Ask HN: Is there a single website with power law impact on trained LLM?

If the SOTA ChatGPT, Claude, Llama, Mistral, Gemini or Gro, were trained on the exact same data except removing 1 website X, is there any website X that would cause a significant drop in performance across a broad range of benchmarks? I'm trying to find if someone has done this test on a SOTA model. Candidates I'm thinking of: Reddit, Wikipedia, Arxiv.