Thank you for this post. Very timely.
the canonical-entity-vs-infinite-prompts distinction is the clearest explanation I've read so far or why "track every prompt variant" tools are a dead end

Question: running the same probe with web search on and off in parallel is the sharpest part of this setup, correct? Since the off version holds steady (only reflects training data), any movement you'd see in the on version over time should map back to retrieval and grounding, not model drift. That makes the on/off delta a pretty clean before-and-after signal for proving whether a specific content/SEO change shifted selection rate, rather than just eyeballing the day-to-day swings.