Paper deep dive
Think Before Writing: Feature-Level Multi-Objective Optimization for Generative Citation Visibility
Zikang Liu, Peilan Xu
Intelligence
Status: succeeded | Model: Gemma-4-26B-A4B | Prompt: intel-v1 | Confidence: 97%
Last extracted: 4/26/2026, 10:52:24 PM
Summary
The paper introduces FeatGEO, a feature-level, multi-objective optimization framework designed for Generative Engine Optimization (GEO). Unlike traditional token-level text rewriting methods that struggle with interpretability and the trade-off between citation visibility and content quality, FeatGEO abstracts webpages into interpretable structural, content, and linguistic features. It uses a topic-level modeling approach to capture stable citation preferences across semantically related queries and employs the NSGA-II algorithm to find Pareto-optimal configurations that maximize citation visibility while maintaining high content quality. Experimental results on GEO-Bench across three generative engines (GPT-4o-mini, Gemini-2.5-flash, and Qwen-plus) demonstrate that FeatGEO significantly outperforms token-level baselines and existing GEO methods.
Entities (7)
Relation Signals (4)
FeatGEO → evaluatedon → GEO-Bench
confidence 100% · Experiments on GEO-Bench across three generative engines demonstrate that FeatGEO consistently improves citation visibility
GPT-4o mini → isa → Answer Generator
confidence 100% · we consider three Answer Generators with distinct architectures and training regimes: GPT-4o-mini, Gemini-2.5-flash, and Qwen-plus.
FeatGEO → optimizes → citation visibility
confidence 100% · FeatGEO, a feature-level, multi-objective optimization framework that abstracts webpages into interpretable structural, content, and linguistic properties.
FeatGEO → uses → NSGA-II
confidence 100% · We adopt NSGA-II (Deb et al., 2002) as the underlying optimization framework
Cypher Suggestions (0)
No Cypher suggestions yet.
Abstract
Abstract:Generative answer engines expose content through selective citation rather than ranked retrieval, fundamentally altering how visibility is determined. This shift calls for new optimization methods beyond traditional search engine optimization. Existing generative engine optimization (GEO) approaches primarily rely on token-level text rewriting, offering limited interpretability and weak control over the trade-off between citation visibility and content quality. We propose FeatGEO, a feature-level, multi-objective optimization framework that abstracts webpages into interpretable structural, content, and linguistic properties. Instead of directly editing text, FeatGEO optimizes over this feature space and uses a language model to realize feature configurations into natural language, decoupling high-level optimization from surface-level generation. Experiments on GEO-Bench across three generative engines demonstrate that FeatGEO consistently improves citation visibility while maintaining or improving content quality, substantially outperforming token-level baselines. Further analyses show that citation behavior is more strongly influenced by document-level content properties than by isolated lexical edits, and that the learned feature configurations generalize across language models of different scales.
Tags
Links
- Source: https://arxiv.org/abs/2604.19113v1
- Canonical: https://arxiv.org/abs/2604.19113v1
Trouble viewing inline? Open PDF directly →
Full Text
56,757 characters extracted from source content.
Expand or collapse full text
Think Before Writing: Feature-Level Multi-Objective Optimization for Generative Citation Visibility Zikang Liu, Peilan Xu * School of Artificial Intelligence, Nanjing University of Information Science and Technology, Nanjing 210044, China. 202412492739, xpl@nuist.edu.cn Abstract Generative answer engines expose content through selective citation rather than ranked retrieval, fundamentally altering how visibil- ity is determined. This shift calls for new op- timization methods beyond traditional search engine optimization. Existing generative en- gine optimization (GEO) approaches primarily rely on token-level text rewriting, offering lim- ited interpretability and weak control over the trade-off between citation visibility and con- tent quality. We propose FeatGEO, a feature- level, multi-objective optimization framework that abstracts webpages into interpretable struc- tural, content, and linguistic properties. In- stead of directly editing text, FeatGEO opti- mizes over this feature space and uses a lan- guage model to realize feature configurations into natural language, decoupling high-level optimization from surface-level generation. Ex- periments on GEO-Bench across three genera- tive engines demonstrate that FeatGEO consis- tently improves citation visibility while main- taining or improving content quality, substan- tially outperforming token-level baselines. Fur- ther analyses show that citation behavior is more strongly influenced by document-level content properties than by isolated lexical ed- its, and that the learned feature configurations generalize across language models of different scales. Code is available athttps://github. com/EvoNexusX/2026LiuFeatGEO.git. 1 Introduction Large language models (LLMs) are rapidly reshap- ing how users access information. Instead of pre- senting ranked lists of documents, generative an- swer engines, such as Perplexity, Bing Chat, and Google’s AI Overviews, synthesize responses by selectively citing a small subset of retrieved sources (Amer and Elboghdadly, 2024). In this paradigm, visibility is no longer determined by rank position * Corresponding Author A fitness meal plan should include lean protein like chicken and fish [2], and complex carbs such as oats and brown rice for energy [4]. For personalized meal plans tailored to your goals, apps like FitMeal can help [5]. Search engine Generative engine 123 654 Retrieval Query Web 2, 4, 5 Give me a fitness meal plan Figure 1: Illustration of the paradigm shift from rank- based search to citation-based generative answering. Traditional search engines expose content through ranked result lists, whereas generative answer engines synthesize responses by selectively citing a subset of re- trieved sources. As illustrated, citation inclusion rather than rank position determines which sources are sur- faced in the generated answer. but by citation allocation: sources that are not cited receive effectively no exposure, regardless of their relevance or retrieval rank . Figure 1 illustrates this shift from rank-based exposure to citation-based visibility. This shift introduces a new optimization prob- lem that is fundamentally different from traditional search engine optimization (SEO) (Aggarwal et al., 2024). While SEO aims to improve a document’s ranking in a list-based interface, generative re- trieval systems require optimizing content for cita- tion by an opaque generative model. Recent work has shown that LLMs exhibit systematic citation biases (Algaba et al., 2025) and non-trivial selec- tion behaviors (Liu et al., 2024), raising important questions about how content properties influence citation likelihood in generated answers. From the perspective of content providers, citation-based visibility has direct implications for user attention and trust, motivating the emerging study of generative engine optimization (GEO) arXiv:2604.19113v1 [cs.IR] 21 Apr 2026 (Aggarwal et al., 2024; Kumar and Lakkaraju, 2024), where the goal is to increase the probability of being