acceptodds
Under review as a conference paper at ICLR 2027

StructGEO: Intent-Guided Structural Organization for Generative Engine Optimization

Abstract

Generative Engine Optimization (GEO) aims to rewrite a document so that generative engines are more likely to include its content in their responses. Recent studies infer potential user intents and then use them to guide rewriting. While effective, they have two limitations. First, because potential queries are inferred independently, they tend to concentrate on a few dominant search intents, resulting in insufficient coverage of less prominent information needs. Second, they focus on what evidence a document contains but overlook how that evidence is organized, even though organization affects whether a generative engine actually uses it. To address both issues, we propose StructGEO, a GEO framework that models search intents hierarchically and organizes supporting evidence to match. Specifically, StructGEO first constructs a top-down intent tree by recursively decomposing broad search intents into fine-grained ones. It then performs structure-aware rewriting by consolidating evidence relevant to similar intents into coherent sections. These steps broaden the coverage of information needs and keep related evidence coherent rather than scattered throughout the document. Extensive experiments show that StructGEO consistently outperforms existing GEO methods, achieving average gains of 21.2% and 17.5% over AutoGEO and IF-GEO, respectively, in Position-Adjusted Word Count (Word & Pos) visibility.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
generative engine optimization, hierarchical intent modeling, retrieval-augmented generation

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_FCYrOpenReview ↗
Rating
2
Confidence
1
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

This paper addresses the topic of Generative Engine Optimization (GEO), the task of rewriting documents to improve their citation within LLM generative‑RAG responses. Existing GEO approaches suffer from two drawbacks: 1) insufficient coverage of less prominent information needs. 2) overlook the structure of evidence in the docment. To mitigate both issues, the authors propose StructGEO. It first builds a top‑down intent tree: coarse‑grained parent intents are recursively decomposed into fine‑grained leaf‑level intents, to avoid intent concentration. Then it performs structure‑aware document rewriting: leaf rewriting, section merging, and doucment merging, to produce the complete rewritten article. Experiments on GEO-Bench and E-Commerce show that StructGEO outperforms existing methods, like AutoGEO, and IF-GEO.

Critical strengths
  1. This paper has two motivations. Both of them are not purely intuitive, and they provide quantitative evidence for query concentration and evidence dispersion in Fig. 2.
  2. In comparison with other baselines, the method achieves substantial improvements across multiple LLM backbones and two datasets. Therefore, the effectiveness of this method is well‑supported with statistical significance.
  3. The experiments exhibit sound fairness control, with dedicated analyses covering query budget, LLM‑call budget, and cross‑backbone bias.
Critical weaknesses
  1. The quality, fidelity and coverage of the intent trees are not directly evaluated. StructGEO fully relies on LLMs to automatically generate parent and child intents, yet the paper lacks either human annotations or automated metrics to measure critical properties: whether the intents are genuinely askable, whether they are grounded in the source document, whether sibling nodes contain redundant intents, and whether key aspects are omitted.
  2. The systematic evaluation of rewrited documents just relies on a LLM‑judge. Human fact verification are absent. This evaluation can measure superficial quality, yet it cannot reliably prove that rewriting does not introduce hallucinations, distort qualifying conditions, or alter numerical values.
  3. Hyperparameter sensitivity analysis is inadequate. The intent tree serves as the starting point of the proposed method and determines downstream evidence allocation and rewriting merging. The paper does not provide sufficient ablation studies on intent-tree hyperparameters such as tree depth, number of child nodes, and similarity threshold between evidence and nodes.
Questions for authors
  1. Can the authors provide a matched flat-intent baseline with the same number of leaf queries, retrieval calls, adaptive-retrieval calls, rewrite calls, and final merge procedure as StructGEO? This is needed to isolate the contribution of the parent-child hierarchy itself.
  2. Can the authors provide some hyperparameter sensitivity experiments over the number of parent intents, tree depth, number of leaves per parent node, retrieval top-k, and threshold τ?
5 comments

We sincerely thank the reviewer for the thoughtful and constructive feedback. We hope our response alleviates your concerns and further reinforces the key contribution of our approach.

 

W1. Intent Tree Quality, Fidelity, and Coverage

We appreciate the reviewer's valuable feedback. To address these concerns, we conducted additional evaluations and experiments. Our detailed responses are provided below.

1. Quality (Real User Query Coverage)

To evaluate whether hierarchical intent modeling better captures actual user information needs, we conducted an additional experiment comparing the coverage of real user queries by flat and hierarchical query generation.

