Casual experiment hint that models seem to search for different stuff (spock.is)

A recent personal experiment has revealed intriguing insights into how large language models (LLMs) interact with web content, particularly regarding the retrieval of specialized files like "llms.txt." Initially, six popular models were prompted to explore a personal website without being directly instructed to seek out this file, with only OpenAI's gpt-5.5 occasionally checking for it. However, after a subtle nudge in the site's design that encouraged LLMs to access the llms.txt file, all models successfully retrieved its content. This finding highlights the importance of effectively advertising structured files to ensure they are accessed, suggesting that LLMs require explicit cues to engage with specialized information. This experiment holds significant implications for the AI/ML community, as it demonstrates that LLMs exhibit differing queries based on the model, indicating they might hold unique 'interests' in the data they retrieve. The author discovered notable variations in information retrieval patterns among models, reflecting their distinct strengths and preferences. This divergence raises questions about model-specific search optimization, potentially leading to a future where websites must cater not just to human visitors but to the varied preferences of different models. As LLM adoption grows, reshaping user interactions with the web, the implications for structuring content—like adopting a “reverse mullet” internet—could fundamentally change web design and information dissemination strategies.