Source-linked AI summary
Query Implied Generative Engine Optimization
Shilpa Ramakrishna, William B. Andreopoulos
TL;DR
Generative search makes citation in generated answers important for visibility, while existing GEO methods often depend on explicit or query-derived signals. QI-GEO infers intent from document semantics and identifies relevant missing concepts; evaluations report improvements in visibility and subjective scores over original documents, including across related query variations.
Problem
Most GEO approaches rely on explicit queries or query-derived signals, although content creators rarely know the full range of questions users may ask.
Source: §I · p. 1, C. Evaluation Frameworks and Metrics · p. 2Method
QI-GEO extracts document entities and relationships, expands them through a knowledge graph, and identifies informational gaps without requiring future queries.
Source: A. Overview of QI-GEO · p. 3, §VII · p. 8Results
Experiments on GEO-Bench and ExtendedGEOBench found improved citation coverage, source visibility, and subjective impression scores over original documents, with improvements effective across semantically related query variations.
Source: §VII · p. 8Takeaways & Limitations
Document-centric intent modeling is presented as a viable alternative to query-based optimization strategies.
Source: §VII · p. 8
Abstract
from arXiv · showhide
The landscape of search has changed drastically with how people look for information online. Traditional search engines are being replaced by Generative Search Engines (GSEs), which use Large Language Models (LLMs) to generate natural language responses to user queries. For content creators, visibility is no longer solely determined by ranking in search results but by being cited within generated responses. But Generative Search Engines are black-boxes, leading to the emergence of Generative Engine Optimization (GEO), a set of techniques aimed at improving content visibility in generative search settings. Most existing approaches rely on the explicit queries or query derived signals to align content to better suit user needs. We propose Query Implied Generative Engine Optimization (QI-GEO) to infers user intent directly from the document. Our approach approximates document's intent space and identifies content that may be missing yet relevant to answer potential user queries. Evaluation on GEO-Bench and Extended GEO-Bench demonstrated improvements across objective and subjective metrics. QI-GEO improved objective scores by up to 15.9% and subjective scores by up to 17.6%, while yielding nearly twice as many citation gains as citation losses. These results suggest that document-derived approximations of user intents can improve visibility without relying on explicit query inputs.
I. INTRODUCTION
Generative Search Engines make citation in generated answers a visibility goal, while query-dependent GEO methods are difficult when creators cannot know the questions users may ask. QI-GEO instead infers intent from document semantics and targets related information gaps.
I. INTRODUCTION
- GSEs generate natural-language answers with source citations, making citation in generated responses a visibility concern beyond search ranking.Traditional SEO practices alone may be insufficient to maintain visibility in these systems.Source: §I · p. 1
A. Generative Engine Optimization
- Most GEO approaches use explicit queries or query-derived signals, although creators rarely know the full range of questions users may ask.This makes query-dependent optimization difficult in black-box generative search systems.Source: §I · p. 1, A. Generative Engine Optimization · p. 2
I. INTRODUCTION
- QI-GEO structures document entities and relationships, expands them with external knowledge, and identifies related concepts missing from the content as optimization opportunities.The approach aims to improve coverage while grounding edits in document semantics rather than generated queries.Source: §I · p. 1
- QI-GEO showed consistent improvements across single-query and multi-query evaluation settings.The evaluation used GEO-Bench and ExtendedGEOBench with objective, subjective, citation-based, and stability metrics.Source: §I · pp. 1–2
B. Query and Intent Modeling
Prior GEO methods approximate user intent through known or generated queries, but the full query distribution for a document is unavailable at optimization time. The problem formulation therefore calls for an intent representation derived directly from the document, evaluated across relevant queries.
