The 'reverse engineering prompts' approach is interesting, but the variance in LLM responses based on temperature and system prompt updates makes consistency a major hurdle for this type of monitoring. One of the biggest technical challenges is distinguishing between when a model retrieves your site via RAG (live search) versus when it relies on stale training data. In the latter case, you can't really optimize for visibility without a new training cutoff, whereas RAG visibility can be influenced by site structure and indexing. Have you found a way to reliably trigger the search-tool use in your pipeline to ensure you're getting live results? Disclosure: I'm building Sivon HQ, where we track similar AI search visibility metrics.