The Retrieval-to-Citation Funnel in AI Search: Evidence from Eight B2B Visibility Projects
Description
When an AI search engine pulls a web source into its context to answer a query, how often does that source actually appear as a citation in the answer? This step, the conversion from retrieval to citation, sits at the center of generative engine optimization but has not been quantified in public data.
This paper measures that rate across eight live B2B visibility projects covering manufacturing, tourism and rail, and other B2B sectors, tracking commercial prompts daily across ChatGPT, Perplexity, and Google AI Overview over one month (~13,200 domain-engine observations, 5,000+ unique source domains).
Three findings: (1) pooled retrieval-to-citation conversion is 55.6%, ranging from 40.9% (ChatGPT) to 76.8% (Google AI Overview), an ordering stable across nearly all projects; (2) engines cite largely disjoint source pools — Jaccard overlap of cited domains stays between 0.12 and 0.21 in every project; (3) which source type converts best is engine-specific, not universal.
This is a practitioner working paper; it has not been peer reviewed. All results are correlational, cover one month, and rest on a single measurement stack (Peec.ai). Client data is anonymized (P1–P8, sector labels only).
Files
funnel_per_client.csv
Additional details
References
- Google. AI features and your website: guidance for optimizing for generative AI features in Google Search.
- Google Search Central documentation, developers.google.com, accessed July 2026.
- Howard, J. The /llms.txt file proposal. llmstxt.org, September 2024, accessed July 2026.
- Ahrefs. Study of llms.txt adoption and bot access across approximately 137,000 websites. ahrefs.com/blog, accessed July 2026.
- Peec.ai. Platform documentation: source, retrieval, and citation metric definitions. docs.peec.ai, accessed July 2026.
- Ullrich, E. llms.txt: Was die Datei bringt, was Google dazu sagt, und wann Sie eine anlegen sollten. eullrich.com/blog, July 2026.