Controlling Output Rankings in LLM-Based Search via Generative Engine Optimization

Haibo Jin; Huimin Zeng; Yifeng Luo; Haohan Wang

cs.CLFeb 3, 2026

Controlling Output Rankings in LLM-Based Search via Generative Engine Optimization

Authors: Haibo Jin, Huimin Zeng, Yifeng Luo, Haohan Wang

Organizations: School of Information Sciences, University of Illinois at Urbana-Champaign, IL, USA

Abstract

The way customers search for and choose products is changing with the rise of large language models (LLMs). LLM-based search, or generative engines, provides direct product recommendations to users, rather than traditional online search results that require users to explore options themselves. However, these recommendations are strongly influenced by the initial retrieval order of LLMs, which disadvantages small businesses and independent creators by limiting their visibility. In this work, we propose CORE, an optimization method that Controls Output Rankings in gEnerative Engines for LLM-based search. Since the LLM's interactions with the search engine are black-box, CORE targets the content returned by search engines as the primary means of influencing output rankings. Specifically, CORE optimizes retrieved content by appending strategically designed optimization content to steer the ranking of outputs. We introduce three types of optimization content: string-based, reasoning-based, and review-based, demonstrating their effectiveness in shaping output rankings. To evaluate CORE in realistic settings, we introduce ProductBench, a large-scale benchmark with 15 product categories and 200 products per category, where each product is associated with its top-10 recommendations collected from Amazon's search interface. Extensive experiments on four LLMs with search capabilities (GPT-4o, Gemini-2.5, Claude-4, and Grok-3) demonstrate that CORE achieves an average Promotion Success Rate of 91.4% @Top-5, 86.6% @Top-3, and 80.3% @Top-1, across 15 product categories, outperforming existing ranking manipulation methods while preserving the fluency of optimized content.

Figures & tables

Figure 1: Overview of how the LLM processes Alice’s query ( blue box ) and how output ranking is controlled by CORE ( orange box ), where the Hello Kitty camera, originally the lowest-ranked result, becomes the top-1 item when applied to CORE-optimized content.

Figure 2: Overview of CORE. (a) Shadow-model optimization uses a shadow model to approximate the synthesizing LLM and directly compute ranking gradients. (b) Query-based optimization interacts with the LLM through iterative feedback, adjusting item text with reasoning-, and review-based content to guide the target item toward the desired ranking.

Model	Method	Home & Kitchen			Tools & Home Improvement			Electronics			Sports & Outdoors			Health & Household
Model	Method	Top-5	Top-3	Top-1	Top-5	Top-3	Top-1	Top-5	Top-3	Top-1	Top-5	Top-3	Top-1	Top-5	Top-3	Top-1
GPT-4o	Baseline	0.0%	0.0%	0.0%	0.0%	0.0%	0.0%	0.0%	0.0%	0.0%	0.0%	0.0%	0.0%	0.0%	0.0%	0.0%
	String	55.0%	45.5%	33.0%	54.5%	44.0%	32.5%	57.0%	46.0%	34.5%	56.0%	45.0%	34.0%	55.5%	45.5%	33.5%
	Reasoning	92.0%	87.0%	81.0%	93.0%	88.0%	82.5%	94.5%	89.5%	84.0%	92.5%	87.5%	82.0%	93.5%	88.0%	83.0%
	Review	89.5%	85.0%	79.0%	90.0%	85.5%	80.0%	91.0%	86.0%	80.5%	90.5%	85.5%	80.0%	91.5%	86.5%	81.0%
Gemini-2.5	Baseline	0.0%	0.0%	0.0%	0.0%	0.0%	0.0%	0.0%	0.0%	0.0%	0.0%	0.0%	0.0%	0.0%	0.0%	0.0%

Table 1: Promotion Success Rate on ProductBench across 15 categories.

