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.
Perplexity reports roughly 80% lower read latency after separating passage processing, publishing and hot retrieval. Here is what publishers can and cannot infer.
A 712-query audit found detector-classified synthetic pages among citations, but 27.1% of cited URLs were not analyzed and page accuracy was not tested.
A source-long-tail recomputation plus a separate LinkedIn engine-mix example show why aggregate AI citation rankings can hide most domains and platform dependence.