Zikang Liu

Liu and Peilan Xu published Think Before Writing in the ACL 2026 long-paper track. The abstract states the problem plainly: generative answer engines expose content through selective citation rather than ranked retrieval, so optimization methods built for classic search are incomplete.

Existing GEO approaches, they write, often rely on token-level text rewriting with limited interpretability and weak control over the trade-off between citation visibility and content quality. FeatGEO is their answer to that gap.

FeatGEO abstracts webpages into interpretable structural, content, and linguistic properties, optimizes over that feature space, and uses a language model to realize chosen feature configurations as natural language. The design decouples high-level optimization from surface generation so practitioners can see which document-level levers moved.

Experiments on GEO-Bench across three generative engines are reported to improve citation visibility while maintaining or improving content quality versus token-level baselines. Further analyses in the paper claim citation behavior tracks document-level content properties more strongly than isolated lexical edits, and that learned feature configurations generalize across language models of different scales.

Public materials associated with the paper point to an open implementation under the FeatGEO / EvoNexusX line of work. Computational cost is non-trivial in reported runs.

The paper treats citation gains that hurt readability as failures, which is why it optimizes several objectives at once.