What Gets Cited: Competitive GEO in AI Answer Engines

Technical Summary: What Gets Cited: Competitive GEO in AI Answer Engines

Problem Statement
AI answer engines utilizing Retrieval-Augmented Generation (RAG) synthesize responses from retrieved documents but cite only a small subset of sources. This creates a visibility bottleneck where ranking alone is insufficient; content owners must also optimize for being explicitly cited in the generated response. While Generative Engine Optimization (GEO) is an emerging field, existing evaluations often fail to address the competitive citation setting practitioners face. Most benchmarks measure single-source visibility within a fixed set, failing to isolate which of two similar candidates wins a citation when they compete directly. Furthermore, observational analyses on live systems conflate content effects with retrieval rank, presentation order, and interface artifacts, complicating causal inference.

Methodology
The authors propose a controlled, competitive GEO framework using a two-document RAG testbed. The core experimental design involves injecting exactly two candidate sources (Variant A and Variant B) into the model context for a specific query, ensuring the model sees no other sources. The study isolates the causal effect of individual content factors through the following protocol:

1. Matched-Pair Design: From 100 anonymized product review articles across 50 categories, the authors constructed 1,440 scenarios. In each scenario, two variants differ in exactly one of 18 content factors while matching on all other facts, prices, specifications, and length (±5%).
2. Factor Taxonomy: The 18 factors span six categories: Content Match (Topic Mismatch, Keyword Gap), Completeness (Price, Specifications, Comparisons), Trustworthiness (Hedged Language, Evidence, Contradictions, Promotional Tone), Readability (Structure, Scattered Information), Competitive Standing (Value Proposition, Depth, Social Proof), and Freshness (Timestamps).
3. Bias Control:
   • Anonymization: Brand names, product models, and publishers were replaced with fictional aliases to prevent pretraining familiarity bias.
   • Counterbalancing: Source order ([A, B] vs. [B, A]) was swapped to separate content effects from position bias (except for the "Lower List Position" factor, where order is the treatment).
4. Execution Scale: The study executed 252,000 trials across six LLMs (Gemini-2.5-Flash, GPT-5-Nano, GPT-5-Mini, GPT-5.2, Claude-3.5-Sonnet, Kimi-K2-Thinking). Each scenario-query-order combination was run 5 times for statistical stability.
5. Statistical Analysis: Citation preference (which source is cited first) was modeled using logistic Generalized Linear Mixed Models (GLMM). The models included nested random effects for scenarios and scenario-order combinations to account for pseudoreplication, yielding Odds Ratios (OR) for each factor after adjusting for position bias.

Key Results
The analysis reveals a clear hierarchy of factors driving citation preference, categorized by their consistency across models:

• Gatekeeper Factors (Unanimous, High Impact): Four factors were significant across all six models with very strong effects (OR > 100). Failing on any of these drastically reduces citation odds regardless of other content strengths:
  1. Topic Mismatch: Content discussing unrelated products vs. query-relevant items.
  2. Price Not Mentioned: Absence of explicit price information.
  3. Recent vs. Old Timestamp: Content dated 2026 vs. 2019.
  4. Lower List Position: Being listed second rather than first in the retrieved set.
• Differentiators (Secondary Impact): Seven factors showed moderate to strong effects (OR 2.1–243) in 4+ models, including completeness (Missing Specifications, Less Comprehensive), trust cues (Hedged Language, Claims With Evidence), and competitive standing (Keyword Gap, No Comparisons).
• Insignificant Factors: Formatting choices (Dense vs. Structured paragraphs, Scattered vs. Organized information) showed negligible impact. Other factors like "Overly Promotional" tone or "Weaker Social Proof" lacked consensus across models.
• Model Sensitivity: Models varied in sensitivity; Kimi-K2 was the most sensitive (83% of factors significant), while Gemini-2.5 and Claude-3.5 were more selective. However, all models agreed on the four gatekeeper factors, suggesting universal citation drivers independent of model architecture.

Contributions
The paper makes three primary contributions:

1. A Controlled Competitive Protocol: A reproducible method to test GEO under competition by presenting two sources, varying a single factor, anonymizing entities, and counterbalancing order.
2. Large-Scale Empirical Evidence: Quantitative measurement of how 18 specific content factors influence citation preference across six diverse LLMs using mixed-effects modeling.
3. Actionable GEO Checklist: A prioritized, evidence-based workflow for practitioners to diagnose and improve citation visibility, distinguishing between retrieval failures (SEO) and content quality issues.

Significance and Claims
The authors claim that this study shifts the focus of optimization from traditional ranking to citation selection in AI search. By isolating causal factors, the work provides a definitive hierarchy of what makes a source "citable" in a competitive environment. The findings suggest that while formatting and minor stylistic edits have little impact, topical relevance, explicit data (prices/specs), freshness, and list position are the primary gatekeepers. The paper validates this approach through an internal pilot at Sprinklr, where teams reported positive qualitative feedback on the workflow's usability for diagnosing content gaps. The authors acknowledge limitations, noting that the two-source design isolates factors better than larger candidate sets but does not fully replicate production RAG pools of 5–10+ documents. Future work is proposed to extend this factorial discipline to larger candidate sets and multiple citation slots.