ChatGPT's 74% Weekly Domain Rotation
Key findings
- 74% of cited domains rotate out within 7 days (SISTRIX, 82,619 prompts)
- Single-run citation checks have 40%+ week-over-week variance
- Minimum viable measurement requires weekly panel cadence over 8+ weeks
The problem with single-run checks
Many AEO tools offer "check if you're cited by ChatGPT" as a one-shot lookup. The user enters their domain, the tool makes one API call, and reports whether the domain appeared.
This produces a measurement with a shelf life of approximately one day.
The SISTRIX finding
SISTRIX ran 82,619 prompts through ChatGPT weekly for 17 consecutive weeks. Their finding: 74% of domains cited in week N are not cited in week N+1 for the same prompt.
This rotation rate means that a single observation has a 74% chance of being different from the next observation of the same prompt. In statistical terms, the week-over-week test-retest reliability is approximately 0.26 — far below any threshold for a valid measurement instrument.
Implications for measurement design
-
Minimum viable cadence is weekly, over 8+ weeks. With 74% weekly rotation, you need at least 8 weekly observations to distinguish signal from noise with 95% confidence.
-
Point-in-time "visibility checks" are marketing, not measurement. A tool that shows you a single ChatGPT response is showing you noise. It cannot tell you whether your brand is consistently cited or whether you happened to appear on the day the check ran.
-
Confidence intervals widen dramatically at low cadence. Our bootstrap analysis shows that 95% CIs for citation share drop below actionable width (less than 10 percentage points) only after 6-8 weeks of daily measurement.