Predicting What Answer Engines Cite: Open Decision Models, Rerankers and a Distilled 4B Student on Ten Engines under Controlled Retrieval

Elisha Bajemon; Andre-Louis Rochet

cs.IROct 28, 2026

Predicting What Answer Engines Cite: Open Decision Models, Rerankers and a Distilled 4B Student on Ten Engines under Controlled Retrieval

Authors: Elisha Bajemon, Andre-Louis Rochet

Organizations: TW3 Partners (Citead)

Abstract

Generative engine optimization needs a scorer that ranks candidate pages by the share of an answer they will receive as citations. We measure pretrained rerankers, a hosted typed decision model (Jev), eight open decision models and distilled students on GEO-Bench queries answered by ten engines that receive the same candidate sources in their prompt. Where several engines answer the same query, they agree with one another at a within-group Spearman correlation of 0.704. Jev and two open 27B decision models, used without training, are the strongest scorers at 0.575 to 0.578. A Qwen3-Reranker-4B fine-tuned with low-rank adapters on teacher labels for other GEO-Bench queries, released as GEO Citation Ranker 4B, reaches 0.542 to 0.549 across three seeds, and its differences with the 27B models and with Jev are not distinguishable from zero on 229 queries. Most of the remaining distance between the student and Jev disappears when Jev scores one passage at a time, as the student does. Students trained on teacher labels and on observed engine citations reach the same level, and labels from Jev alone match the two-teacher target at 0.6B. We release the model, the benchmark, the predictions of every reported scorer and the code.

Figures & tables

Reference points
Scorer	Size	$\rho$ [95% CI]	Top-1	GEO Citation Ranker 4B $-$ scorer
Other engines on the same query a	–	0.704 [0.669, 0.737]	–	$-$ 0.159 [ $-$ 0.201, $-$ 0.116]
Source position in the prompt	–	0.131 [0.068, 0.190]	0.328	–
BM25	–	0.131 [0.063, 0.203]	–	–
Pretrained rerankers, zero-shot
NIM rerank-qa-mistral-4b	4B	0.386 [0.329, 0.449]	0.475	+0.156 [+0.104, +0.212]

Table 1: Citation ranking on 229 queries and 1220 engine-query groups.

\rho

is the mean within-group Spearman correlation with observed citation share; the last column is the paired difference with GEO Citation Ranker 4B, with its 95% query-clustered interval. Decision models are shown with the listwise protocol; Table 2 gives both protocols. A random choice reaches a top-1 accuracy of 0.23. a Cross-engine agreement (Definition 3) on the 1141 groups whose query has at least two engines; its paired difference is computed on those groups. b The five-passage state exceeds the model’s 2048-token context for 23 queries; its row uses the remaining 1011 groups.

Figure 1: Within-group Spearman correlation with observed citation share, with 95% query-clustered intervals, for a selection of the scorers of Table 1 . The shaded band is the cross-engine agreement and its interval. Decision models are shown with the listwise protocol.

Model, zero-shot	Size	Five passages per call	One passage per call	Difference [95% CI]
Jev 1.13 (hosted)	n/a	0.575	0.529	+0.046 [+0.009, +0.087]
JEV-27B	27B	0.578	0.540	+0.038 [+0.004, +0.071]
pplx-decider-v1.1-27b	27B	0.575	0.559	+0.015 [ $-$ 0.013, +0.047]
JEV-9B	9B	0.516	0.468	+0.047 [ $-$ 0.003, +0.099]
NeoHorse-Jev-4B	4B	0.501 b	0.465	+0.049 [0.000, +0.100]
decider-4b	4B	0.480	0.490	$-$ 0.009 [ $-$ 0.049, +0.031]

Table 2: Decision models under the two protocols, with identical state and questions. The difference is listwise minus pointwise on the groups both cover. b 1011 groups in the listwise column (context limit). c The model’s 1024-token context truncates the five-passage state. The student scores one passage per call by construction.