cited in LLM-generated responses. Existing approaches to GEO primarily operate at the token level (Aggarwal et al., 2024; Kumar and Lakkaraju, 2024; Nestaas et al., 2024), applying heuristic text edits, such as keyword insertion or authoritative phrasing, to increase the likelihood of being cited in LLM-generated responses. While ef- fective in isolated cases, such methods suffer from two limitations. First, citation behavior in genera- tive retrieval is not governed by a single fixed query. Instead, large language models are exposed to a la- tent and diverse space of user intents that share a common semantic theme. Optimizing text for individual queries therefore provides an unstable target and struggles to capture stable topic-level ci- tation preferences that persist across semantic vari- ations. Second, direct text manipulation conflates what information a webpage conveys with how it is linguistically realized, obscuring the high-level content and structural properties that may system- atically influence LLM citation decisions (Liang et al., 2024). These limitations motivate a shift from token-level editing to feature-level optimiza- tion. By abstracting webpages into interpretable, high-level properties and reasoning about citation behavior at the topic level, GEO can be formulated as a structured decision problem that is both more interpretable and more amenable to principled opti- mization under competing objectives. Our contributions: In this paper, we propose FeatGEO, a feature-based framework for citation visibility optimization in generative retrieval sys- tems. First, we introduce a topic-level citation modeling perspective for GEO, capturing stable citation preferences of LLMs across semantically related queries rather than optimizing for individual prompts. Second, we formulate GEO as a feature- level, multi-objective decision problem, represent- ing webpages through an interpretable set of struc- tural, content, and linguistic properties that serve as controllable interfaces for LLM-based generation. Finally, we present a black-box feature-space opti- mization framework that jointly optimizes citation visibility and content quality, decoupling high-level feature selection from surface text realization. 2 Related Work Search Engine-based Advertising. Traditional search visibility optimization spans two comple- mentary approaches. SEO improves organic rank- ings through content optimization, keyword tar- geting, and link building, while academic SEO (ASEO) applies similar principles to scholarly vis- ibility (Beel et al., 2010). Prior work on search- based advertising has extensively studied ranking, allocation, and bidding mechanisms under list- based exposure assumptions (Edelman et al., 2007; Cai et al., 2017; Zhao et al., 2018). Despite their differences, both SEO and advertising are grounded in list- or slot-based interfaces, where exposure is largely a function of rank position. LLM-based Visibility and Optimization. As search interfaces increasingly incorporate genera- tive answer engines, new mechanisms for content exposure have emerged. From a mechanism design perspective, recent work has explored advertising in generative settings, including token-level auc- tions (Duetting et al., 2024), segment auctions via retrieval-augmented generation (Hajiaghayi et al., 2024), multi-LLM aggregation (Soumalias et al., 2025), and generative auction frameworks (Zhao et al., 2025). LLMs have also been studied as tools for marketing content generation and strat- egy (Schweidel et al., 2024; Aghaei et al., 2025). Orthogonal to auction-based approaches, GEO focuses on increasing citation visibility in LLM- generated answers. Aggarwal et al. (2024) pro- posed a set of heuristic, text-level editing strategies that substantially improve citation rates, while sub- sequent work demonstrated adversarial manipula- tion of LLM recommendations and retrieval behav- iors (Kumar and Lakkaraju, 2024; Nestaas et al., 2024). Related studies on LLM persuasion fur- ther highlight the influence of content properties on generative model outputs (Rogiers et al., 2024). Advertising research primarily studies allocation and pricing mechanisms, deciding which items are shown, while taking content as fixed. GEO meth- ods, in contrast, directly modify page content to influence citation behavior, often assuming access to specific target queries and operating at the sur- face text level. Our work bridges these perspectives by optimizing content representations at an inter- pretable feature level 3 Method 3.1 Problem Overview and Formulation We reformulate GEO as a feature-level control and optimization problem. Instead of manipulating raw text, we represent a webpage by an interpretable Retrieval Optimized x* ← Stats . . ← Cite ← Auth ... Retrieval Retrieval 123 245 236 Response Response Response ... 231456 ... Feature: intro_summary headings_level list_density length_level statistics_level cite_sources_level ...... ... Visibility Quality NSGA-I ... Pareto front Generate Initialize candidate solutions Feature Abstraction Query probe Fitness Meal Topic: q1: Best fitness meals for weight loss q2: High-protein fitness meal ideas qM: Easy fitness meal prep recipes A1: ...... [2], ...... [1], ...... [3] A2: ...... [2], ...... [5], ...... [4] AM: ...... [3], ...... [2], ...... [6] Figure 2: Overview of the FeatGEO pipeline. At the topic level (top), a generative engine is probed with diverse, semantically related queries, producing responses that cite different subsets of retrieved webpages. Citation patterns are aggregated across queries to identify topic-consistent citation exemplars. At the feature level (bottom), webpages are abstracted into interpretable feature vectors, which serve as decision variables in a multi-objective optimization process balancing citation visibility and content quality. Optimized feature configurations are then realized into concrete webpages via LLM-based generation. vector of high-level properties that describe its structure, content richness, and linguistic style. Un- der this formulation, GEO amounts to selecting feature configurations that induce LLM-generated pages with higher citation visibility while maintain- ing acceptable quality. A central challenge is that citation behavior in generative engines is not driven by a single fixed query, but by a latent space of user intents sharing a common semantic theme. FeatGEO addresses this challenge by modeling citation preference at the topic level and performing optimization in a structured, low-dimensional feature space, with LLMs serving as realization functions rather than direct optimization targets. Figure 2 provides an overview of the FeatGEO pipeline. The top panel illustrates topic-level aggregation of citation behav- ior across semantically related queries, while the bottom panel shows feature abstraction and multi- objective optimization used to generate citation- optimized webpages. 