Measured US AI answer-surface visibility, published as a dataset.

Method, sample, limitations

Method

We put one set of buyer-style questions to every answer surface in the panel and recorded what came back: whether the surface answered at all, and which domains it cited. Valid runs differ by surface, so each surface carries its own denominator and we never pool or compare rates across them. Execution details — the wording of the questions and the judgment configuration — stay private; only the aggregates are published.

Sample

50 questions put to 14 answer surfaces from New York on 2026-07-27: 700 planned runs, 695 of them valid. A second block from the same run pools citations across 15 surfaces — 8662 citations in all, from 50 questions at 1 pass per question.

Measured on our sampled query panel — the questions are ours, so the figures can differ from Naver's own internal counts.

Limitations

This run went one pass per question. That answers what each surface returned and how wide its citation pool was. A ranked citation ladder needs a deeper sample, so that row stays empty and states what it is waiting for. The second block counts every citation handed out, so its numbers sit on a different denominator from the per-observation rates other boards publish. These are surfaces observed in one city on one day, so do not generalize beyond the measured scope. Per-surface denominators are printed next to the table.

Why anyone publishes measured numbers here

Nearly every figure circulating about AI search visibility traces back to a vendor deck, and the trail stops there. No denominator, no window, no way to recompute it. Cite one in an article or a board pack and you have borrowed a number you cannot defend, which is a professional risk that lands on you, never on whoever produced it.

This release is the alternative. Every count sits next to the runs it came from, and denominators that differ by surface are never pooled. Where a figure has not been measured well enough to publish, the row stays empty and carries its reason. Where a number has already appeared on one of our research pages, that page is linked at the bottom, so a reader who wants to argue with us has somewhere to start.

Reuse it freely, with attribution. A company that sells measurement should be the easiest one in its market to fact-check, and the same protocol runs behind our commercial panels — 50 questions, 15 engines, 7 runs each.