Method	GPT-4o			Gemini-2.5			Claude-4			Grok-3
Method	Top-5	Top-3	Top-1	Top-5	Top-3	Top-1	Top-5	Top-3	Top-1	Top-5	Top-3	Top-1
CORE-String	55.0%	45.5%	33.0%	53.5%	43.0%	31.0%	58.0%	46.5%	34.0%	51.5%	42.0%	30.5%
CORE-Reasoning	92.0%	87.0%	81.0%	88.5%	84.0%	77.5%	95.0%	90.0%	83.5%	88.0%	83.0%	77.0%
CORE-Review	89.5%	85.0%	79.0%	94.5%	89.5%	83.5%	93.5%	89.0%	82.0%	90.5%	85.0%	78.5%
STS	54.0%	44.0%	32.5%	53.0%	42.5%	31.5%	57.0%	45.0%	33.5%	52.0%	42.0%	30.0%
TAP	61.0%	47.5%	34.0%	49.5%	39.5%	29.5%	62.5%	48.5%	34.5%	50.0%	40.5%	29.5%

Table 2: Promotion success rate with existing rankings methods on Home & Kitchen.

Model	Baseline	String	Reasoning	Review
GPT-4o	30.84	1,526.84	72.45	31.87
Gemini-2.5	31.11	1,488.32	68.22	32.65
Claude-4	29.95	1,572.15	75.11	31.23
Grok-3	30.47	1,505.67	70.88	32.14

Table 3: Average perplexity scores across categories.

Strategy	Fluency (1–5)	Detection Rate (%)
Baseline	4.7	12.0
String	1.2	98.5
Reasoning	3.7	62.1
Review	4.6	18.4

Table 4: Human evaluation results. Fluency is rated on a 1–5 scale.

Position	GPT-4o			Gemini-2.5			Claude-4			Grok-3
Position	1st Rank	2nd Rank	3rd Rank	1st Rank	2nd Rank	3rd Rank	1st Rank	2nd Rank	3rd Rank	1st Rank	2nd Rank	3rd Rank
1st String	4.0%	20.0%	76.0%	3.0%	21.0%	76.0%	5.0%	19.5%	75.5%	2.5%	22.0%	75.5%
2nd Reasoning	38.0%	44.5%	17.5%	35.0%	46.0%	19.0%	36.0%	45.0%	19.0%	34.0%	47.0%	19.0%
3rd Review	58.0%	35.5%	6.5%	62.0%	33.0%	5.0%	59.0%	34.5%	6.5%	61.0%	33.5%	5.5%
1st String	2.5%	24.0%	73.5%	3.0%	23.0%	74.0%	4.0%	22.5%	73.5%	3.5%	23.0%	73.5%
2nd Review	66.0%	25.5%	8.5%	64.5%	26.0%	9.5%	65.0%	26.5%	8.5%	63.5%	27.0%	9.5%

Table 5: Ranking distributions under six insertion orders on Home & Kitchen.

Config	Method	GPT-4o			Gemini-2.5			Claude-4			Grok-3
Config	Method	Top-5	Top-3	Top-1	Top-5	Top-3	Top-1	Top-5	Top-3	Top-1	Top-5	Top-3	Top-1
Generator = GPT-4o	Reasoning	92.0%	87.0%	81.0%	86.0%	80.0%	74.5%	91.0%	85.5%	80.0%	85.5%	79.5%	73.0%
Generator = GPT-4o	Review	89.5%	85.0%	79.0%	92.0%	87.5%	81.5%	92.0%	87.0%	81.0%	88.0%	83.0%	77.0%
Generator = Gemini-2.5	Reasoning	89.5%	84.0%	77.5%	88.5%	84.0%	77.5%	91.0%	86.0%	80.0%	84.5%	78.0%	72.5%
Generator = Gemini-2.5	Review	87.0%	80.5%	76.0%	94.5%	89.5%	83.5%	84.0%	79.0%	73.5%	89.5%	84.0%	78.0%
Generator = Claude-4	Reasoning	90.0%	85.5%	79.5%	87.5%	82.0%	76.0%	95.0%	90.0%	83.5%	86.5%	79.5%	74.0%

Table 6: Promotion success rate with different Generator–Optimizer configurations.