We first extract information requirements from each real user query, then evaluate whether each generated query satisfies them. A real user query is fully covered if at least one generated query satisfies all its requirements.

DatasetFlat QueriesStructGEORelative Gain
GEO-Bench27.32%30.28%+10.83%
E-Commerce15.87%20.19%+27.22%

StructGEO achieves relative coverage improvements of 10.83% on GEO-Bench and 27.22% on E-Commerce over Flat Queries.

2. Fidelity (Source Grounding)

To evaluate fidelity (source grounding), we assess whether the generated search intents are grounded in their source documents through semantic similarity. Specifically, we use each intent as a retrieval query to rank five candidate documents, including the original target document, using Qwen3-Embedding-8B. The target document achieves average retrieval ranks of 1.444 on GEO-Bench and 1.538 on E-Commerce, indicating strong semantic relevance between the generated intents and their source documents.

3. Document Coverage

We additionally use an LLM judge using GPT-4o mini to evaluate how well the generated intents cover the original document's content. Specifically, the judge assesses whether information in the original document can meaningfully contribute to answering the generated intents. The resulting coverage rates are 90.05% and 92.07% on GEO-Bench, and 87.01% and 89.09% on E-Commerce, using GPT-5 mini and Gemini-3.1 Flash-Lite, respectively. These results suggest that the generated intents meaningfully cover a substantial proportion of the source content.

4. Sibling Redundancy

We evaluated sibling intent redundancy in intent trees generated by GPT-5 mini, using GPT-4o mini as the LLM judge. Two sibling intents are considered redundant only if they request the same information with identical scope, regardless of differences in wording. The redundancy rates are 0.16% on GEO-Bench and 0.29% on E-Commerce. These results suggest that most sibling intents represent distinct information needs.

5. Coverage of Key Aspects

To evaluate whether key information is preserved in the final rewritten documents, we conducted a claim-level evaluation following FActScore. StructGEO achieves 88.4% preservation. These results suggest that StructGEO largely preserves key source information despite prioritizing user-relevant search intents.

 

W2. Factual Faithfulness and Human Verification

We appreciate the reviewer's important concern. To complement our LLM-based quality evaluation, we further assessed potential hallucinations. We (1) conducted an LLM-based, claim-level faithfulness evaluation of rewritten documents against their sources and (2) validated the LLM judge's assessments through human annotation.

Faithfulness evaluation. We conducted an atomic claim-level evaluation, as in FActScore [1], on the E-Commerce dataset, using GPT-5 mini as the optimizer and GPT-4o mini as the judge. An LLM extracts atomic claims from each rewrite, and the judge checks whether each claim is supported by the original.

MethodFaithfulness (%)
AutoGEO85.3
IF-GEO81.0
StructGEO88.6

StructGEO achieves the highest faithfulness (88.6%), as each stage is grounded in the original source evidence.

Human verification. To assess the reliability of our automated factual consistency evaluation, two human annotators independently classified 200 claims as either "supported" or "unsupported." The LLM judge achieved agreement rates of 77.5% and 74.0% with the two annotators, respectively, compared to an inter-annotator agreement of 86.5%. These results indicate a degree of consistency between the LLM judge's assessments and human judgments.

Q1. Contribution of Hierarchical Intent Modeling

To better isolate the contribution of hierarchical intent modeling, we introduced two controlled variants using the StructGEO rewriting framework on E-Commerce (Section 6.3, Fig. 4).

MethodWord & PosQuery Coverage (%)Word & Pos Gain (%)Coverage Gain (%)
Topic Tree34.4810.82+19.8+86.6
Flat Queries37.9915.87+8.7+27.2
StructGEO41.3020.19--
  • Flat Queries: We applied the StructGEO rewriting framework to 15 independently generated IF-GEO queries, treating them as direct children of the root node. StructGEO achieves 8.7% higher Word & Pos and 27.2% higher real user query coverage than this variant.
  • Topic Tree: We constructed a three-level hierarchy from document topics rather than user search intents, matching StructGEO's tree configuration. StructGEO achieves 19.8% higher Word & Pos and 86.6% higher query coverage than this variant.

These results suggest that StructGEO's improvements stem not merely from additional rewriting or document formatting, but also from hierarchically modeling user information needs.

