One useful extra control is to treat each query run as a sample, not a verdict: hold wording, locale, language, signed-in state, and mode constant, then repeat the same prompt across several days. Save the exact prompt and timestamp alongside the cited URLs, not only the brand mention, since a citation to the page you’re evaluating is a different outcome from a mention or a homepage link. I’d compare citation frequency and source overlap over a rolling window; for technical queries, verify each cited claim against the primary documentation. This makes the five-query test much easier to reproduce when answers fluctuate.