Model	GPT-4o			Gemini-2.5			Claude-4			Grok-3
Model	Top-5	Top-3	Top-1	Top-5	Top-3	Top-1	Top-5	Top-3	Top-1	Top-5	Top-3	Top-1
Llama-3.1-8B	55.0%	45.5%	33.0%	53.5%	43.0%	31.0%	58.0%	46.5%	34.0%	51.5%	42.0%	30.5%
Vicuna-13B	50.5%	41.0%	28.0%	48.5%	39.0%	27.0%	54.0%	42.5%	29.5%	46.5%	38.0%	26.0%
Mistral-7B	46.0%	37.5%	24.0%	44.5%	35.5%	23.0%	49.0%	39.0%	26.0%	42.0%	34.5%	21.5%

Table 7: Promotion success rate using different shadow models.

Appendix figures & tables10 assets

Supplementary material from the paper’s appendix.

Appendix

Method	GPT-4o			Gemini-2.5			Claude-4			Grok-3
Method	Top-5	Top-3	Top-1	Top-5	Top-3	Top-1	Top-5	Top-3	Top-1	Top-5	Top-3	Top-1
CORE-String	54.5%	44.0%	32.5%	52.5%	42.5%	31.0%	57.5%	46.0%	33.5%	52.0%	42.5%	31.0%
CORE-Reasoning	93.0%	88.0%	82.5%	87.0%	83.0%	77.0%	92.5%	88.0%	81.0%	89.5%	84.0%	78.5%
CORE-Review	90.0%	85.5%	80.0%	95.0%	90.0%	84.0%	95.5%	90.5%	83.5%	88.0%	83.5%	77.0%
STS	49.0%	39.0%	27.5%	48.0%	38.0%	26.0%	52.0%	41.0%	29.0%	47.0%	37.0%	25.0%
TAP	61.0%	47.0%	33.0%	44.0%	34.0%	23.5%	63.5%	48.5%	34.5%	45.0%	35.0%	23.0%

Appendix

Table 8: Promotion Success Rate with existing ranking methods on Tools & Home Improvement.

Method	GPT-4o			Gemini-2.5			Claude-4			Grok-3
Method	Top-5	Top-3	Top-1	Top-5	Top-3	Top-1	Top-5	Top-3	Top-1	Top-5	Top-3	Top-1
CORE-String	57.0%	46.0%	34.5%	54.0%	43.5%	32.0%	59.0%	47.5%	34.5%	50.5%	41.5%	30.0%
CORE-Reasoning	94.5%	89.5%	84.0%	89.0%	85.0%	78.5%	96.0%	91.0%	85.0%	87.0%	82.5%	76.5%
CORE-Review	91.0%	86.0%	80.5%	96.0%	91.0%	85.0%	94.0%	89.5%	83.0%	91.5%	86.0%	79.0%
STS	47.0%	36.5%	24.5%	45.0%	35.0%	23.0%	50.0%	39.0%	27.0%	44.0%	34.0%	22.5%
TAP	58.5%	44.0%	30.0%	42.0%	33.0%	21.5%	60.5%	45.0%	31.0%	43.0%	32.5%	21.0%

Appendix

Table 9: Promotion Success Rate with existing ranking methods on Electronics.

Real LLM	LLM-based Similarity (1–5)	Human Similarity (1–5)
GPT-4o	4.5	4.3
Gemini-2.5	4.2	4.0
Claude-4	4.7	4.5
Grok-3	3.8	3.7

Appendix

Table 10: Similarity between shadow model (Llama-3.1-8B) and real LLMs on 10 fixed product recommendation samples from Home & Kitchen .

Model	Noise Level ( $\sigma$ )	Top-5	Top-3	Top-1
GPT-4o	0	48.5%	38.0%	26.5%
	0.05	55.0%	45.5%	33.0%
	0.2	41.0%	31.5%	20.0%
Gemini-2.5	0	46.0%	36.5%	25.0%
	0.05	53.5%	43.0%	31.0%
	0.2	39.0%	30.0%	19.0%

Appendix

Table 11: Promotion success rate under different noise levels

\sigma

Metric	Mean	Std
Cosine similarity	0.81	0.06

Appendix

Table 12: Cosine similarity between the optimized continuous embedding and its discretized text