 

W3, Q2. Hyperparameter Sensitivity

We conducted additional sensitivity experiments on five hyperparameters: the number of parent intents, tree depth, number of leaf nodes per parent, retrieval top-k, and evidence similarity threshold τ (Appendix E, Fig. 9). Results are Word & Pos scores on E-Commerce using Gemini-3.1 Flash-Lite.

Leaves per parent1357911
Word & Pos37.7641.2241.5142.8042.2043.15
Tree structure1×153×55×33×3×2
Word & Pos40.2541.5141.3640.81
Retrieval top-k12345
Word & Pos44.3142.4841.5140.3939.18
Threshold τ0.20.40.6
Word & Pos41.5142.6140.86

The results show that StructGEO consistently outperforms the competing GEO methods across all tested configurations, demonstrating its robustness to hyperparameter variations. While performance varies across settings, the default configuration provides substantial visibility gains with a practical optimization budget.

Thank you again for your thoughtful and constructive review. Your valuable feedback has helped us strengthen our work, and we have conducted additional experiments to address your concerns.

As the author-reviewer discussion period approaches its end, we would like to briefly summarize our responses for your consideration.

  • W1. Intent Tree Quality, Fidelity, and Coverage: StructGEO achieves the highest real user query coverage while demonstrating strong source grounding, negligible sibling redundancy, and substantial information preservation.

  • W2. Factual Faithfulness and Human Verification: StructGEO achieves 88.6% faithfulness, and human verification of 200 claims shows that the LLM judge agrees with two annotators at rates of 77.5% and 74.0%, respectively.

  • W3, Q2. Hyperparameter Sensitivity: Our additional sensitivity experiments show that StructGEO consistently outperforms competing GEO methods across all tested hyperparameter settings.

  • Q1. Contribution of Hierarchical Intent Modeling: Using the same rewriting pipeline, StructGEO outperforms Flat Queries and Topic Tree by 8.7% and 19.8%, respectively.

If our responses and additional experiments have sufficiently addressed your concerns, we would sincerely appreciate it if you would consider reassessing your rating. If you have any remaining concerns or questions, we would be happy to address them before the discussion period ends.

Thank you again for your time and thoughtful feedback.

Reviewer_JNJuOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper investigates document rewriting for Generative Engine Optimization (GEO), aiming to improve the visibility of target documents in generative search responses by optimizing their content and organization without access to future real user queries. The authors identify two limitations of existing GEO methods guided by potential queries. First, independently generated queries tend to concentrate on a few search intents, resulting in insufficient coverage of certain information needs. Second, relevant evidence may be scattered throughout a document, making it difficult for generative models to effectively utilize it. To address these issues, the authors propose StructGEO. The method first constructs a search intent tree in a top-down manner, recursively decomposing broad intents into fine-grained ones and grounding each node in relevant evidence from the original document. It then performs structured document rewriting in a bottom-up manner, including intent-guided rewriting at leaf nodes, adaptive evidence retrieval, section merging, and final document integration, thereby organizing related information into a more coherent document structure. The authors evaluate StructGEO on two datasets, GEO-Bench and E-Commerce, using three different LLM backbones. Experimental results show that StructGEO outperforms existing GEO methods across all main evaluation settings.

Critical strengths
  1. The research problem is practically meaningful, and the proposed method is well-designed and comprehensive. This paper addresses the coverage of potential information needs and the organization of supporting evidence in GEO. Beyond rewriting strategies that primarily focus on content expression, the authors further consider how document organization affects the ability of generative search models to utilize relevant evidence. This is a meaningful research perspective. The proposed framework also follows a clear and coherent pipeline, effectively integrating intent modeling, evidence matching, local rewriting, and hierarchical content consolidation.

  2. The experiments are relatively comprehensive, the results are competitive, and the paper is well-developed. The authors conduct systematic experiments on two datasets using three LLM backbones and compare their approach against several existing GEO methods, demonstrating consistent performance improvements. The paper also includes cross-backbone evaluations, ablation studies, query concentration analysis, LLM-call-matched comparisons, and document quality assessments, addressing several potential confounding factors.