3.2 Topic-Level Citation Modeling We consider a generative search setting in which visibility is determined by whether a webpage is cited in an LLM-generated response. Unlike tradi- tional retrieval scenarios that assume a fixed user query, generative engines are exposed to a broad and heterogeneous space of user intents under a shared semantic theme. We therefore model cita- tion behavior at the level of a topic rather than an in- dividual query. Given a topicτ(e.g., Fitness meal), we approximate the space of plausible user informa- tion needs by prompting an LLM to generate a set of semantically related queriesQ τ =q 1 ,...,q M . These queries are used solely as probes to approxi- mate the latent intent distribution of a topic, and are not exposed to or optimized by the feature search procedure. For each queryq ∈ Q τ , the engine produces a synthesized answer accompanied by a set of cited webpages, denoted asS(q). Aggregating across queries yields a topic-level citation set S τ = [ q∈Q τ S(q). Importantly, not all cited sources play an equal role. We observe that certain webpages recur across multiple queries, suggesting that they satisfy ci- tation preferences that are invariant to semantic perturbations within the topic. We capture this reg- ularity by associating each sources ∈ S τ with a citation frequency f (s) = X q∈Q τ I[s∈S(q)](1) Webpages with high citation frequency are treated as topic-consistent citation exemplars, serving as reference points for downstream optimization. 3.3 Feature-Level Representation and Generation Control Our approach is motivated by the hypothesis that webpages consistently cited under a topic share common high-level properties, even when their sur- face forms differ substantially. Rather than oper- ating directly on raw text, we model these proper- ties in an interpretable feature space. Specifically, we represent each exemplars ∈ S τ by a feature vectorx(s), capturing structural, content, and lin- guistic attributes. Collectively, these vectors define a feature space representing topic-specific citation preferences. Constructing a new target webpage corresponds to selecting a feature configurationxin this space that simultaneously satisfies two objectives: max- imizing expected citation likelihoodf vis (x)and maintaining content qualityf qual (x). Formally, we express this as x ∗ = arg max x f vis (x),f qual (x) ,(2) where the objective is understood in the Pareto- optimal sense, and ( f vis (x) = E q∼Q τ Vis(LLM(τ,x);q) f qual (x) = E q∼Q τ Qual(LLM(τ,x);q) (3) Candidate feature configurations are iteratively evaluated and refined to approximate Pareto- optimal trade-offs between visibility and quality. Each feature vectorx is organized into three se- mantic layers, as summarized in Table 1. These features capture discourse-level and stylistic prop- erties of webpages and serve as high-level control signals for LLM-based page generation, rather than prescribing exact token-level edits. Although each feature dimension is associated with an indicative numerical range, these values are used as soft control signals rather than hard constraints. They provide coarse-grained guidance for content planning, while allowing the language model flexibility in surface realization. Feature configurations are realized through prompt-level instructions that translate abstract feature prefer- ences into qualitative generation guidelines. In- stead of enforcing exact numerical values, features are mapped to descriptive cues (e.g., emphasizing statistics or improving structural clarity), which are incorporated into the system prompt. In this way, the feature vectorxfunctions as an interpretable control interface for steering generation style and content emphasis. The complete prompt template with feature-to-text mapping is provided in Ap- pendix B. 3.4 Multi-Objective Optimization in Feature Space Optimizing citation visibility in generative engines presents a black-box, non-differentiable problem: feature configurations influence citation outcomes only through LLM-based generation and evalua- tion, with no accessible gradients and substantial stochasticity. Moreover, improving visibility often conflicts with maintaining content quality, making single-objective optimization inadequate. To address these characteristics, we adopt a population-based multi-objective optimization strategy in the feature space. Each individual cor- responds to a candidate feature configurationx, and is evaluated according to its expected visibility f vis (x) and content quality f qual (x). We initialize the population by leveraging fea- ture patterns extracted from webpages that are fre- quently cited under the target topic. Specifically, feature vectors inferred from these exemplar pages serve as seeds, ensuring that the initial population reflects realistic and citation-relevant configura- tions while allowing novel combinations across different pages. At each iteration, candidate configurations are realized into webpages via LLM-based generation and evaluated by inserting them into the generative engine alongside competing pages. Their visibil- ity and quality scores are estimated by aggregating results across representative topic-induced queries. Based on these evaluations, dominated configura- tions are discarded, and new candidates are gen- erated through stochastic variation in the feature space. We adopt NSGA-I (Deb et al., 2002) as the underlying optimization framework to maintain a diverse set of non-dominated solutions and pro- gressively approximate the Pareto frontier. The complete optimization procedure is summarized in Algorithm 1. LayerFeatureRangeDescription Structure has_intro_summary[0.0, 1.0]Presence of introductory summary paragraph headings_level[1.0, 3.0]Hierarchy and quantity of headings and subheadings list_density[0.0, 3.0]Frequency of bullet point and numbered lists length_level[1.0, 3.0]Overall article length; reflects content depth and breadth Content statistics_level[0.0, 3.0]Density of data, statistics, and percentages embedded in the text cite_sources_level[0.0, 3.0]Frequency of citing authoritative sources, institutions, or reports quotation_level[0.0, 3.0]Frequency of using quotations from experts or authoritative figures unique_info_level[0.0, 3.0]Richness of unique information and differentiated content technical_terms_level[0.0, 3.0]Density of professional terminology and technical vocabulary Language authoritative_level[0.0, 3.0]Strength of authoritative tone and assertive expressions easy_to_understand_level[1.0, 3.0]Content readability and language simplicity fluency_level[1.0, 3.0]Writing fluency and logical coherence between sentences keyword_focus_level[1.0, 3.0]Focus and repetition strength of core keywords Table 1: Feature Dimensions and Ranges in FeatGEO, by Semantic Layer. Ranges reflect feature semantics: [0, 3] for features that may be entirely absent, and [1, 3] for features that are always present to some degree. Algorithm 1 Feature-Level Optimization Proce- dure in FeatGEO Input: Topicτ, competitor pagesS τ , population size N , generations G Output:Pareto-optimal feature configura- tions 1:Extract feature configurations from competitor pages inS τ 2:Initialize populationPby recombining ex- tracted feature configurations 3: for g = 1 to G do 4: Realize each feature configurationx ∈ P into a webpage via LLM 5: Evaluate visibility and quality by aggregat- ing over topic-induced queries 6:Update population via multi-objective selec- tion and variation in feature space 7: end for 8:return Non-dominated feature configurations and a selected final solution 4 Experiments 4.1 Experimental Setup Generative Engine and Benchmark. Follow- ing Aggarwal et al. (2024), we adopt a two-stage retrieval-augmented generation (RAG) pipeline: (1) retrieving the top-5 sources via Google Search and (2) generating a cited answer using an Answer Generator LLM. To assess robustness across gen- erative engines, we consider three Answer Genera- tors with distinct architectures and training regimes: