The basic rules

• Search is case-insensitive.
• Query terms are separated by spaces.
• Every term must be a substring of a word in an item’s title.
• All terms must match for a result to appear.
• Terms that match in the order you typed them score higher than unordered matches.

Every match is scored based on factors that include match length, position and order.

Results are sorted best-first and non-matches are taken out.

For a query git push:

Why a word must fully match

For the items above gitpush (without spaces) won’t return results.

Requiring each term to be a substring of a word keeps results precise: you can type fragments (gi pu) and still match “Github Push Command”, but you never get noise that shares just a few letters from other fuzzy algorithms (e.g. “Github Pages Usage”).

The trade-off is that a typo returns nothing — retype and the score model does the rest.

Optimization

Giving your bookmarks, bookmarklets and scripts good titles will go a long way in making the search get you better results.

Bookmarks typically carry the page title, it’s often short and it might not have the words to match your mental reference to that page.

So take your time to curate them and give them better titles if you’re having too many clashes.

Smart ranking (recency + frequency)

Search relevance is only half the story. Every item also carries a usage score that decides the order of otherwise-equal results:

• Exponential decay — an item loses half its score every 30 days. Recent use keeps an item prominent; old use fades.
• Frequency bonus — items used often get a persistent, logarithmically scaled boost, so a frequently-used item can outrank a more recent but rarely used one.
• Auto-cleanup — entries older than 120 days are pruned, and the store caps at 50 items, so the ranking stays fast.