Every claim, on the record.

Claude would cite Honolulu AC companies' own websites less than 5% of the time. Measured at 2%. The commit timestamp predates the first measurement run. Committed to a dated, timestamped record before the measurement ran.

Wording note, added 2026-09-10. The sentence above is reproduced exactly as committed and is not edited, because editing a pre-registered claim after the result is known would defeat the purpose of pre-registering it. The verb is now known to be wrong: Claude searches nothing and returns no URLs, so it names rather than cites, and the 2% is a share of answers that name the business, not of cited URLs. The prediction and the result are unaffected. See the two-layer correction on the methodology page, dated 2026-09-06.

Across October 2026, fewer than 10 percent of the pages ranking in classic search's organic top five will also be cited by the AI answer engines asked the same question on the same day. Open. Result due 2026-11-01. We do not know it yet either.

What would make us wrong: Pooled overlap at or above 20 percent means classic ranking and AI citation are substantially the same thing on these questions, and we say so in public.

Disclosure. Exploratory 14-day sample seen 2026-09-15 before committing. Overlap 1.5 to 3.4 percent per engine, on the pre-correction definition of cited_urls. The October window is out of sample and uses the corrected instrument.