Every product decision maps to a citable claim—numbers you can audit, not marketing adjectives. Start here when you need to explain why a feature exists.
Papers on the left; product surfaces on the right. Lines are schematic—each card below cites the same mapping in prose.
Aggarwal, Mittal, Murthy, Tian, Patel et al. (KDD 2024)
Foundational. Shows +41% visibility from GEO methods, +115% lift specifically for rank-5 sites. The headline result that says: structural changes can move you from invisible to cited.
Content Agent uses Aggarwal's top techniques: Cite Sources, Statistics, Quotation Addition.
Yu et al. (2025)
Operational. Defines six numerical thresholds (heading depth, paragraph length, format diversity, emphasis density, internal-link density, content density) and three engine paradigms (STS, IR, ISG) with per-paradigm weight profiles.
Content Analyzer encodes every threshold and paradigm weight verbatim. Semantic preservation guardrail uses Yu's eq. 16 thresholds.
Anonymous (under review, 2025)
Strategic. Shows LLM-SEs cite 37% unique domains vs Google, and that every LLM-SE except ChatGPT prefers less popular domains. Inverts the 'big brands will dominate AEO' objection.
Authority Agent targets less-popular but topical domains. Citation Graph computes popularity-controlled lift over Tranco rank.
Chen et al. (2025)
Quantitative. Large-scale empirical comparison across engines. Documents earned-media bias, cross-engine variability, paraphrase sensitivity, cross-language stability.
Probe Engine runs paraphrase ensembles by default and reports stability scores. Per-platform weighting uses Chen's bias measurements.
Nestaas et al. (ICLR 2025)
Defensive. Catalogues PMA attack classes: visual hidden injection, instruction override, preference biasing, cross-page injection. Proves the position effect and the prisoner's dilemma.
Defensive Mode detector implements all four attack classes. Enterprise upsell scans both your pages and competitor pages.
v2 research ingestion on the roadmap includes deeper Google AI Overviews coverage, ContextCite-style attribution beyond our enterprise defensive tier, and Tranco rank as a first-class popularity prior for the Citation Graph—shipping as those pipelines harden.