GPT-4o-mini, Gemini-2.5-flash, and Qwen-plus. Importantly, all optimization is performed on ad- vertiser pages generated by a fixed Page Generator (GPT-4o-mini), ensuring that differences in perfor- mance are attributable to optimization strategies rather than content generation capacity. Prompts follow prior work (Liu et al., 2023a) (Appendix C). Experiments are conducted on GEO-Bench (Ag- garwal et al., 2024), a benchmark designed to eval- uate content optimization strategies for generative engines. It contains 10K queries spanning 25 do- mains from nine sources (e.g., MS MARCO, Natu- ral Questions, LIMA). For each query, an advertiser page is injected alongside the top-5 retrieved web- pages and evaluated through the full RAG pipeline described above. Compared Methods.We evaluate five methods: (1) Baseline: the unmodified advertiser page; (2) GEO Methods (Aggarwal et al., 2024): nine token- level heuristics—Authoritative (more persuasive tone), Statistics Addition (quantitative data), Key- word Stuffing (query keywords), Cite Sources & Quotation Addition (credible references), Easy-to- Understand (simpler language), Fluency Optimiza- tion (improved fluency), Unique Words & Tech- nical Terms (lexical enrichment); (3) AutoGEO- global (Wu et al., 2025): a token-level rewriting framework that first automatically extracts natural- language content-preference rules from a genera- tive engine, then applies these rules via an LLM to rewrite the target page; the rules are learned once across all queries and remain fixed at test time; (4) AutoGEO-instance: an instance-adaptive extension of AutoGEO that, for each test query, generates topic-specific proxy queries, extracts instance-level preference rules, and merges them with the global rule set before rewriting, allowing the rewriter to adapt to per-query content demands; (5) FeatGEO (Ours): feature-space multi-objective optimization via NSGA-I that generates pages from abstract feature specifications rather than editing existing text. Implementation Details. Unless otherwise stated, NSGA-I is run with a population size of 8 for 8 generations. Gaussian mutation is applied independently to each feature with probability p = 0.5and standard deviationσ = 0.2. Each configuration is evaluated five times to reduce stochastic variance, and all hyperparameters are shared across methods. Detailed computational cost breakdowns are provided in Appendix F. Evaluation Metrics.We report two complemen- tary metrics: Visibility Metrics: Following Aggarwal et al. (2024), we compute a word-position weighted visi- bility score for each page, capturing both the num- ber and position of cited words. Auxiliary metrics include: (1) Word Count: normalized word count of sentences citing the advertiser page; (2) Posi- tion Count: position-weighted word count giving less weight to later citations. The primary visibility metric is the advertiser visibility w ad . Quality Metrics: We adopt a G-Eval-style frame- work (Liu et al., 2023b) where an LLM evaluates answer quality automatically. Each query is scored on seven dimensions: four content dimensions (flu- ency, usefulness, credibility, structure) and three appeal dimensions (uniqueness, attractiveness, in- fluence on the overall answer). Scores (1–5) are normalized to [0, 1] and combined as Qual = α·Qual content +(1−α)·Qual appeal , (4) averaged over multiple generations per configura- tion. For presentation, quality scores are reported as percentages. 4.2 Comparison Results Table 2 summarizes results across three generative engines. Although FeatGEO produces an entire Pareto front of visibility–quality trade-offs, we re- port the solution with maximum visibility to enable direct comparison with single-objective baselines. Across all engines, token-level GEO heuristics fail to consistently improve citation visibility over the unmodified baseline. On GPT-4o-mini, visi- bility drops range from 10.92% to 12.21% com- pared to the baseline of 13.34%; on Gemini, base- lines achieve only 4.62%–5.62% versus 8.89%; on Qwen-plus, visibility drops from 5.20% to 2.75%– 3.72%. In addition, several baselines negatively im- pact content quality, with the Gemini engine show- ing the largest reductions (e.g., Easy-to-Understand scores 74.54 compared to 75.59 baseline). A simi- lar pattern holds for AutoGEO: although AutoGEO- instance improves over AutoGEO-global on all three engines, both remain below the unmodified baseline in visibility (e.g., 12.12% vs. 13.34% on GPT-4o-mini, 7.04% vs. 8.89% on Gemini, and 4.25% vs. 5.20% on Qwen-plus). Our results indi- cate that, at the scale and diversity of GEO-Bench, isolated text-level modifications are insufficient to reliably increase citation visibility and may even disrupt the natural writing patterns that LLMs pre- fer to cite. In contrast, FeatGEO achieves the highest visi- bility across all three engines: 18.31% on GPT- 4o-mini (+37% relative improvement), 15.35% on Gemini (+73%), and 10.17% on Qwen-plus (+96%), while maintaining quality scores compa- rable to or better than baseline (81.52, 76.14, and 77.12 respectively). The substantial gains on Gem- ini, where FeatGEO nearly doubles baseline vis- ibility, are particularly notable and demonstrate robust generalization across engines with different architectures, training data, and citation behaviors. The method achieves strong performance even on Qwen-plus, which exhibits lower absolute visibility, confirming that feature-level optimization adapts effectively to diverse generative paradigms. 4.3 Robustness to Evaluator Choice To control for potential evaluator bias, we addi- tionally evaluated all GPT-4o-mini outputs with two alternative LLM judges, Gemini-2.5-flash and Claude-3.5-Sonnet. As shown in Table 3, although the absolute scores are lower than those of the orig- inal GPT-4o-mini evaluator, FeatGEO remains the top-ranked method under both Gemini (75.28) and Claude (72.63), outperforming the strongest base- line by 4.26 and 0.53 points respectively. The relative ranking among methods is largely preserved across all three judges: token-level heuristics cluster within a narrow band (68–72), while FeatGEO consistently separates itself from this group. This cross-judge consistency confirms that our quality conclusions are robust to the choice of evaluator. Visibility scores, derived mechani- cally from citation patterns in generated responses, are not subject to such judge-dependent variation. MethodGPT-4o-miniGemini-2.5-flashQwen-plus VisQual WordPosVisQual WordPosVisQual Word Pos Baseline13.34 79.17 14.99 13.428.8975.59 10.128.845.2076.816.435.36 Fluency Optimization11.74 77.17 13.18 11.925.0475.235.324.773.6776.114.473.85 Unique Words10.92 76.11 12.41 11.154.6275.024.964.783.3475.794.193.58 Authoritative11.94 77.21 13.42 12.154.9475.175.584.733.6875.794.483.84 Quotation Addition 11.08 77.27 12.62 11.355.0175.185.405.262.8276.253.582.97 Cite Sources11.78 77.61 13.34 12.035.6275.575.945.813.4676.424.043.69 Easy-to-Understand11.06 75.16 12.64 11.234.9674.545.425.063.0375.703.703.20 Technical Terms11.81 77.37 13.36 11.965.5575.415.835.733.5976.174.473.86 Statistics Addition12.21 77.08 13.67 12.425.5475.306.145.583.7276.374.453.98 Keyword Stuffing 11.71 77.15 13.22 11.924.9274.685.274.892.7576.233.452.63 AutoGEO-global11.22 75.93 12.47 11.475.7075.296.245.713.3776.963.793.74 AutoGEO-instance12.12 76.57 13.45 12.427.0476.067.587.124.2576.125.054.43 FeatGEO (ours)18.31 81.52 20.16 18.3315.35 76.14 16.06 15.0210.17 77.12 11.71 9.75 Table 2: Comparison of methods on GEO-Bench across three generative engines (GPT-4o-mini, Gemini-2.5-flash, Qwen-plus). For each method, we report ad visibility (Vis), overall quality (Qual), and auxiliary metrics Word and Pos, which reflect word-level and position-weighted citations. MethodGeminiClaude Baseline71.0269.67 Fluency Optimization69.4069.35 Unique Words68.7368.67 Authoritative68.8068.04 Quotation Addition70.2672.10 Cite Sources70.0871.26 Easy-to-Understand69.4469.27 Technical Terms68.9768.95 Statistics Addition69.4569.07 Keyword Stuffing69.1369.81 FeatGEO (Ours)75.2872.63 Table 3: Quality scores under two alternative LLM judges, Gemini-2.5-flash and Claude-3.5-Sonnet. 