AI answers change faster than rankings: citation churn across six engines

Search rankings change, but slowly enough that a monthly report is usually accurate. AI answers are different. In Q1 we re-ran 840,000 prompts every week across ChatGPT, Perplexity, Gemini, Claude, Copilot and Google AI Overviews, and measured how often the cited sources changed.

How we measured churn

For each prompt and engine, we compared the domains cited in one week with those cited four weeks later. A domain counts as churned if it was cited in the first week and absent from all three runs four weeks later. We excluded runs where the engine returned no citations in either week.

Nearly half of sources changed within a month

Weighted by citation volume, 46% of cited domains had churned after four weeks. The rate varied widely by engine:

• Perplexity: 52%

• Google AI Overviews: 49%

• Copilot: 47%

• ChatGPT: 44%

• Gemini: 41%

• Claude: 33%

Churn was highest for prompts in fast-moving categories, such as AI tools and fintech. It was lowest for health and other regulated topics, where answers leaned on a stable set of institutional sources.

Recommendations are steadier than their sources

Brand recommendations churned less than the citations behind them. A brand recommended in week one was still recommended four weeks later 71% of the time, even when the cited pages had changed. In practice, engines often reach similar conclusions from a rotating set of sources.

What this means for measurement

1. Run each prompt at least three times per engine per week; single runs overstate movement.

2. Report a four-week rolling Share of Answer rather than a single week.

3. Track citations and recommendations as separate series.

4. Review Answer Diff weekly to catch lost citations before they show up in the trend.

Churn cuts both ways. If a competitor holds a citation you want, the odds are good it will open up within a quarter, and teams with a stronger page already live tend to take it.