Critical weaknesses
  1. The authors observe that potential queries tend to concentrate on a few search intents, but do not sufficiently demonstrate that such concentration leads to inadequate coverage of actual user information needs. Real-world search queries may naturally follow a highly skewed distribution, so more balanced intent coverage does not necessarily lead to better GEO optimization.
  2. StructGEO simultaneously introduces hierarchical intent modeling, local rewriting, evidence supplementation, and multi-stage merging, making it difficult to determine whether the performance improvements stem from better intent modeling or a more sophisticated LLM rewriting pipeline. In particular, constructing a topic tree directly from the document's themes and applying the same hierarchical rewriting pipeline might achieve similar results. The current experiments lack this critical comparison.
  3. The proposed method relies heavily on LLMs to repeatedly rewrite and integrate the original content, which may introduce unsupported claims, factual contradictions, or omissions of important information. This risk is particularly concerning because the rewritten documents will serve as evidence sources for downstream generative search systems.
Questions for authors

Please refer to W1-W3 in the Critical Weaknesses section for my main concerns. I would appreciate further clarification and, where possible, additional experimental evidence addressing these issues, particularly the independent contribution of intent modeling beyond the multi-stage LLM rewriting pipeline (W2).

Overall, I think this is a well-written paper with a reasonable methodology and relatively comprehensive experiments. Although I currently lean toward weak rejection, I am open to raising my rating if the authors can substantively address some of these concerns, especially W2, during the rebuttal.

4 comments

We sincerely thank the reviewer for the thoughtful and constructive feedback. We hope our response alleviates your concerns and further reinforces the key contribution of our approach.

 

W1. Real User Query Coverage

We appreciate the reviewer's valuable feedback. We agree that a more balanced distribution of inferred intents does not necessarily imply better coverage of actual user information needs.

To directly assess this, we conducted an additional experiment comparing the coverage of real user queries by flat and hierarchical query generation, using GPT-4o mini as the judge. We first extract information requirements from each real user query, then evaluate whether each generated query satisfies them. A real user query is fully covered if at least one generated query satisfies all its requirements.

DatasetIF-GEOStructGEORelative Gain
GEO-Bench27.32%30.28%+10.83%
E-Commerce15.87%20.19%+27.22%

StructGEO achieves relative coverage improvements of 10.83% on GEO-Bench and 27.22% on E-Commerce over IF-GEO. These results provide direct evidence that hierarchical intent modeling better captures actual user information needs.

 

W2. Contribution of Hierarchical Intent Modeling

We thank the reviewer for this insightful suggestion. We agree that the contribution of hierarchical intent modeling should be distinguished from that of the rewriting pipeline.

To address this concern, we introduced two controlled variants using the StructGEO rewriting framework on E-Commerce (Section 6.3, Fig. 4).

  • Topic Tree: We constructed a hierarchy from document topics rather than user search intents, matching StructGEO's tree configuration and rewriting procedure.
  • Flat Queries: We applied the StructGEO rewriting framework to 15 independently generated IF-GEO queries, treating them as direct children of the root node.
MethodWord & PosQuery Coverage (%)Word & Pos Gain (%)Coverage Gain (%)
Topic Tree34.4810.82+19.8+86.6
Flat Queries37.9915.87+8.7+27.2
StructGEO41.3020.19--

StructGEO outperforms Topic Tree and Flat Queries by 19.8% and 8.7% in Word & Pos, respectively, with corresponding query coverage gains of 86.6% and 27.2%.

These results highlight the benefits of hierarchically modeling user information needs.

 

W3. Factual Faithfulness and Information Preservation

We appreciate the reviewer's important concern. To assess these risks, we conducted a claim-level evaluation following FActScore [1] on E-Commerce, using GPT-5 mini as the optimizer and GPT-4o mini as the judge. We evaluated two aspects:

  • Faithfulness (unsupported claims, contradictions): an LLM extracts atomic claims from each rewrite, and the judge determines whether each is supported by the original.
  • Preservation (omissions of important information): an LLM extracts key atomic claims from the original, and the judge determines whether each is preserved in the rewritten document.
MethodFaithfulness (%)Preservation (%)F1
AutoGEO85.389.187.2
IF-GEO81.092.986.5
StructGEO88.688.488.5

StructGEO achieves the highest faithfulness consistent with its use of source evidence throughout rewriting. Its preservation is comparable to AutoGEO, while IF-GEO preserves the most but exhibits the lowest faithfulness. Overall, StructGEO achieves the best balance between faithfulness and preservation (highest F1).

[1] Sewon Min, Kalpesh Krishna, Xinxi Lyu, Mike Lewis, Wen-tau Yih, Pang Wei Koh, Mohit Iyyer, Luke Zettlemoyer, and Hannaneh Hajishirzi. FActScore: Fine-grained Atomic Evaluation of Factual Precision in Long Form Text Generation. EMNLP, pp. 12076-12100, 2023.