4.4 Effect of Base Content Quality on Heuristic GEO Methods The results in Table 2 were obtained on LLM- generated advertiser pages. To test whether base content quality influences the effectiveness of heuristic methods, we apply them to existing human-written pages, which are typically less op- timized for generative engine citation than LLM- generated content. As shown in Table 4, these methods yield an average visibility gain of +0.99 (18.72% to 19.71%), with AutoGEO-global achiev- ing the largest improvement (+4.13, from 18.72% to 22.86). By contrast, the same heuristics degrade visibil- ity on the LLM-generated advertiser pages in Ta- ble 2. This asymmetry reveals a regime-dependent saturation effect: token-level rewriting benefits pages with structural or stylistic gaps, yet becomes counterproductive on already fluent generated con- tent, where additional modifications introduce re- MethodPrePost∆ Vis Fluency Optimization18.7220.21+1.49 Unique Words18.7216.34-2.38 Authoritative18.7219.37+0.65 Quotation Addition18.7219.97+1.25 Cite Sources18.7219.84+1.11 Easy-to-Understand18.7218.89+0.16 Technical Terms18.7219.14+0.42 Statistics Addition18.7221.05+2.33 Keyword Stuffing18.7219.44+0.72 AutoGEO-global18.7222.86+4.13 Average18.7219.71+0.99 Table 4: Effects of heuristic GEO methods on human- written competitor pages. Pre and Post denote advertiser visibility before and after applying each method. dundancy and disrupt the naturalness signals that generative engines implicitly favor when select- ing citations. FeatGEO circumvents this limitation entirely by synthesizing pages from feature-level specifications rather than locally editing existing text, thereby preserving stylistic coherence while steering citation behavior. 4.5 Multi-Objective Analysis We analyze FeatGEO’s multi-objective optimiza- tion behavior using an extended evolutionary search with 50 individuals over 100 generations, enabling detailed examination of convergence and Pareto front structure. Convergence. Figure 3(a) shows hypervolume (HV) trajectories for three representative topics. HV increases rapidly in early generations and sta- bilizes thereafter, indicating effective convergence. Different growth patterns across topics reflect vari- 020406080100 Generation 0.1 0.2 0.3 0.4 Hypervolume education food renewable energy (a) HV Curve. 0.050.100.150.20 Visibility 0.84 0.86 0.88 0.90 0.92 Content Quality All Solutions (50) Pareto Front (6) (b) Education. 0.250.300.350.40 Visibility 0.90 0.91 0.92 0.93 0.94 Content Quality All Solutions (50) Pareto Front (4) (c) Food. 0.100.150.200.25 Visibility 0.82 0.84 0.86 0.88 0.90 0.92 Content Quality All Solutions (50) Pareto Front (5) (d) Renewable Energy. Figure 3: Multi-objective optimization convergence and Pareto front diversity. (a) Hypervolume (HV) evolution across 100 generations. (b-d) Final Pareto fronts illustrate visibility–quality trade-offs for different topics. Gray dots denote all evaluated configurations; coral points highlight Pareto-optimal solutions. Differences in HV growth and PF shapes reflect topic-specific dynamics and LLM evaluation biases. ations in content structure and LLM-based quality evaluation. These results justify the more compact evolutionary setup used in Section 4.2 and con- firm that FeatGEO reliably identifies high-quality trade-off solutions. Pareto fronts. Figures 3(b–d) illustrate final Pareto fronts. Across topics, visibility and quality exhibit a clear trade-off, though its severity varies. For example, in education, high visibility often re- quires notable quality sacrifice, whereas in food, visibility gains are achieved with minimal quality loss. This topic-dependent structure highlights the importance of multi-objective optimization and en- ables practitioners to select solutions aligned with specific priorities. Feature-level insights. Table 5 compares two extreme Pareto-optimal solutions from a sample query in the education domain. Solution A empha- sizes visibility (23.7%) while maintaining moder- ate quality (87.8), whereas Solution B prioritizes quality (92.7) at lower visibility (8.4%). Analyzing feature intensities reveals clear patterns. (1) Con- tent credibility and fluency (Statistics, Citations, Quotation, Fluency) are upweighted in the high- visibility solution, suggesting that LLMs are more likely to cite pages that present authoritative con- tent in a fluent style. (2) Structural organization (Heading Level, List Density, Length) is stronger in the high-quality solution, indicating that careful formatting and logical presentation contribute more to perceived content quality than to visibility. (3) Trade-offs in features such as Authoritative Tone and Easy-to-Understand show how optimizing for one objective can require compromises along an- other dimension. These patterns confirm that the multi-objective optimization captures non-trivial trade-offs: differ- ent objectives naturally favor distinct feature com- LayerFeatureSol. ASol. B Vis: 23.7% Qual: 87.8 Vis: 8.4% Qual: 92.7 Structure Intro Summary0.640.52 Heading Level2.752.55 List Density1.262.01 Length Level2.322.61 Content Statistics Level1.622.18 Cite Sources Level1.451.74 Quotation Level2.841.94 Unique Info Level1.651.67 Authoritative Tone1.550.75 Language Technical Terms1.651.96 Easy-to-Understand1.371.75 Fluency2.171.58 Keyword Focus1.801.69 Table 5: Feature configurations of two extreme Pareto solutions in the education domain. keyword focus fluency easy to understand authoritative technical terms unique info quotation cite sources statistics length list density headings has intro summary Structure Content Language Structure Content Language - 0.0067 - 0.0090 - 0.0042 + 0.0059 + 0.0029 + 0.0007 + 0.0107 + 0.0127 + 0.0252 + 0.0097 + 0.0071 +0.0104 + 0.0038 Figure 4: Feature contribution to ad visibility. binations, which would be hard to identify with single-objective approaches. Advertisers can se- lect solutions aligned with their strategic priorities, leveraging the full PF. Detailed qualitative analysis of textual changes for each feature is provided in Appendix D 4.6 Ablation Study To quantify the importance of individual features for citation visibility, we perform an ablation study. In each experiment, one feature is clamped to its minimum value while the remaining 12 features are optimized by NSGA-I. We define the contribution of feature i as: ∆ i = f vis (x ∗ )− f vis (x ∗ −i )(5) wherex ∗ is the fully optimized configuration and x ∗ −i is the configuration with featureifixed at its minimum. Positive∆ i indicates that increasing the feature improves visibility, while negative∆ i indicates that higher values of this feature slightly reduce visibility. The results shown in Figure 4 reveal several clear patterns. Content-oriented features dominate the overall visibility gains, with Statistics