📄 Think Before Writing: Feature-Level Multi-Objective Optimization for Generative Citation Visibility

Summary

Answer engines built on LLMs, such as Perplexity or AI Overviews, show users only the sources they choose to cite. Ranking position matters less than before. A page that is not cited gets almost no exposure. Earlier generative engine optimization (GEO) methods mostly apply token-level edits to the text, like inserting keywords. These edits are hard to interpret and give little control over the balance between being cited and staying high in quality.

The authors propose FeatGEO. It first probes an engine with many LLM-generated queries on a single topic. Pages that are cited repeatedly across these queries become exemplars. Each page is then described by 13 interpretable features grouped into structure, content and language layers. NSGA-II searches this feature space to optimize both visibility and quality, and an LLM writes the actual page from the chosen configuration.

On GEO-Bench with three generative engines, FeatGEO beats token-level baselines and keeps or improves quality. The work matters because it shows that citation depends more on document-level properties than on surface wording.

Key points

• Optimization works at the topic level. Citation frequency is aggregated over a set of semantically related probe queries rather than tuned to one prompt.
• The 13 features include intro summary, heading level, list density, length, statistics, cited sources, quotations, unique info, authoritative tone, technical terms, readability, fluency and keyword focus.
• In the ablation, Statistics and Cite Sources add the most visibility. Structural features help moderately but consistently. Language features have mixed effects, and Unique Info can slightly hurt.
• The trade-off is clear when comparing the two endpoints of the Pareto front for education topics: 23.7% visibility with 87.8 quality versus 8.4% visibility with 92.7 quality.
• The advantage holds when the page generator is switched from GPT-4o-mini to Qwen3-4B, 8B or 14B. Code is released on GitHub (EvoNexusX/2026LiuFeatGEO).