Proof & authorship record
ChatGPT Citation Drift: I Re-Ran 200 Prompts Weekly for 90 Days. Here's How Much Citations Move.
Citations from ChatGPT aren't stable — they drift run-to-run, day-to-day, and across model versions. I re-ran 200 buyer-intent prompts every week for 90 days. Here's the actual variance, and what it means for your AEO program.
- My role
- Sole researcher and author
- Period
- Published 2026-05-15
- Last verified
- 2026-08-01
What I personally did
- Designed the research question and the sampling method.
- Collected and graded the data personally — no outsourced or scraped third-party dataset.
- Wrote every word of the published piece and maintain it after publication.
How the numbers were measured
- The same prompts repeated on a schedule over the study window, with cited domains logged each run.
- Drift measured as change in cited-domain set between runs, not as a change in wording.
Evidence you can check
Repeat-prompt log
Run cadence and prompt wording are published in the study.
Verification: Repeat the documented prompts yourself and compare your citation set to the published one.
Open source
What this record does not prove
Model versions changed during the window; some drift is attributable to model updates rather than to the sites.
AI Answers hub
Ask ChatGPT, Perplexity, or Gemini about this research?
24 citation-ready answers on AEO, GEO, schema, local SEO, and Bangladesh search — written in the 40–80 word format LLMs quote.