What happened on the next measured date?
We grouped the panel logs by question, engine, and date. For each successive pair of measured dates, we took every distinct business named first on the earlier date and checked whether it was named first in any recorded run on the later date.
That produces 385 first-position opportunities, not 385 independent experiments. If a question names two different businesses first across its runs on the earlier date, both enter the denominator. Multiple runs on the later date also give a business more chances to count as held.
| Engine | First-position holds | Hold rate | Date pairs | Gap: min / median / max days |
|---|---|---|---|---|
| ChatGPT | 21/44 | 47.7% | 34 | 1 / 1 / 2 |
| Claude | 3/21 | 14.3% | 14 | 1 / 2 / 2 |
| Gemini | 45/122 | 36.9% | 82 | 1 / 1 / 2 |
| Google AI Mode | 32/86 | 37.2% | 59 | 1 / 1 / 2 |
| Google AI Overviews | 4/25 | 16.0% | 19 | 1 / 1 / 2 |
| Perplexity | 45/87 | 51.7% | 66 | 1 / 1 / 2 |
Claude’s 21 opportunities are a small sample and combine answer modes the source analysis does not separate. Treat that result cautiously. Name matching is automated and can make mistakes. Runs from the same question are correlated, and this is a convenience archive of local markets rather than a representative sample of all AI questions.
First position and presence are different
A business can lose first place and remain in the answer. We therefore count presence separately: every distinct business named anywhere on the earlier date is checked against the union of names on the later date.
| Engine | Presence holds | Hold rate |
|---|---|---|
| ChatGPT | 76/159 | 47.8% |
| Claude | 3/33 | 9.1% |
| Gemini | 232/565 | 41.1% |
| Google AI Mode | 217/577 | 37.6% |
| Google AI Overviews | 31/92 | 33.7% |
| Perplexity | 232/382 | 60.7% |
These counts do not show continuous survival between observations. A business might disappear and return between checks. The archive also does not establish whether each business was being actively managed or whether competitors changed their sites.
What this means for a business
One screenshot proves one observation. It cannot establish a stable position, a trend, or the effect of a change to your website.
To evaluate a change, keep the questions and collection conditions fixed, measure before and after more than once, and report the denominators and failures. Record the intervention date and concurrent changes. A difference is an observed association; attributing it to the intervention requires stronger evidence, such as a suitable comparison group.
Data and method
The anonymous counts, summary, and recomputation script let you reproduce every fraction above. The data notes describe the units and limitations. The private source logs and identity mapping are not published.
The export preserves the original algorithm: rows need a date, engine, question ID, and a nonempty ordered competitor list. Names are normalized; first names and presence names are pooled within each date. Empty lists are excluded rather than counted as disappearances. That selection can overstate persistence, and the first-position field reflects the ordered competitor list, not an independently audited ranking of every business in the answer. The next edition should retain explicit empty outcomes, standardize run counts, and hand-check identities before estimating survival.
Correction — November 7, 2026
The original page described a one-to-two-day expected lifetime. That was a geometric extrapolation with assumptions these observations do not establish. We have removed it. The current source logs reproduce 385 first-position opportunities, including ChatGPT 21/44; the original July edition reported 383 and 21/42. The source archive already recorded that count correction in October. This release publishes the counts and removes unsupported claims about an unmanaged baseline, causal effects, and a seven-day case whose publication evidence was not established here.