Notebook
How much do these numbers move?
Ask ChatGPT the same question twice and you get two different answers. Any tool that does not say so has not checked. Here is ours, checked.
The measurement
On 2026-08-04 we ran 25 questions 5 times each, on the same surface, on the same day: 125 answers, 0 failures.
- Cited domains appearing in all five runs
- 10.5%
- Cited-domain overlap between any two runs
- 35.4%
- Brand/prompt pairs that flip between runs
- 30.7%(46 of 150)
- A single answer agreeing with the five-run majority
- 91.1%
- Prominence movement when a brand is named twice or more
- 0.97 positions
Where a brand lands in an answer is stable enough to trend, and which sources an answer cites is not. So we smooth mention rate and prominence over a seven-day window and never show a citation count for a single day. Anyone charting daily citation data is charting noise.
Where the answers come from
Bought scrapes of the answer surface a real person sees. Not a model API asked what it would have said — that is model introspection, it disagrees with the real surface, and it is how most of this category produces its numbers. Every stored answer records which surface produced it, so the claim is checkable per row rather than a promise on a page.