New paper: GEO-SFE — How content structure shapes AI citation behavior

We’re excited to share our latest paper, “Structural Feature Engineering for Generative Engine Optimization: How Content Structure Shapes Citation Behavior”, now available on arXiv. The paper formalizes a question we keep running into with customers: when an AI answer engine decides which pages to cite, how much of that decision is driven by structure rather than meaning?

→ Read the full paper (PDF)  ·  arXiv abstract

The gap we’re closing

Most published GEO work focuses on semantic rewrites — what the content says. Our team kept seeing the same pattern in production: two pages with near-identical claims, one cited reliably by ChatGPT / Perplexity / Gemini and one ignored. The difference wasn’t the message — it was the scaffolding around it. GEO-SFE is our attempt to make that scaffolding measurable and controllable.

What GEO-SFE proposes

The framework decomposes page structure into three levels and models how each shifts citation probability across different generative engine architectures:

• Macro-structure — overall document architecture (sectioning, heading hierarchy, narrative order).
• Meso-structure — information chunking (paragraph density, list use, fact placement, callouts).
• Micro-structure — visual emphasis (bolding, inline highlighting, typographic cues retained in the rendered DOM).

On top of those levels we build architecture-aware optimization strategies and predictive models that preserve semantic integrity — the goal is to make structural rewrites that the engine notices, without changing what the page actually means.

Results

We evaluated GEO-SFE across six mainstream generative engines and saw consistent gains:

• +17.3% citation rate
• +18.5% subjective answer quality

The improvements held across engine families, which is the result we cared about most — it suggests structural effects generalize rather than overfitting to one model’s quirks.

Authors

• Junwei Yu
• Mufeng Yang
• Yepeng Ding
• Hiroyuki Sato

The work was carried out as a research collaboration tied to the VeReach GEO program. If you want to discuss applying GEO-SFE to your own content surface, get in touch with the team.

What this means for the product

GEO-SFE isn’t a standalone tool — it slots into the rewrite-and-retest loop already inside VeReach GEO. Expect structural diagnostics and rewrite hints to start surfacing as first-class signals alongside the semantic ones, with the same PIRR / PRP / CCD scoring driving the loop.

For the full methodology, the architecture-aware optimization strategies, and the per-engine breakdown, please read the paper: arxiv.org/abs/2603.29979.