Large language models (LLMs) increasingly shape recommendations across chatbots, e-commerce, and user-generated content (UGC) platforms, creating demand for Generative Engine Optimization (GEO) to improve item visibility in generated responses. Existing GEO studies are mainly content-centric, optimizing visibility under general-purpose queries. However, a shared item description usually serves users with different needs and decision goals, resulting in its visibility gains being dependent on the request. To investigate how changing intent affects visibility, we formulate IntentGEO to evaluate shared rewrites across need specificity and decision progression. We revisit existing GEO methods and find that their visibility gains vary across user intents. More critically, learning-based prompt optimization can underperform its handwritten seed on average. Tracing these updates reveals prompt drift from explicit recommendation instructions toward generic promotional tactics. To address this drift, we propose AMO (Anchored Meta-Optimization), which preserves recommendation instructions in a fixed core while learning appended strategies from citation feedback. Across four domains and ten intent states, AMO improves average citation-rank gain by 35% over the strongest prior learning-based baseline and leads on nine states. Ablations show that learned additions improve on the fixed core alone and that AMO outperforms full-prompt optimization initialized from that core. It also achieves the highest gains in an agentic recommendation pipeline and, after separate adaptation, on Doubao, a commercially deployed, closed-source online assistant. These findings highlight user intent as a key dimension of GEO evaluation and support anchored adaptation across recommendation needs and decision stages.
Figures & tables
Figure 1. Intent-aware recommendation GEO. (a) User needs vary in specificity and decision stage. (b) P/Q queries probe these two dimensions. (c) AMO preserves a recommendation core while learning appended strategies. Three panels show recommendation contexts, five levels of need specificity and decision progression, and an anchored meta-optimization loop that learns additions while keeping core recommendation instructions fixed.
Figure 2. Base visibility and GEO effectiveness vary across recommendation requests. Top: Base citation rank (lower is better). Bottom: rank gains of the five strongest original baselines (higher is better); negative gains indicate lost visibility. Two columns compare need specificity P1 through P5 and decision progression Q1 through Q5. Base rank decreases across P levels, while the five GEO baselines lose gain and most become negative at P5. Q-series gains remain positive but vary by intent state.
Method
P5
P-Avg
Q-Avg
All
RecGEO seed
+0.01
+1.14
+0.77
+0.95
MetaOpt
-0.42
+0.91
+0.68
+0.80
MetaOpt-Rec
-0.66
+0.79
+0.56
+0.67
MetaOpt-Prog
-0.42
+0.98
+0.68
+0.83
MetaOpt-Rec-Prog
-0.58
+0.85
+0.57
+0.71
Table 1. RecGEO and four MetaOpt variants. The “-Rec” variants use RecGEO as the seed prompt, and “-Prog” variants replace each original training query with a randomly sampled P- or Q-series variant.
Instruction marker
Seed
Selected
Updates
Full retention
Yes
No
3/36
Recommender voice
Yes
No
0/36
Keywords
No
Yes
36/36
Call to action
No
Yes
36/36
Table 2. Literal instruction markers before and after MetaOpt-Rec optimization. “Updates” counts saved prompts containing each marker.
Method
Need specificity
Decision progression
Average
P1
P2
P3
P4
P5
Q1
Q2
Q3
Q4
Q5
P
Q
All
Base
9.33
8.98
7.37
5.87
2.50
7.27
7.83
8.70
8.33
9.39
6.81
8.30
7.56
Authoritative
+1.36
+0.54
+0.21
+0.40
-0.12
+0.00
+0.17
+0.14
+0.15
+0.35
+0.48
+0.16
+0.32
Statistics
+1.19
+0.48
+0.21
+0.24
-0.18
-0.04
+0.19
-0.14
+0.07
+0.36
+0.39
+0.09
+0.24
Citations
+1.20
+0.56
+0.28
+0.36
-0.14
+0.05
+0.19
+0.11
+0.21
+0.30
+0.45
+0.17
+0.31
Fluency
+1.19
+0.71
+0.34
+0.45
-0.27
+0.14
+0.25
+0.28
+0.24
+0.27
+0.48
+0.23
+0.36
Table 3. Main benchmark results in IntentGEO . The Base row reports citation ranks; other rows report rank gains ( ↑ ). Best gains are bold with a light red background; second-best gains are underlined with a gray background. Ties follow displayed values.
Method
Rank gain ↑
P-Avg
Q-Avg
All
Core only
1.15
0.80
0.98
Full-prompt opt.
0.84
0.64
0.74
AMO
1.22
1.02
1.12
Table 4. Rank gains for AMO and its ablation variants.
