Site search fails code-like queries (symbols, code tokens, stopwords, typos) #1972

What information was incorrect, unhelpful, or incomplete?

Site search (/en-US/search) returns no results, or buries the canonical page, for queries that developers type as code:

• Symbols and operators return zero results, even where the page title contains the symbol: % ("Remainder (%)"), !, ?, ~, *, @, #, !!, ${}, ::. The index strips punctuation, so these queries have no tokens left to match.
• Code tokens in prose don't help ranking: max-age doesn't surface Cache-Control, and ::before, :has() or @media rank their reference pages low or not at all.
• Exact API names that are stopwords return zero results: a (the <a> element) and then (Promise.prototype.then()).
• Typos, plurals and compound words return zero results: adress, lenght, localcompare, eventlisteners, webworker.

In site search measurements from 4 weeks in September 2026, 22 of the 39 most frequent zero-result queries (about half of those searches) have a matching MDN page. Most of the rest are topics MDN doesn't cover, where zero results is correct.

What did you expect to see?

• Symbol and operator queries find the operator, selector or syntax page (e.g. % → Remainder, ${} → Template literals, :: → pseudo-elements).
• Code tokens rank the page that documents them first (e.g. max-age → Cache-Control).
• Exact API names match, even when they are common words.
• Near-miss spellings still find the intended page when the exact query has no results.
• No regressions for plain-word queries (e.g. html, fetch and flexbox keep their current top results).

Do you have any supporting links, references, or citations?

Related to:

• Searching for max-age to find information about cache-control doesn't work well #1577
• Search prioritizes article title too much #621
• Helping users with !, $, *. and certain other characters not indexed in the help system. dex#93
• Unable to search for :: dex#94

Do you have anything more you want to share?

A fix needs changes in both Dex (index fields) and Rumba (query):

• Index inline code, and add title fields that keep punctuation.
• Apply the inline-code and symbol matching only to code-like queries, so plain-word queries are unaffected.
• Add a fallback query (fuzzy matching, exact API names) that runs only when the main query returns no results.

In a local evaluation against the same measurements, this fixes 17 of the 22 queries above, with no measurable change to how high clicked results rank for other queries. The as-is version of mdn/rumba#885 regresses plain-word queries significantly, which is why the matching needs to depend on the query's shape.

The quick search in the header uses a separate client-side title index and is out of scope here.