Thank you again for your thoughtful and constructive review. Your valuable feedback has helped us strengthen our work, and we have conducted additional experiments to address your concerns.

As the author-reviewer discussion period approaches its end, we would like to briefly summarize our responses for your consideration.

  • W1. Real User Query Coverage: StructGEO improves real user query coverage over IF-GEO by 10.83% on GEO-Bench and 27.22% on E-Commerce.

  • W2. Contribution of Hierarchical Intent Modeling: Using the same rewriting pipeline, StructGEO outperforms Topic Tree and Flat Queries by 19.8% and 8.7%, respectively, while achieving the highest real user query coverage.

  • W3. Factual Faithfulness and Information Preservation: Our FActScore-based evaluation shows that StructGEO achieves the highest faithfulness (88.6%) and F1 score (88.5%), demonstrating the best balance between factual faithfulness and information preservation.

If our responses and additional experiments have sufficiently addressed your concerns, we would sincerely appreciate it if you would consider reassessing your rating. If you have any remaining concerns or questions, we would be happy to address them before the discussion period ends.

Thank you again for your time and thoughtful feedback.

Dear Reviewer JNJu,

As the author-reviewer discussion period has less than 12 hours remaining, we would like to briefly follow up on our previous responses.

We have provided detailed clarifications and additional experimental results to address your concerns, and we would greatly appreciate any feedback you may have before the discussion period ends. If you have any remaining concerns or questions, we would be happy to provide further clarification.

If you feel that our responses and additional experiments have sufficiently addressed your concerns, we would sincerely appreciate it if you could consider updating your assessment and rating.

Thank you again for your time and thoughtful consideration.

Best regards,

Authors

Dear Reviewer JNJu,

Thank you again for your thoughtful review.

We have provided detailed responses and additional experiments to address your concerns. We would be sincerely grateful if you could kindly take a moment to review our responses and consider the additional evidence in your assessment.

We truly appreciate your time and consideration.

Best regards,

Authors

Reviewer_g5v8OpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The authors proposes StructGEO, which builds a tree of search intents for a document, rewrites the evidence for each leaf intent, and merges the rewrites bottom-up into one section per broad intent. The design is motivated by two analysis: queries inferred by IF-GEO concentrate on one intent, and interleaving related sentences lowers visibility. On GEO-Bench and E-Commerce with three backbones, StructGEO reports average Word & Pos gains of 21.2% over AutoGEO and 17.5% over IF-GEO.

Critical strengths
  1. The paper tackles an important problem of how a document should be structured for generative engines.
  2. The idea is simple and intuitive. The motivating experiment in Figure 2b, where only the order of sentences is changed, is interesting.
  3. The evaluation is thorough, with three backbones, three runs with significance tests, a call-matched comparison with IF-GEO and a cross-backbone study.
  4. The paper is well-written and easy to follow.
Critical weaknesses
  1. Effect of document length: It is unclear whether the improvements come from the structure or simply from longer documents. For instance, the leaf prompt asks the model to "develop each point requested by the intent in depth". Further, with GPT-5 mini, StructGEO's rewrites are longer than IF-GEO's in every bin of Figure 5b, up to about 10x the original document. Appendix C only compares lengths within each method, and not across methods. Reporting the average length for each method, and comparing the methods at similar lengths, would be helpful.
  2. Role of the hierarchy: The ablation in Table 3 does not show whether the intent hierarchy itself is needed. When a merge stage is disabled, its inputs are simply concatenated, so the gain from each merge stage also includes one more LLM rewriting pass. Similarly, the Reorganization-only variant converts the document to Markdown with headings, so part of its gain may simply come from the formatting. A baseline that assigns specific intents to random broad intents, or one that runs the same pipeline on IF-GEO's flat list of queries, would show whether the hierarchy matters.

