Track the change and prove what moved it.

Measure the same thing twice: the same questions, on the same engines, at the same cadence, against the same frozen list of competitors (chapter 3). Change the question set between runs and you are comparing two different samples; start new questions as their own series instead.

Then compare the runs question by question. Checking whether two 95% intervals overlap is far stricter than it looks: Payton, Greenstone and Schenker showed in 2003 that two such intervals from identical populations overlap more than 99% of the time, so the overlap check behaves like a 1% test and swallows real drops. On 200 questions, a fall from 42% to 30% shows overlapping 95% intervals, yet a standard two-proportion test gives p = 0.012. If 30 questions dropped you and 6 picked you up, McNemar's paired test on those flips gives p of about 0.00006 (the worked example).

Wait for it to hold. In our daily tracking, 47% of the changes in a single answer had reversed a day later, and where we could follow a change for two more days it held all three days 36.7% of the time. One tracked competitor read 75% on a Monday and 45% on the Friday of the same week, which looks like a 30 point fall; across the week it was named in 89 of 140 answers, 63.6%, and nothing had fallen (the cadence study).

Act on a change once it holds for two or three days, or once the intervals of two readings separate. An alert that fires on every crossing fires all the time on numbers this noisy.

1. Week 1, baseline. Pick your questions, ask every engine, record who is named and what is cited (chapter 2). Run it twice, a day apart.
2. Week 2, access. Check each bot against your key pages, fix CDN blocks and JavaScript-only content (chapter 4).
3. Week 3, identity and sources. Publish your one-sentence definition everywhere, and pitch the five most-cited pages that leave you out (chapter 5).
4. Week 4, re-measure. Ask the same questions on the same engines. Keep what moved beyond the interval; rethink what did not.