C. Evaluation Frameworks and Metrics
- GEO-Bench contains 10,000 queries, and ExtendedGEOBench groups semantically related queries for robustness and multi-query evaluation, [6].Existing visibility evaluation also uses objective metrics such as Word Count and PAWC, subjective LLM-as-a-judge measures, and stability metrics including WTR, WCP, and DR,,.Source: C. Evaluation Frameworks and Metrics · p. 2
B. Query and Intent Modeling
- Known queries, generated queries, and query-derived signals are common ways to approximate user intent, but content creators may not know the complete query space.RAID uses user-role simulation and reflection, while IF-GEO uses reverse retrieval to generate potential queries [6],.Source: C. Evaluation Frameworks and Metrics · p. 2, B. Query and Intent Modeling · p. 2
III. PROBLEM FORMULATION
- A GSE retrieves candidate documents and generates a natural-language response; document visibility concerns its incorporation and citation in that response,.Visibility may be measured with objective citation-contribution and placement measures or subjective perceived-influence measures,.Source: §III · p. 2
- The formulation aggregates visibility gains over a query set, using an aggregation function such as the mean or minimum.The stated challenge is to derive an alternative intent representation directly from the document.Source: §III · pp. 2–3
- Because the query distribution P(D) is unavailable during optimization, the objective is to maximize expected visibility gain over queries associated with document D.Prior methods assume relevant queries are known or can be approximated through query generation, [6],.Source: §III · p. 2
IV. METHODOLOGY
QI-GEO infers user intent from document semantics rather than explicit queries, then expands the document’s semantic space to identify relevant content it does not cover. It ranks those gaps and turns promising ones into targeted edits.
A. Overview of QI-GEO
- QI-GEO infers user intent from document semantics without relying on explicit query inputs.It treats entities and relationships in a document as signals of information needs associated with its topic.Source: A. Overview of QI-GEO · p. 3
- The framework expands document relationships using external knowledge to find relevant concepts that are not explicitly covered.The expansion approximates the broader semantic space surrounding the document.Source: A. Overview of QI-GEO · p. 3
- QI-GEO ranks candidate informational gaps and converts promising ones into targeted edits intended to improve coverage while preserving the document’s original content and structure.Source: A. Overview of QI-GEO · p. 3
B. Relation Mining and Knowledge Graph Construction
QI-GEO represents document semantics as grounded entity–relation triplets, expands them through a knowledge graph, and ranks uncovered concepts as candidate informational gaps. Ranking favors concepts supported by multiple document entities, while coverage checks filter out information already expressed.
B. Relation Mining and Knowledge Graph Construction
- Coreference resolution and relation extraction convert the document into subject–relation–object triplets, which are linked to knowledge-graph identifiers to form its semantic foundation.Coreference resolution preserves relationships spanning sentences; triplets provide a compact representation compatible with graph entities and relations.Source: B. Relation Mining and Knowledge Graph… · p. 3
C. Knowledge Expansion and Coverage Mapping
- One-hop knowledge-graph expansion retrieves directly related concepts, while a coverage map identifies entity–relation pairs already expressed in the document.Expansion prioritizes definitional, causal, or explanatory relations over metadata or navigational relations; concepts absent from the map become candidate gaps.Source: C. Knowledge Expansion and Coverage… · p. 3
D. Gap Discovery and Ranking
- Convergence-based ranking gives greater importance to candidate concepts independently supported by multiple document entities.Relation groups receive facet weights according to reader-facing usefulness, and source-entity importance reflects normalized mention frequency, favoring entities central to the document.Source: D. Gap Discovery and Ranking · pp. 3–4
- Candidates must pass exact and semantic coverage checks to qualify as true informational gaps; the highest-ranked remaining gaps are selected for rewriting.Source: D. Gap Discovery and Ranking · p. 4
E. Document Rewriting
QI-GEO turns selected informational gaps into localized, anchored edits rather than unrestricted rewrites. This links each revision to a specific gap while aiming to preserve the original document’s intent, structure, and content.
