Princeton-led preprint proposes standard metrics for generative search attribution

A February 2026 preprint from researchers at Princeton and industry collaborators defines two metrics aimed at standardizing GEO measurement across platforms. The work responds to a gap merchants already feel: referral analytics undercount AI citations, while rank trackers miss answer-surface visibility entirely.

Counted when a brand URL appears in the cited source set of an AI answer, weighted by estimated answer impressions. GI treats "being named with a link" as the generative analogue of an impression—not a click.

Median days until a URL drops out of the top three citation slots for a fixed prompt set, analogous to SERP decay curves. Short CHL suggests volatile or thin content; longer CHL suggests durable answer-ready pages.

Authors analyzed roughly 8,000 prompts across shopping and informational intents. FAQ and schema completeness explained about 34% of variance in CHL in their sample—larger than raw domain rating alone. They caution that platform-specific reranking can shift CHL quickly after model updates.

• Stop relying solely on last-click GA for AI discovery
• Track prompt-level citation persistence, not one-off mentions
• Treat content updates as experiments with pre/post CHL windows

Visora incorporates compatible proxies—Citation Rate, SOAV trends, and weekly snapshots—in Monitor dashboards for Pro users. If your team adopts GI/CHL language internally, Visora exports map cleanly for board reporting without custom spreadsheets.

Source: preprint summary + author slides, February 2026.