cs.CLOct 16, 2026

GLOBE: Evolving Structured Agentic Memory for E-commerce Product Optimization in Generative Search

Authors: Hyunseo Kim, Suil Choe, Sukkyoung So, Daeyeon Yi

Organizations: AI Lab, CTO Division, LG Electronics, South Korea

Read PDF
Ask AI
Cite

Copy citation

DOINot available for this paper

Abstract

Optimizing product listings for generative search requires language agents not only to rewrite effectively but also to continually self-improve by retaining successful strategies across diverse categories. However, existing approaches often treat structured listings as free-form text, leaving effective strategies implicit in prompts, search trajectories, or rewritten outputs rather than accumulating them as a reusable agentic memory. To address this limitation, we introduce GLOBE, an agentic framework that drives continual self-improvement by storing optimization knowledge in an evolving handbook of explicit, field-scoped behaviors. In each episode, an Editor applies relevant behaviors and receives rank-based feedback, which a Reflector converts into behavior updates. Crucially for dependable execution in real-world settings, a Moderator removes redundant proposals, a Curator commits accepted changes, and consistently harmful behaviors are deactivated. This cycle transforms episodic feedback into a persistent, field-scoped optimization memory. Experiments show that structured behaviors outperform monolithic prompt optimization, while feedback-driven evolution and moderation further improve effectiveness and reduce downside risk. The learned behaviors also remain effective on unseen product categories, suggesting that agentic memory evolution supports interpretable and transferable product optimization.

Figures & tables

Figure 1: Comparison between monolithic prompt optimization and Globe ’s structured handbook optimization. A generic rewrite prompt performs an undifferentiated text transformation, whereas each Globe behavior specifies both a concrete editing operation and its permitted target field, enabling constrained reuse across products.
Figure 2: Overview of the Globe framework. Each iteration deploys the active handbook Hn\mathcal{H}_{n} over a batch of KK query-product pairs: the editor rewrites the target listing and the ranker returns episode-level rank feedback. The Reflector then proposes behavior additions or modifications, the Moderator removes redundant proposals, and the Curator commits the survivors into Hn+1\mathcal{H}_{n+1} , which replaces Hn\mathcal{H}_{n} as the handbook deployed over the next batch. Repeating this deployment-and-evolution cycle accumulates episodic feedback into persistent editing knowledge.
MethodComponentsEffectivenessRisk & ConsistencyHandbook Size
HandbookFeedbackModeratorMRG ↑\uparrowSR ↑\uparrowDSR ↓\downarrowCV ↓\downarrow
Prompt Opt.–––0.35144.1%3.9610.117–
Static HB✓––1.34358.6%2.6280.0429
Globe w/o Mod.✓✓–1.51960.6%2.3890.05451
Globe✓✓✓1.61161.6%2.1720.03447
Table 1: Performance comparison of optimization methods. Higher values are better for MRG and SR ( ↑\uparrow ), while lower values are better for DSR and CV ( ↓\downarrow ). Best results are shown in bold and second-best results are underlined.
Figure 3: Performance of the trained handbook as the number of training queries increases.
MethodGPT-5.4 (Self)Qwen3.5-122B-A10BGemini-3.5-flash
MRG ↑\uparrowSR ↑\uparrowDSR ↓\downarrowCV ↓\downarrowMRG ↑\uparrowSR ↑\uparrowDSR ↓\downarrowCV ↓\downarrowMRG ↑\uparrowSR ↑\uparrowDSR ↓\downarrowCV ↓\downarrow
Prompt Opt.0.35144.1%3.9610.117−- 0.03122.1%1.7000.044−- 0.05121.5%1.3790.019
Static HB1.34358.6%2.6280.0421.33550.3%0.4350.0480.96045.5%0.2900.015
Globe w/o Mod.1.51960.6%2.3890.0541.16746.6%0.1150.0100.97546.6%0.3710.021
Globe1.61161.6%2.1720.0341.52254.1%0.1050.0341.07848.9%0.2460.013
Table 2: Cross-ranker evaluation of optimization strategies trained with GPT-5.4.
Figure 4: Cross-category generalization, measured as MRG over the seed handbook.
Query: “Aux cable for the car that won’t break after a month of use.” Target Field: features
Delta operations committed into H4\mathcal{H}_{4}
modify B012Name the listing’s material and build detail most relevant to the likely wear point (from: name the listing’s material and build details)
add B016When the query describes how past products failed , add a top-placed clause mapping construction evidence to that exact failure mode
Resulting feature bullets
H2\mathcal{H}_{2}rank 4→54\rightarrow 5 , gi=−1g_{i}=-1 (i) tight spaces and connector strain often cause early cable failure (ii) reduces bend stress at the jack for better day-to-day durability
H4\mathcal{H}_{4}rank 4→04\rightarrow 0 , gi=+4g_{i}=+4 (i) tight spaces and repeated bending often cause cable failure [B016] (ii) reduces strain at the connector to address a common breakage point [B012]
Table 3: Behavior evolution from feedback to ranking effect for the target product CableCreation 6 FT Flat 3.5mm Aux Cable . Full example in Appendix E .
Appendix figures & tables11 assets

