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.
From this desk
Start with the most useful page
The featured page offers a strong introduction. The remaining cards help readers move through this editorial section without turning the archive into a service menu.
Read the featured guide
Run the citation-ready passage test
Check whether a passage states a bounded answer, names its evidence, exposes important limits, and remains understandable outside the surrounding page.
Citations
Make the answer easy to evaluate
Separate eligibility from citation-worthiness, expose sources, clarify entities, and write passages that survive retrieval out of context.
Evidence
Classify claims before drafting
Distinguish facts, observations, interpretations, recommendations, and opinions so each receives appropriate support.
Workflow
Use AI without publishing generic work
Bound the brief with sources, require an original contribution, and keep the final editorial judgment with a person.
Latest AEO & AI Search stories
Latest answer-engine articles
New stories are added when they contribute a verified change, method, test, or useful framework—not a citation promise.
-
-
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.
-
Anthropic Traced 8,913 AI-Written Articles to 70 Fake News Sites
Anthropic documented a coordinated network of fabricated news sites, paired social accounts and rigid publishing constraints. The useful lesson is the network footprint, not a guess about whether one paragraph sounds AI-written.
-
Google Tests ‘Dive Deeper’ Overviews Inside Discover Video Cards
The limited Discover experiment adds a topic overview between some video cards and related web stories. Its effect on publisher clicks remains unmeasured.
-
A Court Filing Reports 51% to 94% Lower Publisher CTR in Copilot
A plaintiffs’ filing reports much lower publisher click-through rates in Copilot than Bing Web Search. Here is what the figures prove, omit and change.
-
Claude in Chrome Compliance API: What Browser-Agent Transcripts Capture
Anthropic’s Enterprise compliance export records prompts, responses and text tool activity for Claude in Chrome, with important exclusions and retention limits.
-
GPT-5.5 Retires From ChatGPT and Codex on October 14: Migration Checklist
GPT-5.5 is leaving ChatGPT, ChatGPT Work and Codex on October 14, while the API remains unaffected. Use this surface-by-surface migration plan.
-
Google Tests Publisher Payments for Contributions to AI Answers
Google is testing payments when publisher content significantly contributes to AI answers. Here is what is confirmed, undisclosed and worth measuring.
-
What Perplexity’s CobbleDB Reveals About AI Search Retrieval
Perplexity reports roughly 80% lower read latency after separating passage processing, publishing and hot retrieval. Here is what publishers can and cannot infer.
A useful standard
What this hub does—and does not—measure
AI answers vary by platform, query, location, account state, and time. A citation is an observation in one answer context, not a permanent ranking.
Eligibility is the first gate
A page still needs ordinary crawlability, indexability, useful content, and accessible text before any answer system can evaluate it.
Evidence creates inspectability
Primary sources, precise entities, dates, methods, and visible limits make a claim easier for readers and systems to check.
Retrieval is not causation
Referral and citation reports can document appearance. They do not prove why one source was selected or guarantee the next answer.
References and next steps