A support team builds a knowledge base, drafts many articles with an AI assistant to save time, and publishes them. Months later, agents and customers report that search is unreliable. Typing a phrase from an article returns nothing.
The temptation is to blame the search tool. Sometimes the cause is in the content itself, in characters that look like ordinary text but do not match what a person types.
How Search Matches Text
Search indexes break text into tokens and match the tokens a user types against those in the index. A hidden character in the middle of a word can split it into two tokens or change it into a string that matches nothing a person would ever type. To the reader, the word looks perfectly ordinary.
The effect is selective and therefore confusing. Most of an article matches, but one key phrase does not, and that is often the phrase a customer is most likely to search for.
Where Contamination Enters a Knowledge Base
- Articles drafted in a chat tool and pasted into the editor.
- Product names and error codes copied from generated tables.
- FAQ answers imported in bulk from an AI-assisted spreadsheet.
- Macros and canned replies that agents copy across tickets and articles.
Why Error Codes Are Especially Vulnerable
Customers search for exact strings, such as an error code or a button label. Those strings are short and precise, which makes them the most sensitive to a single stray character. An article that documents Error 4021 may be unfindable by anyone who types exactly that.
The Cost to the Support Operation
When search fails, customers open tickets for questions the knowledge base already answers. Every deflected ticket that does not get deflected is a direct cost. Teams that track ticket deflection often see an unexplained drop and look for causes in navigation or design before looking at the text itself.
A knowledge base is only as useful as its findability, and findability depends on text being exactly what it appears to be.
A Practical Cleanup Approach
For existing articles and new drafts alike, running text through this tool removes invisible characters and normalizes spacing before it is saved, so search behaves the way editors expect.
Teams with a large backlog can prioritize. Start with the twenty most-viewed articles, then the ones with the most ticket references, then work outward. That sequence captures most of the benefit in the first few hours.
Verifying the Fix
After cleanup, test the search with exact phrases from the repaired articles, including error codes and product names. Compare results against the previous behavior. Teams that keep a short list of known test queries can rerun it after every large import.
It is also worth checking the analytics for searches that returned no results. A cluster of zero-result queries that match article titles is a strong sign that contamination is still present.
Embedding the Step in Authoring Guidelines
Authoring guidelines should state where cleanup happens. For most teams, the best point is just before an article is saved in the knowledge base editor. A line in the style guide and a reminder in the template are usually enough.
Support teams that use several writing tools together can find them inside Phrasly AI, which keeps drafting, checking, and cleanup in one place for authors who publish regularly.
Keeping Macros and Canned Replies Clean
Agents copy macros across tickets, and tickets sometimes feed back into articles. A hidden character in a macro therefore spreads through channels. Cleaning the macro library once and instructing agents to paste from the library rather than from old tickets prevents recontamination.
- Clean the macro library and mark it as the source of truth.
- Discourage copying text from old tickets into new articles.
- Review top-used macros for stiffness at the same time.
A Healthy Search Feedback Loop
Zero-result searches are one of the most useful signals a support team has. Reviewing them weekly reveals both missing content and findability problems. When a zero-result query matches an existing article title, the cause is usually something in the text, not a gap in the knowledge base.
Teams that review this list regularly tend to catch contamination early, and they spend less time debating whether search needs replacing.
Measuring Findability
Support teams can measure the health of their knowledge base with a few simple numbers. The proportion of searches that return results, the click-through rate on the first result, and the share of tickets that cite an article all say something about whether customers can find what they need. A sudden drop in any of them is worth investigating, and text quality is one of the candidates.
It helps to run a standard set of test queries each month, using the exact phrases customers use. Keeping the list stable allows comparison over time, and any query that suddenly fails to find a known article points directly to the page to inspect.
Training Agents to Report Search Problems
Agents are often the first to notice that search is behaving strangely, because they use it all day. A simple way to capture their observations, such as a shared channel or form, turns anecdotes into data. Tagging each report with the phrase searched and the article expected makes it easy to find a pattern.
- Record the exact phrase typed and the article expected.
- Note whether the article is found by title but not by body text.
- Share the report with whoever owns the knowledge base.
- Recheck after any cleanup to confirm the fix.
A Practical Prioritization Plan
Teams with hundreds of articles cannot fix everything at once, and they should not try. Rank articles by views, by the number of tickets that reference them, and by how recently they were created from AI-assisted drafts. Clean the top group first, then work down the list as time permits.
After each batch, rerun the standard test searches and note improvements. Seeing searches succeed that previously failed is motivating, and it builds the case for keeping the process in place.
- Start with the most viewed articles.
- Include articles that contain error codes and product names.
- Retest search after every batch.
What Customers Gain
Customers benefit directly from findable articles. They resolve problems faster, avoid waiting for a reply, and form a more positive impression of the company. For support teams, the benefit is fewer repetitive tickets and more time for the cases that genuinely need a person.
A knowledge base that works is one of the most cost-effective parts of a support operation. Protecting it from a subtle technical problem is a small investment with a visible return, and it keeps the content that the team worked hard to write in front of the people who need it.
Support leaders who add text cleanup to their authoring routine often find that their search analytics improve within weeks, which is a rewarding and easily measured result.
A knowledge base succeeds when customers find answers without opening a ticket. Cleaning the text before it is saved removes a hidden obstacle to that goal and keeps search aligned with what readers actually type.
You may also read