Other Weaknesses:

  1. Faithfulness: The leaf prompt asks the model to explain "why it matters, and under what conditions it applies". The Reorganization-only prompt forbids adding external facts, but the leaf prompt does not. So the rewrites may contain claims that are not in the source document. The judge in Appendix B does not check this, and the quality of the final answer (GEU) is also not reported. A faithfulness check on a small sample of rewrites would strengthen the paper.
  2. Self-preference bias: The main results use the same model as both the optimizer and the generative engine. Table 2 shows that StructGEO benefits the most from this setting (14.1% higher than cross-backbone, compared to 7.4% for IF-GEO), so part of the reported improvement may come from self-preference. In the cross-backbone setting, StructGEO is still 13.3% better than IF-GEO. I think this is the more reliable number and could be reported as the main result.
  3. Motivation M1 only shows a correlation. Documents whose queries fall under one intent may simply be narrower documents, which could also explain their lower visibility. The controlled experiment in Appendix D is stronger evidence and could be moved to the main paper.
  4. The cost is compared in number of LLM calls (Figure 4). Since StructGEO's rewrites are much longer, a comparison in tokens would be more informative.
Questions for authors

Please see the weaknesses above.

5 comments

We sincerely thank the reviewer for the thoughtful and constructive feedback. We hope our response alleviates your concerns and further reinforces the key contribution of our approach.

 

W1. Effect of Document Length

We appreciate the reviewer's valuable suggestion. We agree that it is important to distinguish the effects of document length from those of structural optimization.

We have added the average rewritten document lengths for all three methods (Appendix D, Table 6) and conducted additional comparisons under comparable document-length conditions.

We select queries where the IF-GEO rewrite length is within 75-150% of the corresponding StructGEO rewrite length. Among these samples, IF-GEO produces shorter and longer rewrites in 51.04% and 48.73% of cases, respectively.

DatasetBackbone# SamplesIF-GEOStructGEO
E-CommerceGPT-5 mini123.735.5340.22
E-CommerceGemini-3.1236.335.9143.99
E-CommerceQwen3.857.332.9337.54
GEO-BenchGPT-5 mini218.735.1239.63
GEO-BenchGemini-3.1488.034.1634.91
GEO-BenchQwen3.874.733.4936.28

StructGEO outperforms IF-GEO across all settings, achieving average Word & Pos gains of 16.57% and 7.79% on E-Commerce and GEO-Bench, respectively, with an overall gain of 12.18%.

These results suggest that StructGEO's visibility improvements cannot be explained solely by document length.

 

W2. Role of the Hierarchy

We thank the reviewer for this insightful suggestion. We agree that the contribution of hierarchical intent modeling should be distinguished from that of the rewriting pipeline.

To address this concern, we introduced two controlled variants using the StructGEO rewriting framework on E-Commerce using GPT-5 mini (Section 6.3, Fig. 4).

  • Topic Tree: We constructed a hierarchy from document topics rather than user search intents, matching StructGEO's tree configuration and rewriting procedure.
  • Flat Queries: We applied the StructGEO rewriting framework to 15 independently generated IF-GEO queries, treating them as direct children of the root node.
MethodWord & PosGain (%)
Topic Tree34.48+19.8
Flat Queries37.99+8.7
StructGEO41.30-

StructGEO outperforms Topic Tree and Flat Queries by 19.8% and 8.7%, respectively.

To further investigate this, we conducted an additional experiment comparing real user query coverage between flat and hierarchical query generation, using GPT-4o mini as the judge. We first extract information requirements from each real user query and then evaluate whether each generated query satisfies them. A real user query is considered fully covered if at least one generated query satisfies all its requirements.

StructGEO achieves 20.19% real user query coverage, outperforming Flat Queries (15.87%, +27.2% gain) and Topic Tree (10.82%, +86.6% gain). These results suggest that hierarchical intent modeling better captures real user information needs than flat query generation or document-centric topic hierarchies.

 

W3. Faithfulness

We thank the reviewer for this valuable suggestion. To assess whether rewriting introduces unsupported claims, we conducted a claim-level evaluation following FActScore [1] on E-Commerce, using GPT-5 mini as the optimizer and GPT-4o mini as the judge.

Specifically, an LLM extracts atomic claims from each rewrite, and the judge determines whether each claim is supported by the original. We report the percentage of supported claims.

MethodFaithfulness (%)
AutoGEO85.3
IF-GEO81.0
StructGEO88.6

This result demonstrates that StructGEO achieves higher factual faithfulness and factual consistency than competing GEO methods.

[1] Sewon Min, Kalpesh Krishna, Xinxi Lyu, Mike Lewis, Wen-tau Yih, Pang Wei Koh, Mohit Iyyer, Luke Zettlemoyer, and Hannaneh Hajishirzi. FActScore: Fine-grained Atomic Evaluation of Factual Precision in Long Form Text Generation. EMNLP, pp. 12076-12100, 2023.