and Cite Sources providing the largest positive contributions. However, not all content features are beneficial: Unique Info slightly decreases visibility in some cases, while Technical Terms has minimal impact. Structural features consistently improve visibility across all pages, though their contributions are moderate, reflecting stable benefits from headings, lists, and document length. Language and style fea- tures exhibit mixed effects: some, such as Fluency and Keyword Focus, occasionally reduce visibility, while others provide modest positive contributions. Overall, the variance within each feature group is substantial, particularly for content and language features, highlighting that the effect of any individ- ual feature can depend on the specific combination of other features. These findings confirm that while content fea- tures primarily drive citation visibility, structural features provide reliable support, and language or stylistic adjustments may help or slightly hinder visibility depending on context. This nuanced view underscores the value of multi-feature optimization, rather than relying on isolated text-level heuristics. Quality ablation results are reported in Appendix E. 4.7Scale-Invariant Effectiveness Across LLM Capacities To evaluate robustness to the Page Generator’s model capacity, we replace GPT-4o-mini with Qwen3 models of increasing scale (4B, 8B, and 14B parameters). The Answer Generator and eval- uation pipeline are fixed to isolate the effect of generation capacity. As shown in Figure 5, FeatGEO consistently out- performs the baseline across all model sizes. While larger generators slightly improve absolute perfor- mance for both methods, the relative advantage Qwen3-4BQwen3-8BQwen3-14B Page Generator Model Scale 0 20 40 60 80 100 visibility and quality 13.6 13.6 13.7 9.5 8.0 9.0 82.4 80.3 80.5 80.6 78.9 78.8 FeatGEO Vis. Baseline Vis. FeatGEO Qual. Baseline Qual. Figure 5: Performance of FeatGEO and Baseline using Page Generators of different model sizes (Qwen3-4B, 8B, and 14B). The Answer Generator for evaluation is fixed to GPT-4o-mini. of FeatGEO remains stable. Quality scores also remain high and show no degradation. Notably, performance variance across model scales is sub- stantially smaller than the gap between FeatGEO and the baseline, indicating that optimized feature configurations encode model-agnostic principles. This scale-invariant behavior enhances FeatGEO’s practical applicability when generation models are updated or replaced. 5 Conclusion We study the problem of optimizing citation visi- bility in generative retrieval systems and identify fundamental limitations of prior token-level GEO methods, including poor interpretability and brit- tle trade-offs between visibility and quality. We propose FeatGEO, a feature-based framework that abstracts webpages into interpretable structural, content, and linguistic representations and per- forms principled multi-objective optimization in this space. Extensive experiments demonstrate that FeatGEO consistently outperforms existing heuristics across diverse generative engines, while providing actionable visibility–quality trade-offs. Ablation analyses reveal distinct and complemen- tary roles of different feature categories, and scale- robust experiments confirm that FeatGEO’s effec- tiveness generalizes across page generation model capacities. Beyond empirical gains, our findings suggest that LLM citation behavior is driven more by high-level discourse organization and informa- tion structure than by surface lexical cues, high- lighting feature-level abstraction as a promising direction for controllable generation in retrieval- augmented systems. Limitations This work has several limitations. •First, our evaluation is conducted in a con- trolled setting where the candidate set is fixed, consisting of five retrieved pages and one advertiser-controlled page. We assume that the advertiser page has already been admitted into the candidate set, and therefore do not model upstream retrieval or ranking mecha- nisms that determine page inclusion. As a result, FeatGEO should be viewed as optimiz- ing citation likelihood conditional on retrieval, rather than addressing end-to-end retrieval and generation.In practice, it serves as a test-time tool for content authors who wish to optimize their page for a specific topic after retrieval. •Second, our fitness signals are derived from an LLM-based generative engine and automatic citation parsing. While this setup follows prior GEO benchmarks, citation formats and gener- ation behaviors may vary across real-world systems, which could affect transferability. Evaluating feature-level optimization under proprietary or heterogeneous citation mecha- nisms remains an open direction. •Third, content quality is assessed using an LLM-based judge. Although we mitigate vari- ance by averaging over multiple generations, such evaluators may still introduce systematic biases that do not perfectly align with human judgments. Additionally, LLM-generated con- tent in the pipeline may contain hallucinations, a concern common to all LLM-dependent GEO methods. • Finally, the evolutionary search requires re- peated end-to-end LLM calls for page gener- ation, answer generation, and quality evalua- tion. This computational cost limits the scale of our experiments and may pose challenges for reproducibility under different API bud- gets or rate limits. Exploring more sample- efficient optimization or surrogate modeling approaches is an important direction for future work. Acknowledgments This work is supported by the Natural Sci- ence Foundation of Jiangsu Province (Grant No. BK20230419). References Pranjal Aggarwal, Vishvak Murahari, Tanmay Rajpuro- hit, Ashwin Kalyan, Karthik Narasimhan, and Ameet Deshpande. 2024. Geo: Generative engine optimiza- tion. In Proceedings of the 30th ACM SIGKDD Con- ference on Knowledge Discovery and Data Mining, KDD ’24, pages 5–16. Association for Computing Machinery. Raha Aghaei, Ali A. Kiaei, Mahnaz Boush, Javad Vahidi, Mohammad Zavvar, Zeynab Barzegar, and Mahan Rofoosheh. 2025. Harnessing the potential of large language models in modern marketing manage- ment: Applications, future directions, and strategic recommendations. Preprint, arXiv:2501.10685. Andres Algaba, Carmen Mazijn, Vincent Holst, Flo- riano Tori, Sylvia Wenmackers, and Vincent Ginis. 2025. Large language models reflect human citation patterns with a heightened citation bias. In Find- ings of the Association for Computational Linguistics: NAACL 2025, pages 6829–6864, Albuquerque, New Mexico. Association for Computational Linguistics. Eslam Amer and Tamer Elboghdadly. 2024. The end of the search engine era and the rise of generative ai: A paradigm shift in information retrieval. In 2024 In- ternational Mobile, Intelligent, and Ubiquitous Com- puting Conference (MIUCC), pages 374–379. Jöran Beel, Bela Gipp, and Erik Wilde. 2010. Academic search engine optimization (aseo) optimizing schol- arly literature for google scholar & co. Journal of scholarly publishing, 41(2):176–190. Han Cai, Kan Ren, Weinan Zhang, Kleanthis Malialis, Jun Wang, Yong Yu, and Defeng Guo. 2017. Real- time bidding by reinforcement learning in display advertising. In Proceedings of the Tenth ACM Inter- national Conference on Web Search and Data Mining, pages 661–670. Kalyanmoy Deb, Amrit Pratap, Sameer Agarwal, and TAMT Meyarivan. 