stale suggestions are refused. If the article changed after the suggestion was made, accepting it is blocked and the check needs to run again, so a reviewer can't overwrite someone else's edit.
• One open suggestion at a time. An article with an unreviewed rewrite isn't sent through another rewrite until the first is accepted or dismissed.
• Judgement calls stay manual. Merges, contradictions and splits always need a person. There is no automatic version of "which answer is right".
• Auto-apply is opt-in. For brand alignment, readability, validity and restructuring, you can choose to apply rewrites automatically for a run. It is off unless someone turns it on.
• Everything is recorded. Each applied change appears in the article's history, so you can see what changed, when and why.

How often should you run each check?

Run checks on change, not on the calendar alone. An article that hasn't changed and passed last time doesn't need checking again, and re-running it only creates noise. A practical rhythm:

• On every new or edited article: readability and brand alignment, so new content starts healthy.
• When source documents change: validity across the affected knowledge area. A new policy or price list is the commonest cause of wrong answers.
• Before publishing new articles: similarity and contradictions, which compare drafts too, so a new article doesn't duplicate or contradict an old one.
• Monthly or after a migration: complexity, and restructuring for articles behind their template.

Edits make earlier results stale. A check marked "re-check recommended" isn't a fault; it means the article changed since anyone looked.

How does content health relate to knowledge gaps?

They are two halves of one job. Gap detection starts from what customers asked and finds what is missing. Health checks start from what you have written and find what is wrong. A stale article and a contradiction are both gaps from the customer's point of view, which is why the knowledge gap analysis guide counts them among the four kinds of gap.

Use the conversations to decide where health work matters most. A readability finding on an article that answers thousands of contacts a month comes before one on an article nobody uses. Knowledge Quest runs all six checks over the knowledge base you already have, alongside gap detection from your conversations, and sends approved changes to your knowledge base by live sync or export. Your knowledge base stays; its content gets healthier.