A reading is only worth the method behind it. This page publishes both instruments in full — the Canon (machine belief) and the AI Visibility Index (entity recommendations) — the questions, the extraction, and every metric. Anyone can reproduce a number from the transcripts.
Canon questions are put to each model with no web-search or retrieval tools enabled. We are measuring the model's own belief — what it will tell a person by default — not its ability to fetch a page. Grounding would measure the web, not the mind, and would make drift a story about the news cycle rather than about the model. (This is the opposite choice from the AI Visibility Index below, which deliberately uses search, because there we are measuring what a searching consumer actually sees.)
Every question is asked to every tracked model 3 times — repeated sampling is what makes stability measurable. A cheap-model pass then maps each free-text answer to one label from that question's taxonomy, with a short rationale so any classification can be audited against the original transcript. Answers that decline to take a position are flagged as refusals.
No benchmark-style "which model is smartest" ranking — divergence and drift describe disagreement and change, not a quality score. Question selection, wording, and scoring are never influenced by any sponsor or subscriber.
The entity lens asks the consumer-intent questions people actually ask when hiring a local business, with web search enabled (this is what a real searching customer gets). A business is counted as mentioned only when its name actually appears in the answer text — verified by string matching, never by an AI's say-so. Its visibility is its mention rate across the question bank, weighted by how many consumers use each assistant:
Directories and aggregators (Yelp, Angi, Google, and the like) are excluded; alias spellings of one company are grouped; a name an AI lists but does not actually write is not counted.