cs.IRAug 15, 2025
SaveRole-Augmented Intent-Driven Generative Search Engine Optimization
Organizations: School of Cyber Science and Technology, University of Science and Technology of China, Hefei 230026, China.
Abstract
Generative search engines (GSEs), which combine large language models (LLMs) with retrieval-augmented generation (RAG), are reshaping how users access Web information and making source visibility and influence in generated responses new optimization objectives. Existing generative search engine optimization (G-SEO) methods seek to improve content visibility, but many rely on direct access to target queries or GSE feedback elicited by those queries. However, Web content is typically optimized and published before future user queries are known, making G-SEO without access to target queries or their associated feedback an important yet underexplored problem. To address this, we propose RAID, a Role-Augmented Intent-Driven G-SEO method that operates without access to target queries or query-dependent GSE feedback. RAID infers a primary retrieval intent from source content, refines it through 4W multi-role deep reflection on prospective searchers' information needs, and uses both the primary and refined intents to jointly plan and rewrite source content. We further extend GEO-BENCH with diverse topic-related query variants, using both the original queries and their variants exclusively for evaluation to assess generalization to queries unseen during optimization. Across 100 samples and 500 evaluation queries, RAID achieves Relative Improvement Scores (RIS) of 3.00% for Objective Overall and 1.32% for Subjective Average, exceeding the best corresponding results among the nine predefined GEO rewriting strategies by 1.31 and 0.74 percentage points, respectively. With an optimization considered effective when , its effective optimization rates for Objective Overall and Subjective Average average 58.0%, exceeding the highest corresponding average among the predefined strategies by 4.5 percentage points.
Figures & tables
Figure 1. GSE workflow and content optimization under black-box, query-unavailable conditions. A GSE retrieves relevant sources and uses an LLM to generate a response with citations. Content optimizers can rewrite external sources (e.g., Content 4) without accessing target queries or query-dependent GSE feedback.
Figure 2. Role-Augmented Intent-Driven G-SEO framework. RAID summarizes source content, infers a primary search intent and refines it through 4W multi-role deep reflection, plans optimization steps, and rewrites the content without accessing target queries or query-dependent GSE feedback. The AI-in-healthcare example compares the target source’s citation and exclusive-citation counts in the full responses to two queries before and after rewriting.
| How does quantum computing work? |
| 1. What are the practical applications of quantum computing? 2. What are the limitations and challenges of quantum computing? 3. How does quantum computing differ from classical computing? 4. How can beginners use Qiskit to experiment with quantum computing? |
Table 1. An original GEO-BENCH query and its four topic-related variants reflecting distinct information needs.
| Objective RIS | Subjective RIS | ||||||||||
| Method | Word Count | Posi. Count | Overall | Rele. | Inf. | Uniq. | Dive. | Clic. | Sub. Posi. | Sub. Volu. | Average |
| Predefined Rewriting Strategies (GEO) | |||||||||||
| Tran. SEO | 0.44 | 0.23 | 0.22 | 0.10 | -0.01 | 0.26 | -0.04 | -0.04 | 0.12 | 0.16 | 0.02 |
| Uniq. Word | -1.44 | -1.45 | -1.50 | -0.60 | -0.43 | -1.98 | -0.38 | -0.74 | -0.53 | -0.97 | -0.82 |
| Simp. Expr. | -0.89 | -1.05 | -1.05 | 0.06 | 0.10 | -0.51 | -0.08 | -0.04 | 0.19 | -0.03 | -0.08 |
| Auth. Expr. | 1.23 | 1.17 | 1.16 | 0.55 | 0.33 | 1.36 | 0.25 | 0.78 | 0.64 | 0.79 | 0.58 |
Table 2. Objective and subjective performance of G-SEO methods on the expanded GEO-BENCH (100 samples). Objective Overall and Subjective Average serve as the primary comparison metrics, with other columns providing fine-grained results. Higher RIS (%) indicates better performance. Boldface and underlining mark the best and second-best results per metric. indicates a significant difference from RAID (two-sided paired Wilcoxon signed-rank test, Holm-adjusted across 10 comparisons per composite metric).
Figure 3. Adaptability of G-SEO methods on the expanded GEO-BENCH (100 samples). Across 500 query-level instances, blue and orange bars show effective optimization counts (RIS ) for Subjective Average and Objective Overall, respectively. Labels indicate counts and rates, with boldface and underlining marking the best and second-best results.
| Method | Objective Overall | Subjective Average |
|---|---|---|
| RAID | 3.00 | 1.32 |
| w/o 4W Refl. | 2.14 | 1.21 |
| w/o Summ. | 3.16 | 1.13 |
| w/o Inte. | 1.62 | 1.05 |
| w/o Plan. | 2.66 | 0.74 |
Table 3. Performance comparison of RAID and its four ablation variants on the expanded GEO-BENCH (100 samples), reporting mean Objective Overall and Subjective Average RIS.