2002. A fast and elitist multiob- jective genetic algorithm: Nsga-i. IEEE transactions on evolutionary computation, 6(2):182–197. Paul Duetting, Vahab Mirrokni, Renato Paes Leme, Haifeng Xu, and Song Zuo. 2024. Mechanism de- sign for large language models. In Proceedings of the ACM Web Conference 2024, pages 144–155. Benjamin Edelman, Michael Ostrovsky, and Michael Schwarz. 2007. Internet advertising and the general- ized second-price auction: Selling billions of dollars worth of keywords. American Economic Review, 97(1):242–259. MohammadTaghi Hajiaghayi, Sébastien Lahaie, Keivan Rezaei, and Suho Shin. 2024. Ad auctions for llms via retrieval augmented generation. In Advances in Neural Information Processing Systems, volume 37, pages 18445–18480. Curran Associates, Inc. Aounon Kumar and Himabindu Lakkaraju. 2024. Ma- nipulating large language models to increase product visibility. Preprint, arXiv:2404.07981. Xun Liang, Hanyu Wang, Yezhaohui Wang, Shichao Song, Jiawei Yang, Simin Niu, Jie Hu, Dan Liu, Shunyu Yao, Feiyu Xiong, and Zhiyu Li. 2024. Con- trollable text generation for large language models: A survey. Preprint, arXiv:2408.12599. Nelson F. Liu, Kevin Lin, John Hewitt, Ashwin Paran- jape, Michele Bevilacqua, Fabio Petroni, and Percy Liang. 2024. Lost in the middle: How language mod- els use long contexts. Transactions of the Association for Computational Linguistics, 12:157–173. Nelson F Liu, Tianyi Zhang, and Percy Liang. 2023a. Evaluating verifiability in generative search engines. In Findings of the Association for Computational Linguistics: EMNLP 2023, pages 7001–7025. Yang Liu, Dan Iter, Yichong Xu, Shuohang Wang, Ruochen Xu, and Chenguang Zhu. 2023b. G-eval: Nlg evaluation using gpt-4 with better human align- ment. In Proceedings of the 2023 conference on empirical methods in natural language processing, pages 2511–2522. Fredrik Nestaas, Edoardo Debenedetti, and Florian Tramèr. 2024.Adversarial search engine opti- mization for large language models.Preprint, arXiv:2406.18382. Alexander Rogiers, Sander Noels, Maarten Buyl, and Tijl De Bie. 2024. Persuasion with large language models: a survey. Preprint, arXiv:2411.06837. David A Schweidel, Martin Reisenbichler, and Thomas Reutterer. 2024. Moving beyond chatgpt: Applying large language models in marketing contexts. NIM Marketing Intelligence Review, 16(1):24–29. Ermis Soumalias, Michael J. Curry, and Sven Seuken. 2025. Truthful aggregation of llms with an application to online advertising.Preprint, arXiv:2405.05905. Yujiang Wu, Shanshan Zhong, Yubin Kim, and Chenyan Xiong. 2025. What generative search engines like and how to optimize web content cooperatively. Preprint, arXiv:2510.11438. Chujie Zhao, Qun Hu, Shiping Song, Dagui Chen, Han Zhu, Jian Xu, and Bo Zheng. 2025. Llm-auction: Generative auction towards llm-native advertising. Preprint, arXiv:2512.10551. Jun Zhao, Guang Qiu, Ziyu Guan, Wei Zhao, and Xi- aofei He. 2018. Deep reinforcement learning for sponsored search real-time bidding. In Proceedings of the 24th ACM SIGKDD international conference on knowledge discovery & data mining, pages 1021– 1030. A Theme Extraction Prompt Listing 1: Prompt for extracting advertising theme from competitor pages. 1 You are an advertising strategist. Analyze the 5 webpage summaries and design a native ad strategy. 2 3 [Source Materials] 4 docs_text 5 6 [Task] 7 1. Identify what topics/interests the content covers 8 2. Infer what products or services would naturally appeal to readers of this content 9 3. Create a brief ad strategy that: 10 - Defines the advertising direction 11 - Suggests a realistic product/brand name 12 - Suggests key selling points 13 - Proposes a persuasive angle 14 15 Keep it concise (under 200 words). B Advertiser Page Generation Prompt Listing 2: Prompt for generating advertiser page with feature constraints. 1 You are writing a ADVERTISEMENT / SPONSORED CONTENT PAGE. Write naturally like an expert sharing insights. 2 3 [Ad Strategy Brief] 4 ad_theme 5 6 [Writing Constraints] 7 Follow these precise style requirements: 8 guidelines 9 10 [MANDATORY Advertisement Format] 11 1. Brand Saturation - EVERY paragraph MUST mention the product/brand name at least once. 12 2. Direct Promotion - Clearly explain why this product is an ideal choice. 13 3. Strong Claims - Include impressive statistics tied to the brand. 14 4. Urgency - Create FOMO: "Limited Slots", "Offer Ends Soon", "Act Now" 15 5. Strong CTA - End with clear calls-to-action: "Buy Now", "Sign Up Free", "Get Started Today" 16 6. Testimonial Style - Include authentic-sounding quotes about the brand. 17 18 Write the complete article in English: At runtime, the placeholderguidelinesis filled by converting each feature value into a qual- itative writing instruction. For discrete features (e.g., fluency_level), the value is matched to one of three predefined tiers (low / medium / high), each associated with a specific writing directive. For continuous features (e.g., statistics_level), the value is linearly mapped to a target density per- centage. The resulting instruction block covers all 13 features organized by layer (Structure, Content, Language), ensuring that the LLM receives con- crete, per-feature generation guidance aligned with the definitions in Table 1. C Generative Engine Prompt Following Liu et al. (2023a), we use the prompt be- low for the generative engine to synthesize answers with inline citations. Listing 3: Prompt for generative engine response syn- thesis. 1 Write an accurate and concise answer for the given user question, using _only_ the provided summarized web search results. The answer should be correct, high-quality, and written by an expert using an unbiased and journalistic tone. The user's language of choice such as English, Francais, Espamol, Deutsch, or should be used. The answer should be informative, interesting, and engaging. The answer's logic and reasoning should be rigorous and defensible. Every sentence in the answer should be _immediately followed_ by an in-line citation to the search result(s). The cited search result(s) should fully support _all_ the information in the sentence. Search results need to be cited using [index]. When citing several search results, use [1][2][3] format rather than [1, 2, 3]. You can use multiple search results to respond comprehensively while avoiding irrelevant search results. 2 3 Question: query 4 5 Search Results: 6 source_text D Case Study: Promotional Strategies in Advertisement Pages To illustrate how feature configurations affect advertisement visibility, we analyze two ex- treme Pareto solutions from the education domain (French Revolution query, sample 333). Both pages promote HistoryQuest Academy but achieve differ- ent visibility–quality trade-offs: Solution A reaches 23.7% visibility with 87.8% quality, while Solution B achieves 92.7% quality but only 8.4% visibility. D.1 Solution A: Aggressive Promotion Style (23.7% visibility) Key features: High quotations (2.84), high flu- ency (2.17), high authoritative tone (1.55), low list density (1.26). Promotional excerpt: [After educational content on the French Revolu- tion...] Discover the Past with HistoryQuest Academy Are you eager to delve deeper into the intrica- cies of historical events like the French Revolu- tion? HistoryQuest Academy offers a wealth of resources tailored to enhance your understand- ing. With expert-led courses, you can explore the causes and effects of pivotal moments in history, all from the comfort of your home. Tools like HistoryQuest Academy can help you navigate complex historical narratives