Supplementary material from the paper’s appendix.

Appendix

ModelMRG ↑\uparrowSR ↑\uparrowDSR ↓\downarrowCV ↓\downarrowHB
GPT-5.41.61161.6%2.1720.03447
Qwen3.5-122B-A10B1.50756.5%2.1060.05829
Gemini-3.5-flash1.70060.8%2.3100.02846
Appendix
Table 4: Performance evaluation of the GLOBE framework across different LLM backends.
IDNameFieldsConfl.Instruction
B001AddQuantitativeStatistics (Statistics Addition)F, D—Replace vague qualitative claims with specific quantitative data from the listing: exact dimensions, capacities, wattage, weight, counts, percentages, runtime hours, or temperature ranges. Present numbers in a way that directly answers likely shopper questions.
B002CiteCredibleEvidence (Cite Sources)F, D—When the listing mentions certifications, standards, test results, material grades, or industry ratings (e.g., UL-listed, FDA-approved, IP67, ASTM-certified, Energy Star), surface and attribute these explicitly as evidence of quality or compliance.
B003IncludeAuthoritativeQuotesOrClaims (Quotation Addition)DB009When the listing contains manufacturer statements, expert-backed claims, or notable brand heritage, present them as direct attributable claims (e.g., ‘Designed by [Brand]’s engineering team for…’ or ‘Meets [Standard] requirements for…’) to add credibility.
B004OptimizeFluencyAndFlow (Fluency Optimization)D—Restructure fragmented bullet dumps, incomplete sentences, and disjointed listing text into coherent, well-flowing prose that reads naturally while preserving all factual content. Ensure logical sentence progression from product identity to key features to use context.
B005SimplifyForAccessibility (Easy-to-Understand)F, DB007Rewrite overly technical or jargon-heavy descriptions in plain language that a general shopper can understand, while keeping the precise specs available. Convert insider terminology to buyer-friendly equivalents (e.g., ‘18/10 stainless steel’ →\rightarrow ‘18/10 stainless steel (restaurant-grade, resists corrosion)’).
B006UseAuthoritativeTone (Authoritative)DB009Write in a confident, expert-advisor tone that positions the product as a considered recommendation rather than a sales pitch. Use decisive language (‘delivers,’ ‘provides,’ ‘handles’) instead of weak hedging (‘might work,’ ‘could be suitable’).
Appendix
Table 5: The seed handbook H1\mathcal{H}_{1} (nine behaviors). Fields lists the permitted target fields targets(b)\mathrm{targets}(b) : T = title, F = features, D = description. Confl. lists the behaviors registered as conflicting, which the editor is instructed not to apply jointly. Italicized names in parentheses give the originating GEO strategy of Aggarwal et al. (2023) ; B008 and B009 are introduced by this work for the structured e-commerce setting. B009 is the single deliberately prohibitive behavior in the framework and acts as a factuality guardrail; the Reflector is prevented from proposing further restriction-only behaviors (Table 11 ).
 Pre-modificationPost-modification  
BehaviorHRnnHRnnΔ\Delta MRGpp
B00159.1%8856.0%736-0.080.625
B003 †45.9%13560.8%51+0.690.029
B00456.0%10055.8%883+0.110.355
B00556.7%9755.4%839+0.070.431
B00856.0%10055.8%882+0.110.356
Appendix
Table 6: Episode-level outcomes of modified behaviors before their first and after their last deployed modify during training. HR is the helpful rate HR(b)=h+(b)/n(b)\mathrm{HR}(b)=h^{+}(b)/n(b) , nn the number of applied episodes, Δ\Delta MRG the change in mean rank gain, and pp a one-sided Mann–Whitney test on episode gains. † marks behaviors whose pre-modification HR was below the pooled helpful rate of the same batches, i.e., behaviors modified while under-performing. Modified behaviors with fewer than 10 applied episodes on either side are omitted.
IDName and instructionFieldsh+/h−/h0h^{+}/h^{-}/h^{0}Active span
B014SurfacePriceAndBudgetFit When budget is a primary query constraint and the listing shows a current price, place the exact price or under-budget fit in the opening sentence or first feature.F, D0/3/2H3→H10\mathcal{H}_{3}\rightarrow\mathcal{H}_{10}
B023NameClosureAttachmentOrMountingMethod When the query cares about how the product fastens, mounts, seals, straps on, or attaches, add a top-placed phrase naming the exact listing-supported attachment method and the surface or item it secures to.F, D3/6/3H5→H6\mathcal{H}_{5}\rightarrow\mathcal{H}_{6}