Method
P3
P4
P5
AMO
14.5
8.5
5.5
RecGEO
11.0
9.0
5.5
MetaOpt-Prog
10.5
9.0
1.5
AutoGEO
6.0
3.5
0.5
Table 5. Visibility gains under paraphrased P3–P5 requests. Δ Pass@5 reports percentage-point gains over Base in the top-five target citation rate.
Method
Natural
Consist
Salience
Avg.
Content Impr.
4.82
4.21
4.83
4.62
AMO
4.71
4.01
4.99
4.57
MetaOpt
4.71
4.02
4.98
4.57
AutoGEO
4.59
4.18
4.82
4.53
RecGEO
4.54
4.03
4.99
4.52
MetaOpt-Rec
4.32
4.00
4.99
4.43
Table 6. LLM-as-judge quality scores for rewritten descriptions. Avg. is the mean across dimensions.
Figure 3. Generation-stage GEO across three assistant workflows. Fixed candidate evidence is supplied directly (a), returned by the agent’s search tool (b), or submitted through Doubao’s official API with native search disabled (c). Three evaluation workflows. In the static setting, the model receives candidate descriptions directly. In the agentic setting, the assistant issues a search query, receives candidate evidence through a tool, and generates a recommendation. The third setting evaluates rewriting through the official Doubao assistant API.
Method
P-Avg
Q-Avg
All
Pass@5 (%)
Base
—
—
—
33.85
Content Impr.
0.621
0.546
0.584
40.67
RecGEO
0.800
0.660
0.730
42.25
AutoGEO
0.599
0.482
0.540
40.72
MetaOpt
0.737
0.505
0.621
41.62
AMO
0.929
0.734
0.832
43.48
Table 7. Agentic recommendation without retraining. P/Q/All report mean rank gains over Base. Pass@5 is the top-five target citation rate.
Method
P-Avg
Q-Avg
All
Pass@5 (%)
Base
—
—
—
43.52
Content Impr.
0.379
0.383
0.381
48.34
RecGEO
0.719
0.980
0.849
54.67
AutoGEO
0.746
1.065
0.906
56.68
MetaOpt
0.419
0.492
0.456
49.40
MetaOpt-Rec
0.710
0.873
0.791
54.02
Table 8. Doubao assistant evaluation with fixed evidence. P/Q/All: mean rank gains over Base; Pass@5: top-five target citation rate (%).
Method
ΔRU
ΔRC
Nnew (%)
Content Impr.
1.470
-0.059
149 (26.0)
RecGEO
2.596
0.145
233 (40.7)
AutoGEO
2.813
0.136
256 (44.8)
MetaOpt-Rec
2.467
0.116
222 (38.8)
AMO
2.846
0.322
264 (46.2)
Table 9. Online Doubao: mean rank gains for Base-uncited ( ∣U∣=572 ) and Base-cited ( ∣C∣=1,418 ) cases. New citations report counts and percentages of U .
Appendix figures & tables8 assets
Supplementary material from the paper’s appendix.
Appendix
Dataset
Train
Val.
Test
Cand.
Books
174
25
50
9.02
Retail
200
25
50
10.00
Videogames
200
25
50
10.00
Ragdoll
0
0
50
8.00
Total
574
75
200
9.26
Appendix
Table 10. Dataset splits before intent expansion. Cand. denotes the average candidate pool size per test instance.
Setting
Value
Primary backbone
gpt-4o-mini
Rewrite / answer / query temperature
0.7 / 0.7 / 0.3
Top- p / random seed
1.0 / 42
Epochs / batches per epoch / batch size
3 / 12 / 50
Rewrite variants per selected prompt
3
LLM judge / temperature
gpt-4o-mini / 0.0
Appendix
Table 11. Shared experimental settings for recommendation evaluation and prompt optimization.
Method
P1
P2
P3
P4
P5
Avg.