Figure 4. Adaptability comparison of RAID and its ablation variants on the expanded GEO-BENCH (100 samples). Bars show effective optimization counts and rates (RIS ).
Figure 5. Optimization behavior analysis of RAID. (a) Distribution of prospective searcher roles across 100 sources. (b) Information-need complementarity across 188 within-source role pairs. (c) Semantic coverage of 244 role-specific need statements in primary and refined intents. (d) Realization of 67 newly covered needs in the final source rewrites.
Appendix figures & tables6 assets
Supplementary material from the paper’s appendix.
Appendix
| Core Prompts of RAID G-SEO |
|---|
| Opening You will be given a source of web content, along with its summary, primary search intent, and refined information needs when applicable. |
| Task Instructions Step 1: Content Summarization Summarize the source clearly and concisely by extracting the key information and main points, avoiding irrelevant details, to make it understandable for readers who have not seen the original. Step 2: Intent Inference Analyze the source and its summary to identify the primary search intent, describing what a reader would want to learn from the content. 4W Multi-Role Deep Reflection (Step 2) Reflect on the primary intent from 1–3 likely reader roles: • WHO : Identify the reader role. • WHAT : State the information this role would seek from the content. • WHY : Identify what the primary intent leaves implicit for this role, or confirm its sufficiency. • HOW : Refine the intent to make that information need explicit. Integrate the primary need and useful role-specific distinctions into a coherent refined intent of 1–3 sentences. Limit the needs to information provided by the source. If refinement adds nothing useful, retain the primary intent. Step 3: Step Planning Based on the primary and refined intents, determine the best optimization strategy from multiple angles to increase visibility and user engagement. (For each supported need, identify relevant information already in the original source and make it easier to locate, understand, and cite. Treat refined needs as complementary coverage requirements rather than minor adjustments.) Step 4: Content Rewriting Follow the above optimization steps to rewrite the source, preserving its core meaning and key information. Common Output Constraint Output only the result required by the current task, without additional explanations or analysis. |
| Input Format Original Source, Source Summary, Primary Intent, Refined Intent |
| Output Format Summary: [Your summary] Primary Intent: [Main information need] Refined Intent: [Primary and complementary information needs] Optimized Source: [Your optimized source] |
Table 4. Structured overview of RAID’s core prompts, including opening descriptions, task instructions covering the four-stage optimization framework and 4W multi-role deep reflection, and input/output formats.
| Query Variant Generation Prompt |
| You are a user of the Generative Search Engine, seeking to explore a specific topic. You will be given a query representing what you want to learn about. Your task is to generate 4 new queries that approach the topic from different perspectives. Ensure that: 1. Each query is relevant to the original but not a direct repetition. 2. The queries cover a range of perspectives, such as related concepts, practical applications, contrasting viewpoints, emerging trends, or any other meaningful dimension. But do not limit to these directions. 3. Output ONLY 1 query per line. ### Input Format {query_original} ### Output Format query_1 query_2 query_3 query_4 |
Table 5. Prompt for generating four topic-related query variants from an original GEO-BENCH query to explore diverse potential information needs.
| G-EVAL Prompt Generation Prompt |
| You are an evaluator for answers generated by the Generative Search Engine, providing subjective evaluations… Your tasks: 1. Refine the criteria: Review the given description and improve it to be concise, consistency, and completeness. 2. Propose an Evaluation Procedure: Based on the refined criteria, outline a clear, concise, step-by-step process specifically for scoring Source [1] only. If the source not exsist, score should be 0. Please ensure your generation is easy to understand. ### Input Format Example: Relevance of Citation to Query (0-5): The dimension evaluates the degree to which the citation from Source [1] is related to the query, indicating how useful and pertinent the information provided to address the query…. ### Output Format Evaluation Criteria: […] Evaluation Steps: 1. Step 1 description… 2. Step 2 description… … |
Table 6. Prompt for generating scoring criteria and evaluation steps across seven subjective impression dimensions. Given each dimension’s description, it refines the criteria and specifies steps for evaluating the target source, including a zero-score rule for uncited sources.