B032NameAestheticStyleMatch When the query specifies a visual style, finish, or design language, place a short early phrase naming the exact listing-supported style cue, color, material look, or silhouette that matches it.F, D6/7/5H7→H8\mathcal{H}_{7}\rightarrow\mathcal{H}_{8}
B046MapQueryAlternativesAndFallbacks When the query allows alternatives, ranges, or optional variants such as ‘X or Y’, ‘at least’, or ‘preferably’, add one early sentence mapping the listing’s exact match, acceptable fallback, or nearest supported option in the same order as the query.F, D1/2/4H10→H11\mathcal{H}_{10}\rightarrow\mathcal{H}_{11}
Appendix
Table 7: Behaviors deactivated during the final Globe run, i.e., with n(b)≥Nmin⁡n(b)\geq N_{\min} and s(b)<0s(b)<0 . Instructions are the last versions deployed before deactivation, and Fields follows Table 5 . Active span gives the handbook version in which each behavior was added and the first version in which it was inactive. Deactivated behaviors remain stored for traceability.
Query: "Aux cable for the car that won’t break after a month of use." (Product: CableCreation 6 FT Flat 3.5mm Aux Cable)
Section: [Features]
Behavior Evolution
Behavior IDHandbook v2Handbook v4
B012 (Modified)When durability or longevity is part of the query, add a dedicated phrase naming the listing’s material and build details —such as full-grain leather, steel housing, stitched seams, cast metal, reinforced joints, or ceramic burrs—and place it near the top.When durability or longevity is part of the query, add a top-placed phrase naming the listing’s material and build detail most relevant to the likely wear point —such as steel housing, stitched seams, reinforced joints, rust-resistant metal, thick insulation, or etched markings.
B016 (Created)— Not present in v2 —When the query describes how past products failed, add a dedicated clause near the top that maps listing materials or construction to that exact failure mode (e.g., reinforced hinges to prevent handle breakage, rust-resistant metal for wet storage, stitched seams to resist splitting).
Product Rewriting
Appendix
Table 8: End-to-end qualitative example of behavior evolution.
Prompt for the generative ranker M\mathcal{M}
System message
You are ChatGPT, a large language model trained by OpenAI.
You are helping a user find the best product for their query.
Be direct and honest in your assessment. Rank products based on how well they match the query.
User message
Here is my query: {query}
Appendix
Table 9: Prompt for the generative ranker M\mathcal{M} . The same prompt is used to obtain the original ranking rM(q,P,pt)r_{\mathcal{M}}(q,\mathcal{P},p_{t}) and the post-rewrite ranking rM(q,P′,pt′)r_{\mathcal{M}}(q,\mathcal{P}^{\prime},p^{\prime}_{t}) ; the two calls differ only in whether the target slot of {formatted products} contains the original or the rewritten listing. For the Prompt Opt. baseline we instead use the full ChatGPT system prompt released with E-GEO Bagga et al. (2025) to reproduce their setting exactly.
Prompt for the Product Editor
System message
You are an expert in e-commerce product optimization.
Your task is to rewrite product descriptions so they rank higher in LLM-based product comparisons.
Follow the provided editing rules precisely. Maintain factual accuracy.
Return ONLY the rewritten description, no explanations.
User message
Appendix
Table 10: Prompt for the product editor. The complete active behavior set is supplied as a strategy pool and the editor selects the subset relevant to the current query–product pair. The scope constraint block is emitted only when at least one listing field lies outside the union of the permitted target fields, which enforces the field scoping of the handbook. The self-reported APPLIED line is parsed during post-processing to obtain Ai\mathcal{A}_{i} .
Prompt for the Reflector
System message
You are an expert in product description optimization and prompt engineering.
Analyze episode logs from a product description rewriting system and propose improvements to the behavior handbook.
The system:
1. Provides ALL behaviors from the handbook to the writer for each query
2. The writer applies whichever behaviors are relevant and self-reports which ones it used
Appendix
Table 11: Prompt for the Reflector. Episode summaries list, for every episode, the query identifier, product category, signed rank change, self-reported applied behaviors, and a query snippet.