Pass@5
Δ R
Pass@5
Δ R
Pass@5
Δ R
Pass@5
Δ R
Pass@5
Δ R
Pass@5
Δ R
Authoritative
27.7 ± 3.1
+1.36 ± 0.15
24.3 ± 0.8
+0.54 ± 0.05
39.5 ± 0.7
+0.21 ± 0.14
58.5 ± 3.2
+0.40 ± 0.24
86.7 ± 1.2
-0.12 ± 0.04
47.3 ± 1.4
+0.48 ± 0.06
Statistics
24.0 ± 1.9
+1.19 ± 0.06
23.5 ± 1.6
+0.48 ± 0.13
38.3 ± 0.2
+0.21 ± 0.03
56.0 ± 0.7
+0.24 ± 0.03
86.5 ± 0.0
-0.18 ± 0.02
45.7 ± 0.4
+0.39 ± 0.02
Citations
24.8 ± 1.3
+1.20 ± 0.14
25.3 ± 0.9
+0.56 ± 0.10
39.7 ± 3.1
+0.28 ± 0.19
57.8 ± 1.7
+0.36 ± 0.11
86.3 ± 1.6
-0.14 ± 0.13
46.8 ± 0.6
+0.45 ± 0.07
Fluency
24.3 ± 1.2
+1.19 ± 0.02
27.3 ± 0.8
+0.71 ± 0.03
41.0 ± 1.1
+0.34 ± 0.06
58.8 ± 2.4
+0.45 ± 0.18
85.2 ± 0.5
-0.27 ± 0.04
47.3 ± 0.3
+0.48 ± 0.03
Unique Words
22.5 ± 1.6
+0.99 ± 0.09
21.5 ± 1.1
+0.35 ± 0.07
36.8 ± 1.0
+0.03 ± 0.08
52.3 ± 0.8
+0.01 ± 0.05
85.2 ± 0.2
-0.26 ± 0.03
43.7 ± 0.5
+0.22 ± 0.02
Appendix
Table 12. Full P-query results: mean ± population SD across three rewrites. Pass@5 in %.
Field
Example
Query
recommend me the very best book about Investing for beginners
Original
Stock Investing For Dummies, 3rd Edition includes information on stock investing in both bear and bull markets; unique investment segments; stock investing for different types of situations; and examples straight from the real world of stock investing as they have occurred in the past three years.
MetaOpt
Stock Investing For Dummies, 3rd Edition includes information on stock investing in both bear markets and bull markets. Stock Investing For Dummies, 3rd Edition includes information on unique investment segments.
AMO
Why Stock Investing For Dummies, 3rd Edition is recommended: Stock Investing For Dummies, 3rd Edition is recommended because it combines dual market-regime coverage (bear and bull), unique investment segments, situation-specific guidance, and three-year real-world examples in a single volume …
Service answer
For new investors seeking a comprehensive, all-market guide, Stock Investing For Dummies, 3rd Edition stands out, … covers strategies for both bull and bear markets, … [5].
Rank
Base: uncited; MetaOpt: uncited; AMO: first cited
Appendix
Table 16. Online Doubao: AMO first, MetaOpt uncited.
Δ MRR
Δ Pass
Method
@5
@10
@1
@3
@5
AMO
.0720
.0752
4.03
9.13
13.98
RecGEO
.0599
.0635
3.58
6.52
11.75
MetaOpt-Prog
.0522
.0546
2.72
6.73
10.45
AutoGEO
.0522
.0537
2.90
6.50
9.65
Content Impr.
.0516
.0542
3.12
6.20
9.57
Appendix
Table 17. Top-K gains from Base. Pass: percentage points.
Method
P-Avg
Q-Avg
AMO (original anchor)
12.25
13.25
AMO (paraphrased anchor)
15.50
13.50
RecGEO
14.25
9.75
MetaOpt-Prog
13.75
8.25
AutoGEO
8.25
7.25
Appendix
Table 18. Anchor-paraphrase robustness. P-Avg and Q-Avg report mean Δ Pass@5 over P1–P5 and Q1–Q5, respectively (percentage points over Base).
Method
GPT-4o-mini
Gemini
Qwen
Content Impr.
0.79
0.41
0.48
RecGEO
0.95
0.54
0.44
AutoGEO
0.77
0.41
1.10
MetaOpt-Prog
0.83
0.50
0.65
AMO
1.12
0.62
1.54
Appendix
Table 19. Mean rank gain ΔR over Base across all intent states and three rewrites. Gemini and Qwen denote Gemini 2.5 Flash-Lite and Qwen3-30B-A3B, respectively.
Setting
Contrast
Difference [95% CI]
Main
AMO–RecGEO
+0.172[+0.072,+0.271]
Main
AMO–MetaOpt-Prog
+0.293[+0.175,+0.412]
Main
AMO–Core only
+0.147[+0.056,+0.236]
Main
AMO–Full-prompt opt.
+0.384[+0.268,+0.502]
Main P5
AMO–Base
+0.063[−0.305,+0.435]
Base
P2–Q3
+0.290[−0.040,+0.635]
Appendix
Table 20. Paired rank comparisons and 95% confidence intervals. Main gains average all states unless marked P5; Base compares unrounded saved P2/Q3 ranks.
Department of Computer Science, University of Toronto · Department of Mechanical & Industrial Engineering, University of Toronto · Faculty of Information, University of Toronto +3