E. Document Rewriting
- Each gap becomes an edit specification with an anchor, a suggestion, and a necessity score.The anchor identifies related existing text, and the suggestion describes the missing concept to incorporate.Source: E. Document Rewriting · p. 4
- Edits are applied sequentially as localized additions, usually a short clause or sentence appended to the selected anchor.The rewriter is instructed to avoid large-scale rewrites to preserve document structure and minimize unintended changes.Source: E. Document Rewriting · p. 4
- Anchored revisions can be traced to their informational gaps and original locations, while the final document incorporates high-value missing information and preserves its original intent, structure, and content.Source: E. Document Rewriting · p. 4
V. EXPERIMENTAL RESULTS
The experiments are framed around whether document semantics can support query-independent optimization in generative search. They ask whether intent can be inferred without explicit queries, whether expansion reveals useful missing information, and whether gains hold across query formulations.
A. Research Questions
- The experiments ask whether user intent can be inferred directly from a document without access to explicit queries.Source: A. Research Questions · p. 5
- They also ask whether semantic expansion can reveal missing but relevant information that improves document visibility in generative search engines.Source: A. Research Questions · p. 5
- A further question is whether visibility gains from query-independent optimization remain consistent across formulations of the same information need.Source: A. Research Questions · p. 5
B. Datasets and Benchmarks
QI-GEO is evaluated on GEO-Bench and ExtendedGEOBench under fixed retrieval, then applied through a pipeline that extracts document relations, expands them in a knowledge graph, identifies gaps, and makes localized edits.
B. Datasets and Benchmarks
- GEO-Bench contains 10,000 query-document instances with five candidate documents and one target per query; ExtendedGEOBench groups related queries with different phrasings.The latter supports evaluation across multiple formulations of the same information need.Source: B. Datasets and Benchmarks · p. 5
- The fixed-retrieval setup keeps the candidate pool unchanged, comparing each optimized target document with its original while all other documents remain unchanged.This isolates document modifications from retrieval variability.Source: B. Datasets and Benchmarks · p. 5
C. Experimental Configuration
- The pipeline extracts document relations, links triplets to a knowledge graph, and expands them through one-hop neighborhood retrieval.REBEL was selected from evaluated relation extractors because it consistently produced useful triplets for expansion.Source: C. Experimental Configuration · p. 5
- Generic and metadata-oriented relations are filtered, while definitional, explanatory, and causal relations are retained before candidate concepts are ranked and filtered.Ranking uses the convergence-based strategy, semantic similarity, and coverage checks.Source: C. Experimental Configuration · p. 5
- Selected informational gaps become edit requests implemented as localized Qwen2.5 7B edits anchored to existing content.The edits introduce missing information without substantially altering the document’s original structure or meaning.Source: C. Experimental Configuration · p. 5
D. Evaluation Metrics
The evaluation combines objective, subjective, and stability measures of generative-search visibility. QI-GEO improves citation coverage, with citation gains exceeding losses across both single-query and multi-query settings.
D. Evaluation Metrics
- Objective measures include Word Count Impression, Position Impression, and PAWC, which weights cited-word contributions by citation position.PAWC combines citation volume and prominence and is the primary objective metric.Source: D. Evaluation Metrics · p. 5
- Subjective impression scores average seven five-point dimensions, including relevance, influence, uniqueness, diversity, and perceived position and volume.The other dimensions are click follow-up likelihood and diversity.Source: D. Evaluation Metrics · p. 5
- Stability is assessed with Mean Gain, Win-Tie Rate, Worst-Case Performance, and Downside Risk to describe average improvement and adverse outcomes.The measures capture performance across query variations, including the largest observed drop and the magnitude of negative outcomes.Source: D. Evaluation Metrics · p. 5
A. Citation Coverage
- Citation coverage rose from 0.657 to 0.726, a 10.5% relative improvement in single-query evaluation, and increased by 8.3% in multi-query evaluation.Table I reports the fraction of queries in which the target document is cited.Source: Table I · p. 6, A. Citation Coverage · p. 6
- Citation gains outnumber losses by more than 2:1 in single-query evaluation, with a similar pattern in multi-query evaluation.The reported pattern reflects a tendency toward more frequent citations, rather than a small number of large improvements.Source: Figure 2 · p. 6, A. Citation Coverage · p. 6
B. Source Visibility
QI-GEO improves PAWC, which reflects both citation volume and prominence. Improvements are positive overall, while the reported WTR and downside risk indicate gains across many queries with generally small losses when visibility declines.
