Diagnosing and Repairing Citation Failures in Generative Engine Optimization

Zhihua Tian; Yuhan Chen; Yao Tang; Jian Liu; Ruoxi Jia

cs.IRMar 10, 2026

Diagnosing and Repairing Citation Failures in Generative Engine Optimization

Authors: Zhihua Tian, Yuhan Chen, Yao Tang, Jian Liu, Ruoxi Jia

Organizations: Virginia Tech · Zhejiang University

Abstract

Generative Engine Optimization (GEO) aims to improve content visibility in AI-generated responses. However, existing methods measure contribution-how much a document influences a response-rather than citation, the mechanism that actually drives traffic back to creators. Also, these methods apply generic rewriting rules uniformly, failing to diagnose why individual document are not cited. This paper introduces a diagnostic approach to GEO that asks why a document fails to be cited and intervenes accordingly. We develop a unified framework comprising: (1) the first taxonomy of citation failure modes spanning different stages of a citation pipeline; (2) AgentGEO, an agentic system that diagnoses failures using this taxonomy, selects targeted repairs from a corresponding tool library, and iterates until citation is achieved; and (3) a document-centric benchmark evaluating whether optimizations generalize across held-out queries. AgentGEO achieves over 40% relative improvement in citation rates while modifying only 5% of content, compared to 25% for baselines. Our analysis reveals that generic optimization can harm long-tail content and some documents face challenges that optimization alone cannot fully address-findings with implications for equitable visibility in AI-mediated information access.

Figures & tables

Figure 1: Comparison between traditional search engines and LLM-powered generative engines. Traditional search engines return a ranked list of hyperlinks based on relevance to the user query, while generative engines synthesize answers using retrieved content and provide citations to source webpages.

Figure 2: A taxonomy of citation failure modes in generative engines, spanning html fetching, prasing and answer generation stages.

Figure 3: Optimization process of $\mathtt{AgentGEO}$ . $\mathtt{AgentGEO}$ starts by verifying citation status against competitor documents. For uncited documents, it triggers a ”Diagnose-then-Repair” loop where: (1) an LLM first diagnoses the reason the document was not cited; (2) based on the diagnosis and a query-specific memory of past iterations, a specific tool is selected to modify a surrogate webpage (a copy of the original); and (3) the modified surrogate is retested against the same competitors. The loop repeats until citation is achieved or a limit is reached. All successful modifications are merged via an aggregation strategy into a single suggestion to optimize the original webpage. This step is then iteratively looped over additional training queries to further optimize the page.

Dataset	Doc-centric	HTML	Train/Test Split
GEO-Bench ( Aggarwal et al., 2024 )	$\times$	$\times$	$\checkmark$
E-commerce ( Wu et al., 2025 )	$\times$	$\times$	$\checkmark$
Researchy-GEO ( Wu et al., 2025 )	$\times$	$\times$	$\checkmark$
CC-GSEO-Bench ( Chen et al., 2025b )	$\checkmark$	$\times$	$\times$
$\mathbf{MIMIQ}$ (Ours)	$\checkmark$	$\checkmark$	$\checkmark$

Table 1: Comparison of

\mathbf{MIMIQ}

and existing GEO benchmarks. Our dataset uniquely combines a document-centric paradigm with a holistic evaluation scope and supports explicit train/test query splits.

	In-context							Attribute first then generate
	CR	Contribution			Faithfulness			CR	Contribution			Faithfulness
Method	CR	Word	Pos	Wordpos	TF-IDF	Embed	Jaccard	CR	Word	Pos	Wordpos	TF-IDF	Embed	Jaccard
Vanilla	56.58	16.28	16.28	16.39	-	-	-	60.20	23.69	28.72	23.57	-	-	-
Technical Terms	55.84	14.78	14.73	14.60	73.32	85.64	21.37	57.33	22.25	27.87	22.03	74.39	86.98	21.98
Cite Sources	60.52	17.01	16.96	16.89	92.76	93.62	65.62	59.21	24.04	29.02	23.79	92.33	92.95	64.60
Keyword Stuffing	62.44	17.83	17.74	17.82	82.85	88.95	30.04	65.62	28.09	33.39	27.81	82.16	86.92	28.39

Table 2: Performance comparision of

\mathtt{AgentGEO}

againse baselines across different citation methods. Bold and underline indicate the best and second-best results, respectively. ”-” indicates the metric is not applicable to the method.

Model	Metric	Vanilla	AutoGEO	$\mathtt{AgentGEO}$ (ours)
GPT	CR	56.58	68.80	79.52
	Word	16.28	17.87	20.29
	Pos	16.28	18.53	21.08
	Wordpos	16.39	17.65	20.28
	Precision	48.63	56.79	58.43
	Recall	64.62	68.99	67.10

Table 3: Performance comparison across different LLM-based generative engines. Bold indicates the best performance.

