Fixing a Search That Returned Nothing

Around a fifth of searches on a product documentation site returned no results, and the team's proposed fix was to replace the search engine.

A technical documentation site with several hundred pages and an exact-match search. Content was written by the engineering teams that built each feature.

• No budget for a hosted search product in the current year.
• Content was owned by many teams and could not be centrally rewritten.
• The site was statically generated, so any solution had to work with a prebuilt index.

Keep the existing search and add three things: an alias field on every page, fuzzy matching with a bounded edit distance, and a no-results state that offers the closest matches and the most popular pages.

The zero-result log showed the failures were not retrieval failures. Users searched for the words they used, and the documentation used the words the engineers used. An alias list mapping the two, maintained by whoever writes the page, solved most of it. Fuzzy matching handled typos. The improved empty state handled the remainder without pretending the search had succeeded.

• The zero-result query log, grouped by intent, which showed clusters of synonyms for the same concept.
• Comparing the top zero-result queries against support ticket wording, which matched closely.
• A count of how many failures were typos versus vocabulary, which came out heavily in favour of vocabulary.

• Aliases require maintenance, and pages whose authors do not add them stay hard to find.
• Fuzzy matching introduces some irrelevant results, which required a relevance threshold.
• The team did not get the better search engine they wanted, and the underlying engine remained basic.

Zero-result searches fell substantially, and the remaining ones became a useful backlog of genuinely missing content. The alias field also improved external search visibility, which was not the goal.