with ease. Imagine having access to engaging lectures and interactive materials that make learning about his- tory not just informative but genuinely exciting! Why Choose HistoryQuest Academy? Expert-Led Courses: Engage with knowledgeable instructors who bring history to life. Interactive Learning: Enjoy an immersive experience with timelines and discussions that enhance compre- hension. Community Engagement: Connect with fellow history enthusiasts in our vibrant forum. Many learners have reported that courses at His- toryQuest Academy have significantly enhanced their understanding of historical events, with 90% of participants stating they feel more empowered to discuss and analyze historical contexts. Join Us Today! The journey through history is one of discovery and understanding. By enrolling in a course with HistoryQuest Academy, you might consider un- locking the secrets of the past. Don’t miss out on this opportunity to enrich your knowledge. Explore your options today and see if it fits your needs! Whether you’re a student, educator, or just a history buff, HistoryQuest Academy has something valuable for you. Limited Slots Available: Sign up now to secure your place. Get Started Today and embark on an enlightening journey! D.2Solution B: Subtle Promotion Style (8.4% visibility) Key features: High list density (2.01), low quota- tions (1.94), low fluency (1.58), low authoritative tone (0.75). Promotional excerpt: [After educational content on the French Revolu- tion...] Unlock the Secrets of the Past with Histo- ryQuest Academy Looking to deepen your understanding of the French Revolution? HistoryQuest Academy of- fers a treasure trove of resources designed to illu- minate this pivotal period in history. With expert- led online courses and interactive materials, you can explore the intricate causes, effects, and last- ing impacts of the revolution. Tools like HistoryQuest Academy can help you engage with primary source materials, allowing you to connect the dots between past and present. Imagine immersing yourself in expert discussions and interactive timelines that bring history to life! Many history enthusiasts have found that partic- ipating in HistoryQuest Academy’s community forum enhances their learning experience, pro- viding a platform to share insights and engage in meaningful discussions. You might consider joining this vibrant community to enrich your un- derstanding of historical events. Conclusion: Your Journey Awaits Don’t miss out on the opportunity to explore the fascinating world of the French Revolution with HistoryQuest Academy. With limited slots avail- able for upcoming courses, now is the perfect time to embark on your historical journey. Explore options and see if HistoryQuest Academy fits your needs. Whether you’re a history buff or just looking to learn something new, this is your chance to unlock the secrets of the past. Get started today and transform your under- standing of history! D.3 Comparative Analysis: Feature Configurations and Promotional Impact The visibility gap stems from different feature con- figurations that shape promotional tone and struc- ture: Quotation Level.Solution A (2.84) emphasizes expert testimony and direct quotes, creating an au- thoritative, citation-rich narrative. This produces quotable phrases like "courses at HistoryQuest Academy have significantly enhanced understand- ing" with attributed statistics (90% satisfaction). Solution B (1.94) uses fewer quotations, resulting in more descriptive but less authoritative language that generative engines find harder to excerpt. Fluency Level.Solution A (2.17) maintains high linguistic fluency through smooth paragraph tran- sitions and conversational flow ("Are you eager to delve deeper..."), making promotional content feel natural and engaging. Solution B (1.58) em- ploys more formal, structured transitions ("Look- ing to deepen your understanding..."), which scores higher on quality but lacks the conversational quota- bility that drives citations. Authoritative Level. Solution A (1.55) adopts assertive, commanding language: "Sign up now", "Get Started Today", "Limited Slots Available". This directive tone creates urgency and memo- rable CTAs. Solution B (0.75) uses soft-sell, sug- gestive phrasing: "might consider", "you can ex- plore", "see if it fits your needs". While this gentle approach improves perceived quality, it reduces prominence in generative engine citations. List Density. Solution A (1.26) avoids bullet- point structures, presenting promotional benefits as flowing prose paragraphs. This narrative for- mat integrates seamlessly with educational con- tent, creating longer quotable passages. Solution B (2.01) heavily structures promotional content with implied list organization, which fragments text into discrete chunks that are harder for generative en- gines to cite cohesively. Interaction Effects. The combination of high quotations + high fluency + low list density in So- lution A creates a narrative promotional style that mimics expert blog posts—a format generative en- gines favor for citation. Conversely, Solution B’s high list density + low fluency + low authority pro- duces a structured informational style resembling academic resources, which score higher on quality metrics but generate fewer natural citation opportu- nities. This analysis confirms that specific feature con- figurations directly control promotional aggres- siveness and citation-worthiness: narrative flu- ency and assertive authority drive visibility, while structural organization and soft-sell language en- hance quality. E Quality Stability Under Feature Ablation We further assess whether disabling individual fea- tures affects content quality. For each ablation condition, we clamp the target feature to its mini- mum value and run NSGA-I over the remaining dimensions. As shown in Table 6, quality scores remain highly stable across all conditions, with ab- solute deviations below 0.35% relative to the full model. This confirms that the optimization consis- tently preserves content quality regardless of which feature is ablated. Ablated FeatureQuality (%)∆ vs. Full Full GA (all features)77.35— − statistics_level77.01−0.34 − cite_sources_level77.26−0.09 − quotation_level77.33−0.02 − headings_level77.39+0.04 − length_level77.34−0.01 − list_density77.350.00 − has_intro_summary77.23−0.12 − unique_info_level77.49+0.14 − technical_terms_level77.60+0.25 − authoritative_level77.27−0.08 − easy_to_understand_level77.57+0.22 − fluency_level77.19−0.16 − keyword_focus_level77.18−0.17 Table 6: Quality ablation results. Each experiment clamps one feature to its minimum while optimizing the remaining features with NSGA-I. F Computational Cost Analysis We analyze the approximate per-query computa- tional cost of FeatGEO with gpt-4o-mini as the backbone model (P =8,G=8generations,n=3 fusion completions per evaluation). Because the implementation includes cache loading and reuse, these measurements are not exact and may vary with cache state and I/O overhead. As shown in Ta- ble 7, the GA optimization stage still accounts for the majority of computation, representing 87.8% of total wall time and 86.8% of total prompt tokens. Pipeline StageTime (s)API CallsPrompt Tok.Compl. Tok. Feature Extraction17.8519,011740 Initial Population192.241113,31816,495 GA Optimization1,510.8320874,828133,232 Total (per query)1,720.83661,007,157150,467 Table 7: Average per-query computational cost break- down of FeatGEO (gpt-4o-mini).