Figure 4: Citation rate improvement over vanilla baseline across different topics. The numbers in left indicate the original citation rate for each topic.

Panel of Figure 8 — Algorithm 1 Workflow of $\mathtt{AgentGEO}$

Appendix figures & tables6 assets

Supplementary material from the paper’s appendix.

Appendix

Panel of Figure 5 — Figure 5: Distribution of topics and content lengths in the MIMIQ benchmark.

		Citation	Contribution			Faithfulness
Setting	Method	CR	Word	Pos	Wordpos	TF-IDF	Embed	Jaccard
MIMIQ-OOD	Vanilla	57.14	25.19	25.01	24.94	-	-	-
	Easy-to-Understand	68.06	24.48	24.79	24.70	81.60	86.42	30.24
	Statistics Addition	58.57	20.72	21.03	20.93	96.07	94.39	70.75
	Quotation Addition	60.71	21.06	21.19	21.24	96.70	94.64	75.14
	Fluency Optimization	59.69	20.80	21.31	21.11	92.40	92.73	49.20

Appendix

Table 4: Performance comparison in OOD and HTML compatibility settings. Bold indicates the best performance and underline indicates the second-best performance. ”-” indicates the metric is not applicable to the method.

	Citation	Contribution			Faithfulness
Method	CR	Word	Pos	Wordpos	TF-IDF	Embed	Jaccard
Vanilla	69.98	26.09	25.98	25.93	–	–	–
$\mathtt{AgentGEO}$	84.29	35.18	34.89	34.54	94.91	90.32	75.36
Improv.	+14.31	+9.09	+8.91	+8.61	–	–	–

Appendix

Table 5: Cross-method robustness evaluation.

Panel of Figure 6 — Figure 6: Performance analysis on webpage constraints. (a) Comparison of citation rates across varying webpage lengths. (b) Evaluation of content preservation using similarity metrics (TF-IDF, Cosine Embedding, and Jaccard).

	Citation	Contribution			Faithfulness
Config	CR	Word	Pos	Wordpos	TF-IDF	Embed	Jaccard
Vanilla	56.58	16.28	16.28	16.39	-	-	-
w/o Memory	75.94	22.21	22.28	22.09	96.89	94.39	84.51
w/ Memory	79.32	22.69	22.66	22.53	96.95	95.45	84.86
Batch 1	74.53	25.56	25.57	25.52	93.41	88.77	77.05
Batch 5	80.86	23.39	23.40	23.23	96.87	94.14	83.35

Appendix

Table 6: Comparison of optimization performance across different memory settings and batch configurations. Bold indicates the best performance.

# Training Queries	Mean	Std. Dev.	Abs. Improv.	Rel. Improv.
0	0.6837	0.2050	–	–
10	0.7200	0.2112	0.0363	5.30%
20	0.7214	0.1994	0.0377	5.51%
40	0.7614	0.1910	0.0777	11.36%
60	0.7729	0.1767	0.0891	13.04%
80	0.8044	0.1625	0.1207	17.65%

Appendix

Table 7: Citation Performance Across Training Set Sizes

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

Think Before Writing: Feature-Level Multi-Objective Optimization for Generative Citation Visibility
Apr 21, 2026Zikang Liu, Peilan XuGenerative Engine Optimization Multi-Objective Optimization
School of Artificial Intelligence, Nanjing University of Information Science and Technology, Nanjing 210044, China.
GEO-Flag: Detecting and Measuring GEO-Optimized Web Content
Aug 17, 2026Junjie Chu, Ye Leng, Mingjie Li +3Generative Engine Optimization IR Evaluation
CISPA Helmholtz Center for Information Security · HPE · University of Waterloo
From Experience to Skill: Multi-Agent Generative Engine Optimization via Reusable Strategy Learning
Apr 21, 2026Beining Wu, Fuyou Mao, Jiong Lin +7Generative Engine Optimization LLM Agent Skill Learning
Hangzhou Dianzi University · Central South University · Shanghai Artificial Intelligence Laboratory +3

Apr 21, 2026J/K move · Enter open · S save

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.

Generative Engine Optimization Multi-Objective Optimization

Open paper AskPDF arXiv

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.

Abstract

Generative Engine Optimization

Explore similar work

Think Before Writing: Feature-Level Multi-Objective Optimization for Generative Citation Visibility

GEO-Flag: Detecting and Measuring GEO-Optimized Web Content

From Experience to Skill: Multi-Agent Generative Engine Optimization via Reusable Strategy Learning