Nginx logs from targeted prompts prove that ChatGPT, Claude and similar models perform provider-side origin fetches at inference time using dedicated user-agents such as ChatGPT-User/1.0 and Claude-User/1.0, often in multi-IP bursts with no referrer (Khallad, 2026; OpenAI, 2024).

The primary source documents reproducible GET sequences for candidate pages, robots.txt checks by Claude, and redirect following, establishing that retrieval is not limited to pre-training indexes. Cloudflare's 2024 automated traffic report shows AI bots now account for more than 20 percent of observed crawler volume, consistent with the burst patterns logged (Cloudflare, 2024). Earlier coverage omitted the inference-time distinction versus training crawls and under-reported cumulative server load when multiplied across millions of domains.

Synthesis with The New York Times documentation of 2023-2024 scraping litigation reveals the same user-agent tokens now trigger real-time RAG-style fetches rather than static citation from training data alone (NYT, 2023). This shifts uncompensated infrastructure costs to content originators while model providers avoid caching penalties. Logs therefore supply primary evidence that conventional referral metrics conceal the dominant scraping vector.