$\tau$	GPT-4o			Gemini-2.5			Claude-4			Grok-3
$\tau$	Top-5	Top-3	Top-1	Top-5	Top-3	Top-1	Top-5	Top-3	Top-1	Top-5	Top-3	Top-1
0.5	78.5%	70.0%	62.0%	76.0%	68.0%	60.0%	82.0%	74.0%	65.0%	72.0%	65.0%	57.0%
0.6	88.0%	82.0%	76.0%	86.0%	80.0%	74.0%	91.0%	85.0%	78.0%	83.0%	77.0%	70.0%
0.7	89.5%	85.0%	79.0%	94.5%	89.5%	83.5%	93.5%	89.0%	82.0%	90.5%	85.0%	78.5%
0.8	89.5%	85.0%	79.0%	94.0%	89.0%	83.0%	93.0%	88.5%	82.0%	90.0%	84.5%	78.0%
0.9	89.0%	84.5%	79.0%	94.0%	89.0%	83.0%	93.0%	88.5%	82.0%	90.0%	84.5%	78.0%

Appendix

Table 13: Promotion success rate under different similarity thresholds

\tau

on Home & Kitchen using the Review-based strategy

Defense	Method	GPT-4o			Gemini-2.5			Claude-4			Grok-3
Defense	Method	Top-5	Top-3	Top-1	Top-5	Top-3	Top-1	Top-5	Top-3	Top-1	Top-5	Top-3	Top-1
/	String	55.0%	45.5%	33.0%	53.5%	43.0%	31.0%	59.5%	48.0%	35.0%	51.5%	42.0%	30.5%
	Reasoning	92.0%	87.0%	81.0%	88.5%	84.0%	77.5%	94.0%	89.0%	82.0%	88.0%	83.0%	77.0%
	Review	89.5%	85.0%	79.0%	94.5%	89.5%	83.5%	95.0%	90.5%	83.5%	90.5%	85.0%	78.5%
Perplexity-based	String	0.0%	0.0%	0.0%	0.0%	0.0%	0.0%	0.0%	0.0%	0.0%	0.0%	0.0%	0.0%
	String	(55.0% $\downarrow$ )	(45.5% $\downarrow$ )	(33.0% $\downarrow$ )	(53.5% $\downarrow$ )	(43.0% $\downarrow$ )	(31.0% $\downarrow$ )	(59.5% $\downarrow$ )	(48.0% $\downarrow$ )	(35.0% $\downarrow$ )	(51.5% $\downarrow$ )	(42.0% $\downarrow$ )	(30.5% $\downarrow$ )

Appendix

Table 14: Promotion success rate under different defense strategies across models.

Dataset	Method	GPT-4o			Gemini-2.5			Claude-4			Grok-3
Dataset	Method	Top-5	Top-3	Top-1	Top-5	Top-3	Top-1	Top-5	Top-3	Top-1	Top-5	Top-3	Top-1
Document	String	58.5%	48.5%	33.5%	56.0%	46.0%	31.0%	61.5%	50.0%	34.0%	54.0%	44.5%	30.5%
	Reasoning	91.0%	86.0%	80.0%	88.0%	83.0%	76.5%	93.0%	88.5%	81.0%	86.5%	82.0%	75.0%
	Review	88.0%	83.0%	76.0%	90.5%	85.5%	78.0%	89.5%	85.0%	78.0%	86.0%	81.0%	74.5%
Travel	String	53.5%	43.0%	30.0%	51.0%	41.0%	28.5%	55.0%	44.5%	30.5%	49.5%	39.5%	27.0%
	Reasoning	87.5%	82.0%	75.5%	84.0%	78.0%	71.0%	89.5%	83.5%	76.0%	82.5%	76.0%	69.0%

