Generative Search Ads Split Relevance, CTR and Revenue
A Meituan paper reports a seven-day test of objective-specific generative ad retrieval, with separate relevance, click and commercial policies.
Latest research
Completed tests, case studies, and evidence reviews
Case studies, experiments and analysis of original research.
-
-
Same AI Question, Different Language: A 67,200-Response Audit
An audit of GPT, Claude and Gemini found that responses to the same Ukraine-war statement bank varied across 112 language and script conditions.
-
Algebraic Retrieval: 38 Tests Passed, With One Reproduction Limit
We ran the portable Algebraic Retrieval suite and verified its receipt hash. The full 11,429-document Java fixture remains outside this local test.
-
ChatGPT, Claude, Grok and DeepSeek Search the Web Differently
A 171,264-conversation study shows that four AI platforms differ at every search stage, from deciding to browse through selecting citations.
-
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.
-
Chrome Now Measures Ad Count, Density and Weight in CrUX
Chrome added four experimental ad metrics to CrUX. We measured a fixed publisher panel and found why count, density, CPU and network weight need separate diagnoses.
-
What 30,000 Reddit Answers Reveal About Which Voices Survive AI Search
A 30,000-output study finds retrieval drives most variation while Reddit Answers favors early, top-level and more formal comments in its selected evidence.
-
Google Research Cut Retrieval Fan-Out Time 12× to 20×
Google researchers report that Retrieve-for-Train generated ten retrieval directions 12× to 20× faster in fashion and music benchmarks, while the diffusion model traded some recall for greater diversity.
-
Q2D-Web Shows Why AI Citations Depend on First-Stage Retrieval
A 190-million-document benchmark shows why retrieval recall, unique evidence coverage and downstream citations must be measured separately.