Student	Training labels	Training queries	$\rho$ [95% CI]
0.6B	Jev and LLM judge	757	0.457 [0.402, 0.512]
0.6B	Jev only	757	0.467 [0.417, 0.521]
0.6B	LLM judge only	757	0.444 [0.388, 0.498]
0.6B	Jev and LLM judge	189	0.427 [0.366, 0.485]
0.6B	Jev and LLM judge, cross-query distractors	757	0.472 [0.416, 0.525]
4B	Jev and LLM judge	757	0.542 [0.491, 0.591]

Table 3: Label ablations, all with the training recipe of Section 4 and seed 42. a Trained on the evaluation queries by five-fold cross-validation, each group scored by the model that did not see its query; the two-stage student starts from the teacher-trained 4B adapter.

Appendix figures & tables1 asset

Supplementary material from the paper’s appendix.

Appendix

Engine	Vendor	Weights	Queries	Groups	Zero labels	Ranker 4B	Jev	JEV-27B	NIM
mistral-medium-3.5-128b	Mistral AI	open	229	228	22%	0.538	0.571	0.574	0.395
gemini-3.1-flash-lite	Google	closed	150	149	20%	0.599	0.650	0.621	0.425
gpt-5.5	OpenAI	closed	150	143	28%	0.533	0.547	0.552	0.340
gpt-5.4	OpenAI	closed	150	149	24%	0.537	0.579	0.576	0.384
gpt-5.4-mini	OpenAI	closed	150	149	28%	0.556	0.544	0.557	0.405
gemma-4-31b-it	Google	open	100	99	23%	0.607	0.622	0.650	0.411

Appendix

Table 4: Engine configurations. Queries answered, non-degenerate groups, share of sources with a zero citation share, and mean within-group Spearman of four scorers per engine; Ranker 4B is GEO Citation Ranker 4B.

Generative Engine Optimization

Generative Engine Optimization studies how to make web content visible, cited, and accurately represented in AI-generated search answers.

See all 41 papers

What Gets Cited: Competitive GEO in AI Answer Engines
May 25, 2026Rahul Vishwakarma, Shushant Kumar, Ratnesh JamidarGenerative Engine Optimization
Sprinklr · Gurugram, India · Dubai, UAE
System One models Are Not Better Rerankers
Oct 23, 2026Michael Dinzinger, Kanishka Ghosh Dastidar, Michael GranitzerLLM Reranking Information Retrieval
University of Passau, Passau, Germany · Interdisciplinary Transformation University Austria (IT:U), Linz, Austria
What Drives AI Citations? Content, Retrieval and Model Knowledge in Generative Engine Optimization (GEO)
Oct 10, 2026Mahe Chen, Nick KoudasGenerative Engine Optimization
University of Toronto, Canada

May 25, 2026J/K move · Enter open · S save

What Gets Cited: Competitive GEO in AI Answer Engines

Rahul Vishwakarma, Shushant Kumar, Ratnesh Jamidar

Sprinklr · Gurugram, India · Dubai, UAE

Generative Engine Optimization

Open paper AskPDF arXiv

AI answer engines generate answers from retrieved pages but cite only a few sources. This makes visibility depend not just on ranking, but on being cited. We study competitive Generative Engine Optimization (GEO): when two retrieved candidates compete, what makes one more likely to be cited first? We build a controlled two-document retrieval-augmented generation (RAG) testbed that injects exactly two candidate sources into the model context and measures which source is referenced by the first citation marker in the output. Across six LLMs we execute 252,000 trials, repeated paired comparisons under one factorial program over 18 content factors. In each trial the two sources differ in exactly one factor; we use brand anonymization and counterbalanced source order to separate content effects from position bias. Mixed-effects models show that topical relevance and list position are the biggest drivers of being cited first. Including explicit price information and a recent timestamp also helps consistently. Completeness and trust cues add smaller gains, while formatting-only edits have little impact. We release a reproducible evaluation protocol and a prioritized GEO checklist for practitioners, and we exercised it in an early internal pilot at Sprinklr, where teams reported positive qualitative feedback on workflow usability.

Abstract

Generative Engine Optimization

Explore similar work

What Gets Cited: Competitive GEO in AI Answer Engines

System One models Are Not Better Rerankers

What Drives AI Citations? Content, Retrieval and Model Knowledge in Generative Engine Optimization (GEO)