RESEARCH | GEO-SFE
AI citations follow structure more than semantics.
GEO-SFE breaks page structure into measurable, controllable features so content earns more generative-engine citations without changing what it claims. Across six major generative engines it improved citation rate by +17.3% and subjective answer quality by +18.5%.
Results
- Citation rate
- +17.3%
- Subjective answer quality
- +18.5%
- Major generative engines evaluated
- 6
Figures are results under the paper’s experimental conditions. They do not guarantee the same lift for any individual brand or site.
Three structural layers
- Macro-structure: Document-level design: sectioning, heading hierarchy and argument order.
- Meso-structure: Information chunking: paragraph granularity, lists, fact placement and callouts.
- Micro-structure: Visual emphasis: bold, inline highlights and other cues that survive rendering.
How VeReach uses it
GEO-SFE is not a standalone tool. It runs inside VeReach GEO’s rewrite-and-remeasure loop, so structural diagnoses and rewrite suggestions are re-measured alongside semantic signals.
Paper
Structural Feature Engineering for Generative Engine Optimization: How Content Structure Shapes Citation Behavior
Junwei Yu, Mufeng Yang, Yepeng Ding, Hiroyuki Sato
This page summarizes the paper. See the paper for methods, per-engine breakdowns and evaluation conditions.
More VeReach papers
Research on how GEO should be evaluated and how far AI-based evaluation can be trusted.
- Conversational Capture: A Trajectory-Level Framework for Evaluating Generative Engine Optimization in Multi-turn Human-Agent Interaction
Junwei Yu, Jieyu Zhou, Mufeng Yang, Yepeng Ding, Hiroyuki Sato · September 30, 2026
A framework for evaluating GEO across a whole multi-turn conversation instead of a single answer. It formalizes conversational capture, where a source cited early becomes more likely to be cited again through retrieval conditioned on history and through users’ follow-up questions, and shows theoretically that single-turn evaluation hides this effect.
Read on arXiv - JuryFlow: Disagreement-Guided Human-in-the-Loop Multi-Agent Evaluation
Mufeng Yang, Junwei Yu, Yepeng Ding · September 30, 2026
An evaluation framework for LLM-as-judge systems that treats disagreement between judges as an uncertainty signal instead of voting it away. A human makes one targeted correction where judges disagree, and that correction is propagated to related cases and reused as rubric entries. The authors report better agreement with gold labels than single-judge and majority-vote baselines on MT-Bench and LLMBar.
Read on arXiv