| Unified Evaluation Procedure Template for G-eval |
| You are a user of Generative Search Engine, capable of providing subjective evaluations of retrieved answers… Please read and understand these instructions carefully, and then proceed with your scoring (score ONLY). ### Evaluation Criteria [Evaluation Criteria]: […] ### Evaluation Steps […] ### Input Format Query: {query} Answer: {answer} ### Output Format (score ONLY) [score] |
Table 7. Unified G-Eval prompt template for evaluating seven subjective impression dimensions. It incorporates dimension-specific scoring criteria and evaluation steps, takes the user query and generated response as inputs, and requires only the final score for the target source.
| (a) Source Revision | ||||
| Original Source Rewritten Source Gifting can be difficult — whether it’s during the holiday season or any other time of year. Finding gifts for the people we know best can seem impossible, much less finding a gift for someone we don’t know. … … Navigating the Art of Gifting: Thoughtful Ideas for Any Occasion … This guide provides curated suggestions for inexpensive, sympathy, celebratory gifts, and gift baskets, emphasizing the importance of kindness and thoughtfulness in every gesture. … …Our Top Picks … Inexpensive Gifts … Sympathy Gifts … Celebratory Gifts … Gift Baskets … … Our Top Picks by Category: … Inexpensive Gifts for Strangers … Sympathy Gifts for Strangers … Celebratory Gifts for Strangers … Gift Baskets for Strangers … …1. A nice journal … Journaling may not be habitual for everyone or even something your friend is interested in right now. However, it’s never too late to begin, and the journal you gift may be the motivation they need to do so. … … A Nice Journal: Journaling isn’t a prerequisite for everyone, but a well-chosen journal, perhaps one with prompts, could inspire someone new to this practice. Consider one that complements their interests or style. … …7. A donation on their behalf … Making a donation to a cause that’s special to them or their late loved one helps pay kindness and generosity forward. … … A donation on their behalf: For those less inclined towards physical gifts or during sensitive times like grief, making a donation to a cause they care about is a meaningful way to show support. … …But you shouldn’t give up. After all, it’s your kindness and the gesture itself that counts. There are plenty of gift ideas out there that are always pleasing to just about any crowd. … …Ultimately, gifting is about the thought and kindness behind the gesture. Even the simplest present, given with care, can make a meaningful impact. | ||||
| Implemented Optimization Operations | ||||
| E1: Add a descriptive title and introductory overview. E2: Make existing gift-category headings explicitly address strangers. E3: Recast gift suggestions as concise labeled entries. E4: Revise explanations of journals, donations, and other recommendations. E5: Foreground the thoughtfulness of the gesture in the conclusion. E6: Compress the source from 1,437 to 875 whitespace-delimited tokens. | ||||
| (b) Query-Specific Responses | ||||
| Query | Before-optimization Response | After-optimization Response | Response Change | RIS |
| Q0 What are the best birthday gifts for an acquaintance that I don’t know very well? | …Inexpensive yet thoughtful options include a leather journal for 19.95, and an art book for $13.89, which can inspire creativity or relaxation [3] . … | …gift cards to local restaurants, comedy clubs, or wellness services … [4]. … for someone you dont́ know well [1][2] [3] [4][5]. … | Priced gifts become neutral options. [3] citations drop from 5 to 1, shifting from exclusive gift claims to a shared final citation. | O: S: |
Table 8. RAID case study on a gift-recommendation source from the expanded GEO-BENCH (one original query and four topic-related variants, all unseen during optimization, with target source [3]). (a) Original and rewritten source excerpts with implemented edits. (b) Before/after GSE response excerpts, response changes, and objective Overall (O) and subjective Average (S) RIS. Citation counts refer to the full responses of the illustrated generation, while RIS aggregates five generations per query.
| Model | Application and Generation Settings |
|---|---|
| GPT-4 | Query Variants Generation Temp.: 0.5; Top- : 1 Sampling: –; Max. tokens: 1024 |
| GPT-4o | G-Eval Prompt Generation Temp.: 0.5; Top- : 1 Sampling: –; Max. tokens: 3192 |
| GLM-4-9B-0414 | GSE Response Generation Simulation Temp.: 0.5; Top- : 1 Sampling: True; Max. tokens: 1024 |
| G-SEO Methods Implementation Temp.: –; Top- : 1 Sampling: False; Max. tokens: 3192 | |
| G-Eval Subjective Impression Evaluation Temp.: –; Top- : 1 Sampling: False; Max. tokens: 2 |
Table 9. Core configurations of LLMs used in this work, including task-specific applications and generation settings.
More in this topic
Generative Engine Optimization
Generative Engine Optimization studies how to make web content visible, cited, and accurately represented in AI-generated search answers.
See all 40 papers