Prompt for the Moderator (functional deduplication)
System message
You are a behavior handbook curator. Your job is to find NEW behaviors that are functionally redundant with existing ones. Two behaviors are redundant if they trigger on the same query pattern AND produce the same type of edit, even if they use different wording or slightly different examples.
User message
## Handbook ( {n} behaviors)
[NEW] | [EXISTING] [ {behavior id} ] {behavior name}
instruction: {instruction}
Appendix
Table 12: Prompt for the functional-deduplication stage of the Moderator. It is invoked only on proposals that survive embedding-based deduplication ( τ=0.9\tau=0.9 ), and it sees the projected handbook state in which each behavior is tagged as [EXISTING] or [NEW]. Existing behaviors can never be dropped by this stage.
Prompt for the Prompt Opt. baseline (initial rewriting prompt)
System message
You are an expert ml researcher having previous background in SEO and search engines in general. You are working on novel research ideas for next generation of e-commerce websites. These websites will have language models augmented with search engines, with the task of recommending products based on the catalog of products backed by the search engine. This new set of systems will be collectively called language engines (generative search engines). This will require e-commerce businesses to update their SEO techniques so that the product ranks higher in the llm generated answer. Specifically they will use GEO (Generative Engine Optimization) techniques to boost their visibility in the final rankings outputted by the Language Engine.
User message ( authoritative variant, used as the epoch-0 prompt)
Transform the following product description into an authoritative style without adding or removing any core information.
The revised description should reflect confidence, expertise, and assertiveness, while maintaining the original content’s meaning and relevance.
The description should be assertive in its statements, such that reader believes that this is more valuable product than other products.
Appendix
Table 13: Prompt for the Prompt Opt. baseline. The rewriting prompt is initialized with the authoritative strategy of E-GEO Bagga et al. (2025) — their default — and is then iteratively replaced by the meta-optimizer output in Table 14 . The whole listing is passed as a single {description} field, so the baseline has no notion of field scope.
Prompt for the Prompt Opt. baseline (meta-optimizer)
System message
You are an expert in prompt engineering, marketing, and writing. Your goal is to improve the rewriting prompt so that when it rewrites a product description, that product ranks higher compared to other products.
User message
I need you to improve a prompt that rewrites product descriptions to make them rank higher in product comparisons.
CONTEXT:
We have a system where:
Appendix
Table 14: Meta-optimizer prompt of the Prompt Opt. baseline. After each batch, the meta-optimizer receives the aggregate rank-gain statistics of the current rewriting prompt together with the history of previous epochs, and returns a replacement prompt. This is the prompt-level analogue of GLOBE’s handbook evolution: the optimization knowledge is carried entirely by a single free-form instruction rather than by structured, field-scoped behaviors.
More in this topic

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

Explore similar work

CardsList

  1. MemRerank: Preference Memory for Personalized Product Reranking

    Mar 31, 2026Zhiyuan Peng, Xuyang Wu, Huaixiao Tou +2LLM PersonalizationUser Preference Modeling

    Santa Clara University · Independent Researcher

  2. AgenticGEO: A Self-Evolving Agentic System for Generative Engine Optimization

    Mar 2, 2026Jiaqi Yuan, Jialu Wang, Zihan Wang +3Generative Engine Optimization

    School of Computer Science and Engineering, Beihang University, Beijing, China · Independent Contributor, CA, United States

  3. Generalizable Self-Evolving Memory for Automatic Prompt Optimization

    Mar 23, 2026Guanbao Liang, Yuanchen Bei, Sheng Zhou +5Automatic Prompt OptimizationLanguage Model Self-Improvement

    Zhejiang University · University of Illinois Urbana-Champaign · Alibaba Group