B. Source Visibility
- Mean PAWC increased from 0.2137 to 0.2407, a 12.6% relative improvement in single-query evaluation, and rose by 15.9% in multi-query evaluation.PAWC captures citation frequency and position, with optimized documents tending to appear earlier and contribute more cited content.Source: Table II · p. 6, B. Source Visibility · p. 6
- The distribution shifts toward positive gains in both settings, and average improvement is not driven by a small number of outliers.QI-GEO does not improve every query.Source: B. Source Visibility · p. 6
- QI-GEO achieves a WTR of 0.70 and downside risk of 0.011, with losses generally small when visibility declines.The WTR means optimized documents perform at least as well as the original for approximately seven out of ten queries.Source: B. Source Visibility · p. 6
C. Subjective Evaluation
QI-GEO raises mean subjective scores in both evaluation settings, and its visibility gains persist across semantically related query variants. The evaluation also reports positive subjective-score stability and gains across differing query phrasings.
C. Subjective Evaluation
- An LLM judge evaluates subjective impressions across relevance, influence, uniqueness, diversity, click follow-up likelihood, subjective position, and subjective volume.Source: C. Subjective Evaluation · p. 6
- 17.6% and 17.1%: QI-GEO’s mean subjective scores rise from 2.332 to 2.743 in single-query evaluation and from 2.137 to 2.503 in multi-query evaluation.Table III reports mean subjective scores for original and QI-GEO-rewritten documents.Source: Table III · p. 6, C. Subjective Evaluation · p. 6
- The lowest subjective-score range shrinks while higher ranges become more populated, shifting initially weak-visibility documents toward moderate and high categories.Source: Figure 3 · p. 7, C. Subjective Evaluation · p. 7
- A subjective-score WTR of 0.81 means optimized documents score at least as well as originals for more than four out of five queries.Some negative outcomes remain, though the overall distribution is strongly skewed toward positive gains.Source: C. Subjective Evaluation · p. 7
D. Multi-Query Robustness
- 66.0% of documents have positive mean gain across five query variants; 73.7% improve or maintain visibility on at least three, and 22.2% improve on all five.Source: D. Multi-Query Robustness · p. 7
- When optimization succeeds, average visibility gain is +0.1634 versus an average loss of -0.0629 when it fails; downside risk is 0.0097.Source: D. Multi-Query Robustness · p. 7
- 87.5% of documents gain visibility on at least one query phrasing, although the improvement varies across formulations.Figure 5 visualizes the range of gains across variants for each document.Source: Figure 5 · p. 7, D. Multi-Query Robustness · p. 7
VII. CONCLUSION AND FUTURE WORK
QI-GEO infers user intent from document semantics, expanding entities and relationships through a knowledge graph to identify potentially relevant informational gaps. Experiments report improved visibility and subjective scores, while future work points to richer and more specialized knowledge graphs.
VII. CONCLUSION AND FUTURE WORK
- QI-GEO models user intent from document semantics, using entities, relationships, knowledge-graph expansion, and informational-gap identification without future queries.Source: §VII · p. 8
- Experiments on GEO-Bench and ExtendedGEOBench show improved citation coverage, source visibility, and subjective impression scores, with gains effective across related query variations.Source: §VII · p. 8
- Future work could improve gap coverage through richer graph expansion, jointly model inferred intents, and use domain-specific graphs for specialized domains.The paper names law, finance, healthcare, and geospatial information as examples of specialized domains.Source: §VII · p. 8