Controlled tests by Sympha Foresight in which AI models chose which sources to cite. These are model experiments, not live search-engine measurements.
In a controlled Sympha Foresight test across 172 real sources and 35 queries, a citability scorecard predicted which sources AI models chose to cite, with AUC 0.803 (GPT-4o-mini) and 0.896 (Claude Sonnet).
In the same research, a source's high Google ranking did not correlate with high citability.
In a 216-vote controlled test, reputation cues did not change which sources AI models chose to cite. Substance drove citation, not brand fame.
In controlled tests, putting the brand name inside the quotable claim sentence raised named-in-answer attribution by 5.5 percentage points.
In controlled tests, one substance-grounded rewrite beat multi-agent deliberation for citability (multi-agent scored −0.096 with zero wins in five paired tests).
In a held-out controlled test, one substance-grounded rewrite raised multi-model citation win-rate by +0.44 (31 paired wins, zero losses).
See how these findings apply to your site: Sympha Foresight →