 

W4. Self-preference Bias

We thank the reviewer for this valuable suggestion. We will consider placing greater emphasis on the cross-backbone results in the final manuscript.

 

W5. Motivation M1

We appreciate the reviewer's helpful observation. To clarify the connection between the two analyses, we moved the explanation of the controlled experiment from the final paragraph of Section 6.3 to the Motivation section.

W6 Computational Cost in Tokens

We thank the reviewer for this constructive suggestion. We agree that token consumption provides a more informative measure of computational cost than LLM-call counts alone.

We have added a token-based cost comparison (Appendix B, Fig. 7), measuring the total input, output, and reasoning tokens consumed per document across the entire optimization pipeline.

StructGEO achieves a Word & Pos score of 41.3 using 98.9K tokens, outperforming IF-GEO at both a comparable token budget (34.1 at 94.5K tokens) and a larger budget (34.7 at 137.4K tokens) on E-Commerce using GPT-5 mini. These results suggest that StructGEO's performance gains cannot be explained solely by token consumption.

Thank you again for your thoughtful and constructive review. Your valuable feedback has helped us strengthen our work, and we have conducted additional experiments to address your concerns.

As the author-reviewer discussion period approaches its end, we would like to provide this brief summary for your consideration.

  • W1. Effect of Document Length: We conducted length-matched comparisons, showing that StructGEO outperforms IF-GEO by 12.18% on average even at comparable document lengths.

  • W2. Role of the Hierarchy: Under the same rewriting pipeline, StructGEO outperforms Flat Queries and Topic Tree by 8.7% and 19.8%, respectively, while achieving the highest real user query coverage.

  • W3. Faithfulness: Our FActScore-based evaluation shows that StructGEO achieves higher faithfulness (88.6%) than AutoGEO (85.3%) and IF-GEO (81.0%).

  • W4-W5. Presentation: We appreciate your helpful suggestions regarding the presentation of our results and have clarified the connection between Motivation M1 and the controlled experiment in the Appendix.

  • W6. Computational Cost: Our token-based comparison shows that StructGEO achieves higher visibility than IF-GEO, even when IF-GEO uses a comparable or larger token budget.

If our responses and additional experiments have sufficiently addressed your concerns, we would sincerely appreciate it if you would consider reassessing your rating. If you have any remaining concerns or questions, we would be happy to address them during the remaining discussion period.

Thank you again for your time and thoughtful feedback.

Comments on the paper

1 comment
AuthorsOpenReview ↗

Summary of Revisions

We have revised our manuscript and conducted additional experiments to address the reviewers' concerns regarding the contribution of hierarchical intent modeling, the effects of document length and token budget, and experimental robustness.

1. Intent Generation Prompt Refinement

We refined the intent generation prompts in Appendix I to prevent overly broad parent intents and excessively narrow child intents. We also allowed one child intent to represent the general search intent of its parent. We reran all experiments of StructGEO with the revised prompts. StructGEO achieves average Word & Pos gains of 31.9% and 27.8% over AutoGEO and IF-GEO, respectively.

2. Contribution of Hierarchical Intent Modeling

  • Intent Modeling Ablation (Fig. 4a): Under the same rewriting procedure, StructGEO outperforms the Topic Tree and Flat Queries variants by 19.8% and 8.7%, respectively.
  • Real User Query Coverage (Fig. 4b): StructGEO achieves relative coverage gains of 86.6% and 27.2% over Topic Tree and Flat Queries, respectively.

3. Document Length and Computational Budget Analysis

  • Document Length Analysis (Appendix D, Tables 6-8): We compared methods under comparable document-length conditions. StructGEO achieves an average gain of 33.91% and 10.55% over AutoGEO and IF-GEO, respectively, using Gemini-3.1 Flash-Lite.
  • Token Consumption Analysis (Appendix B, Fig. 7): We extended the cost analysis to include input, output, and reasoning tokens. StructGEO maintains higher visibility even under larger IF-GEO token budgets.

4. Additional Analyses and Reproducibility

  • Hyperparameter Sensitivity (Appendix E): We extended the analysis to tree depth, parent nodes, retrieval top-k, and similarity threshold.
  • Reproducibility (Appendix J): We included the prompts for Topic Tree construction and user query coverage evaluation.

We believe these revisions comprehensively address the reviewers' feedback. We will also incorporate further revisions based on the ongoing discussion and additional feedback.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.