ColdGEO: Feature-Aware Cold-Start Source Selection in Generative Search Engines

Abstract

Generative search engines and AI answer engines must decide which web sources to cite in response to a query, but a large share of candidate sources have little or no prior citation history, making cold-start source ranking a critical and underexplored problem for generative engine optimization (GEO). We present ColdGEO, a feature-aware ranking model for cold-start source selection that combines content-structure signals, readability metrics, and query–passage semantic similarity within a gradient-boosted LambdaRank framework (LightGBM and XGBoost). On the GEO-bench benchmark, ColdGEO attains a cold-start NDCG@5 of 0.6174, significantly outperforming cosine-similarity, BM25, and structure-only baselines (Wilcoxon signed-rank test, p = 0.0202, n = 974 queries). An ablation study identifies and removes four features with negligible contribution to cold-start accuracy (heading count, citation markers, domain authority, intro-summary presence) without measurable loss in ranking quality. To assess whether these gains generalize beyond a single corpus, we conduct a cross-dataset validation on MS MARCO v2.1 and find that the GEO-bench-trained model transfers poorly out of the box (cold NDCG@5 = 0.2861), underperforming even a parameter-free cosine-similarity baseline (0.6968). A per-feature distribution-shift analysis attributes this failure chiefly to a citation-frequency feature that encodes dataset-specific sampling artifacts rather than transferable relevance signal, compounded by a 15× scale mismatch in passage length between the two corpora. We introduce two domain-agnostic replacements — a locally-computed, per-query TF-IDF similarity in place of citation history, and log-normalized length together with length-relative structural density — which reduce the citation feature’s distributional shift by roughly an order of magnitude (4.46σ to 0.44σ) and improve MS MARCO cold-start NDCG@5 by 12.2% relative (0.2861 to 0.3211), while leaving GEO-bench performance and statistical significance unchanged. The residual gap is informative in its own right: the domain-adapted model remains statistically indistinguishable from a structure-only baseline on MS MARCO (p = 0.3865), indicating that cold-start ranking gains are real but corpus-dependent. We characterize which feature classes transfer reliably across domains (semantic similarity) and which do not (raw length, citation history), offering practical guidance for building cold-start source rankers that generalize.