Progress Update

Slick Ranking Progress Update - Friday May 15 2026

Slick Ranking Progress Update - Friday May 15 2026

What Slick Ranking Addresses

Slick is a private search engine. It does not track users or use personal data, which means local and geo features have no reliable signal and perform poorly. The system works with a relatively small index of roughly 2 million documents. Even at this scale, it delivers competitive results on head queries and reasonable performance on many long-tail ones.

You can try Slick here: slicksearchhq.com


Version Evolution

Launch version used a single-stage score. It handled straightforward cases reasonably but struggled with navigational queries and long-tail topics. Brand homepages often lost to off-target pages on the same domain, and some how-to queries returned loosely related content.

Intermediate version added targeted heuristics and intent rules. This improved brand navigation for major sites and introduced filters for definitions, who-is queries, and ecosystem pages. Navigational accuracy increased for several well-known domains, but Wikipedia-style mismatches on definitional queries and token overlap problems remained.

Latest version uses dual recall (lexical plus dense) with fusion, followed by cross-encoder reranking and a normalized heuristic blend into a final score. Vector similarity no longer mixes directly into the main heuristic sum. Semantics now operate primarily through recall and reranking. The shift produced clearer improvements on both head and long-tail queries.


Real Ranking Examples

Query Result
apple Earlier versions surfaced product or search pages. Now consistently returns the main homepage.
stackoverflow Earlier versions ranked dataset repos above the official site. Now ranks first.
facebook Still lands on a locale version (en-gb.facebook.com) rather than the apex domain. Better than before, but not ideal.
amazon Improved from influencer shop pages to the main homepage.

Informational and question

Query Result
what is artificial intelligence Previously returned a researcher biography. Now surfaces a direct definitional article.
what is machine learning Similar shift away from mismatched Wikipedia entries toward explanation-focused content.
how does the internet work Improved over clearly wrong articles, but can still favor IoT-related guides over core network explanations.

Transactional

Query Result
download vscode Returns the correct download page, though its final score remains lower than expected.
best smartphones 2025 Ranks relevant review roundups higher than before.

Improvements in the Latest Version

The hybrid setup reduced several earlier failure modes:

Navigational accuracy increased for major brands (Amazon, Microsoft, LinkedIn, Stack Overflow, GitHub)

Definitional queries show fewer wrong biography or category page matches

Long-tail queries handle better due to improved recall and reranking

Empty result sets no longer occur in standard test suites

Cross-encoder scoring provides more consistent top-result quality across intents

Top results are more stable across most query types, especially navigational queries where official homepages are consistently surfaced. These gains hold even with the modest 2 million document index.


Remaining Issues

Canonical domain selection can still pick locale or subdomain variants (facebook, twitter)

Explainer queries like “how does the internet work” show semantic drift toward adjacent topics

Who-is queries rarely surface standard biography pages

Local search is especially weak, as expected for a private engine with no geo or user data

Comparison queries like react vs vue need better coverage of dedicated comparison content

A few correct transactional results, such as the VS Code download page, receive lower blended scores than they should

These gaps are most visible on edge navigational cases and queries requiring precise intent separation.


Summary

The ranking system has made steady progress in navigational reliability and definitional handling. The move to dual recall plus reranking addressed several weaknesses from earlier single-stage scoring. Work continues on the remaining gaps.

The full list of 75 test queries, along with more detailed analysis, is available in the Slick Statistics repository on GitHub.