Appendix

Table 15: Top-

k

promotion success rates beyond product search

Transfer	Method	Gemini-2.5			Claude-4			Grok-3
Transfer	Method	Top-5	Top-3	Top-1	Top-5	Top-3	Top-1	Top-5	Top-3	Top-1
w/o	String	53.5%	43.0%	31.0%	59.5%	48.0%	35.0%	51.5%	42.0%	30.5%
	Reasoning	88.5%	84.0%	77.5%	94.0%	89.0%	82.0%	88.0%	83.0%	77.0%
	Review	94.5%	89.5%	83.5%	95.0%	90.5%	83.5%	90.5%	85.0%	78.5%
w/	String	52.0%	41.5%	29.5%	54.0%	44.5%	31.0%	49.0%	39.0%	28.0%
	String	(1.5% $\downarrow$ )	(1.5% $\downarrow$ )	(1.5% $\downarrow$ )	(5.5% $\downarrow$ )	(3.5% $\downarrow$ )	(4.0% $\downarrow$ )	(2.5% $\downarrow$ )	(3.0% $\downarrow$ )	(2.5% $\downarrow$ )

Appendix

Table 16: Transfer performance across models with and without transfer

Model	Avg. Loops	Tokens / Loop	Total Tokens / Item	Cost / Item
GPT-4o	3.1	1,284.7	3,982.6	$0.0398
Gemini-2.5	4.6	1,912.3	8,806.6	$0.0110
Claude-4	3.7	1,356.8	5,018.2	$0.0452
Grok-3	4.1	1,487.9	6,100.4	$0.0549

Appendix

Table 17: Computation cost per item across different models

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

Aligning Large Language Models with Searcher Preferences
Mar 11, 2026Wei Wu, Peilun Zhou, Liyi Chen +6Personalized Language Model Alignment Generative Retrieval
School of Artificial Intelligence and Data Science, University of Science and Technology of China, Hefei, China · Xiaohongshu Inc., Beijing, China · Thrust of Artificial Intelligence, The Hong Kong University of Science and Technology (Guangzhou), Guangzhou, China +1
CHASE: How Content Ecosystems Are Reshaped When Ranking Is the Only Target
Aug 31, 2026Qianwen Gao, Zichang Su, Yiwen Hou +2Generative Engine Optimization
University of California, Berkeley · Zhejiang University
Role-Augmented Intent-Driven Generative Search Engine Optimization
Aug 15, 2025Xiaolu Chen, Jie Bao, Haojie Wu +3Generative Engine Optimization
School of Cyber Science and Technology, University of Science and Technology of China, Hefei 230026, China.

Mar 11, 2026J/K move · Enter open · S save

Aligning Large Language Models with Searcher Preferences

Wei Wu, Peilun Zhou, Liyi Chen, Qimeng Wang+5

School of Artificial Intelligence and Data Science, University of Science and Technology of China, Hefei, China · Xiaohongshu Inc., Beijing, China · Thrust of Artificial Intelligence, The Hong Kong University of Science and Technology (Guangzhou), Guangzhou, China+1

Personalized Language Model Alignment Generative Retrieval

Open paper AskPDF arXiv

The paradigm shift from item-centric ranking to answer-centric synthesis is redefining the role of search engines. While recent industrial progress has applied generative techniques to closed-set item ranking in e-commerce, research and deployment of open-ended generative search on large content platforms remain limited. This setting introduces challenges, including robustness to noisy retrieval, non-negotiable safety guarantees, and alignment with diverse user needs. In this work, we introduce SearchLLM, the first large language model (LLM) for open-ended generative search. We design a hierarchical, multi-dimensional reward system that separates bottom-line constraints, including factual grounding, basic answer quality and format compliance, from behavior optimization objectives that promote robustness to noisy retrieval and alignment with user needs. Concretely, our reward model evaluates responses conditioned on the user query, session history, and retrieved evidence set, combining rule-based checks with human-calibrated LLM judges to produce an interpretable score vector over these dimensions. We introduce a Gated Aggregation Strategy to derive the training reward for optimizing SearchLLM with Group Relative Policy Optimization (GRPO). We deploy SearchLLM in the AI search entry of RedNote. Offline evaluations and online A/B tests show improved generation quality and user engagement, increasing Valid Consumption Rate by 1.03% and reducing Re-search Rate by 2.81%, while upholding strict safety and reliability standards.

Abstract

Generative Engine Optimization

Explore similar work

Aligning Large Language Models with Searcher Preferences

CHASE: How Content Ecosystems Are Reshaped When Ranking Is the Only Target

Role-Augmented Intent-Driven Generative Search Engine Optimization