Exa ATLAS and Tavily Test Different Kinds of Search Completeness
Exa ATLAS and Tavily’s company benchmark use different tasks and scores. A worked example shows why their results cannot form one search leaderboard.
Tagged collection
Latest articles tagged “Original Research”
Open an article for the complete method, evidence, limitations, and a focused reader discussion.
-
-
AuthGR: What Authority-Aware Generative Retrieval Changes
An ACL 2026 industry paper connects authority-aware retrieval with stronger ranking judgments and higher click behavior inside a commercial search system.
-
What Perplexity’s Real-World Mistakes Experiment Actually Proved
Perplexity reduced tool failures with selected correction traces, but the strongest online comparison did not show a significant satisfaction gain.
-
Top-500 Retrieval Still Missed Useful Evidence
Re:CAP recovered relevant documents missed by flat retrieval, including gaps left by a 1,500-candidate ensemble. That changes how retrieval coverage should be audited.
-
Baidu and Google Gave Similar AI Answers From Different Source Webs
A 500-query bilingual study found high answer similarity but very low cited-host overlap between Baidu and Google AI overviews.
-
13.4 Billion Web Questions Show FAQ Frequency Is Not Search Demand
The most repeated questions on the web increasingly come from templates and commerce pages. Publishers need behavioral demand evidence before creating another FAQ.
-
AI Answer Citations Can Fail at Four Separate Layers
Four recent papers expose separate failure points between retrieval and a trustworthy AI citation. SearchEngineAnswer reproduced PageRecall’s public tier-0 checks and built a claim-level audit ledger.
-
A GEO Measurement Paper Links AI Visibility to Business Impact, but Tests Simulated Answers
A new GMMM preprint connects generated-answer inclusion, question volume, system use and notice probability, but its empirical demonstration is simulated.
-
Schema.org 30.1 Product Fields Pass Google Parsing but Fail the Public Validator
A dated compatibility test of six Schema.org 30.1 commerce properties finds a lagging public validator, successful Google parsing and a nested Product risk.
