PaperlayerView exact revision on arXiv

Source-linked AI summary

Agentic Share-of-Search: A Multi-Agent AI System for Competitive Decision-Making in LLM-Mediated E-Commerce

Spandan Ghose Chowdhury

arXiv:2609.11190v1cs.AIcs.IR

TL;DR

As discovery shifts toward AI-mediated recommendations, sellers lack infrastructure for measuring SKU-level competitive visibility and diagnosing gaps. This paper introduces an ASoS-based multi-agent decision-support system and evaluates its diagnostic agent in a 100-trial signal-ablation feasibility study. The agent recovers the ablated signal in 39.0% of trials overall and 63.9% among top-three-correlation ablations.

Abstract

from arXiv · show

AI shopping assistants increasingly redirect consumer discovery, creating an urgent need for tools that support seller-side competitive decision-making. We present a multi-agent AI system that automates competitive visibility measurement and root cause diagnosis in LLM-mediated ecommerce. The system introduces Agentic Share-of-Search (ASoS) as the decision target, deploys query agents across leading AI platforms, and uses a ReAct-based diagnostic agent to recommend prioritized merchandising interventions. A 100-trial ablation study, presented as a feasibility evaluation of this prototype, shows the agent recovers the ablated signal in 39% of trials (95% CI: 30.0% - 48.8%, 5.5x over chance), rising to 63.9% among high-correlation ablations.

INTRODUCTION

As product discovery shifts to AI recommendations, sellers lack SKU-level tools for measuring competitive visibility and connecting gaps to merchandising actions. The paper frames ASoS as retailer-attributed visibility and presents a decision-support pipeline for measuring and diagnosing those gaps.

INTRODUCTION

  • AI-mediated recommendations are replacing some search-result discovery, while existing seller tools lack LLM-response measurement, SKU-level visibility, or attribution from signals to interventions.Traditional recommender metrics evaluate buyer-side utility, whereas the seller-side question is how often products appear relative to competitors and what actions to take.Source: From consumer-side recommendation to… · p. 1, INTRODUCTION · p. 1

What ASoS measures: the philosophy of the metric

  • ASoS measures retailer-attribution-weighted visibility rather than catalog-derived product presence, targeting the retailer attribution that merchandising teams can influence.The metric is intended to reflect where an AI assistant says consumers should buy, weighted by recommendation rank.Source: What ASoS measures: the philosophy of… · p. 2

An AI system for managerial competitive decisions

  • The system measures retailer visibility in AI-generated recommendations, diagnoses catalog-signal gaps, and gives human merchandising teams ranked intervention recommendations.It is presented as a prototype and feasibility study, not a validated substitute for managerial judgment.Source: An AI system for managerial… · p. 2

Contributions

  • The contributions include ASoS and counting rules, a multi-agent measurement architecture, a ReAct diagnostic agent, a 100-trial synthetic-ablation evaluation, and a controlled-experiment design.The experiment design is intended to connect identified gaps and interventions to measured visibility uplift.Source: Contributions · p. 2

LITERATURE REVIEW

The paper adapts agentic root-cause analysis and shopping-agent infrastructure to seller-side e-commerce decisions. It positions product-level diagnosis as a bridge between visibility tracking, optimization, and traditional share-of-search.

AI agents for business decision support

LLM web-browsing agents and shopping-agent benchmarks

  • Shopping-agent benchmarks assess agents as buyers, while this work repurposes query-agent infrastructure for seller-side decision support.The cited benchmark examples include DeepShop (Lyu et al., 2025), Shopping Companion (2026), and customer-simulation frameworks (Zhu et al., 2025).Source: LLM web-browsing agents and… · p. 3

Generative engine optimization and AI brand visibility

  • The paper bridges GEO research and commercial visibility tracking by introducing a product-level decision-support agent; commercial tools track brand-citation frequency but operate at brand rather than SKU level.Commercial tools offer no intervention attribution. GEO content presentation tactics can boost a source’s visibility in LLM responses by up to 40% (Aggarwal et al., 2024).Source: Generative engine optimization and AI… · p. 3

Share-of-search as a competitive metric

  • ASoS extends Share-of-Search beyond branded Google and Bing queries to AI-assistant surfaces and unbranded category queries.Share-of-Search operationalized branded organic search volume as a leading indicator of market share.Source: Share-of-search as a competitive metric · p. 3

METHODOLOGY

The methodology builds a competitive query-and-SKU measurement pipeline, enriches catalog signals, and uses an association-based diagnostic agent to prioritize merchandising recommendations. Its measurement and evaluation are framed as prototype work; a controlled intervention experiment is future work.

Stage 1: Query universe and competitive set

  • The system defines a query universe, competing retailers, and representative SKUs across price tiers, then targets four AI platforms by design.The mock evaluation exercises three platform clients and the live pilot covers Gemini only; Google AI Overviews is deferred to future live measurement.Source: Stage 1: Query universe and… · p. 3

Stage 2: Multi-agent query execution and ASoS aggregation

  • Query agents capture repeated platform responses, extract and resolve product mentions, and aggregate rank-discounted weighted appearance into retailer ASoS.The pipeline runs weekly and flags volatility; counting rules exclude responses without extractable mentions, deduplicate repeated SKUs at their earliest rank, and split attribution weights across retailers when needed.Source: Stage 2: Multi-agent query execution… · p. 4

Stage 3: Competitive signal database

  • The signal database combines pricing, content quality, reviews, catalog coverage, shipping, and availability attributes that merchandising teams can modify.Quantitative attributes use deterministic parsers, while qualitative content-quality signals use rubric-based, structured-output model scoring.Source: Stage 3: Competitive signal database · p. 5

Stage 4: Decision module: ReAct diagnostic agent

  • For below-benchmark retailer-cluster pairs, the diagnostic agent correlates SKU appearance with catalog signals and ranks recommendations by correlation magnitude and observed gap.A portfolio-wide validation checks whether the hypothesized signal gap holds beyond the cluster sample; the output is an association-based diagnostic, not a causal effect estimate.Source: Stage 4: Decision module: ReAct… · pp. 5–6
  • The ReAct loop validates leading candidates against portfolio-wide signal deltas and emits a recommendation for the first validated candidate or after testing up to five.It is single-shot per retailer-cluster pair, with correlation and deterministic SOP-validation tools.Source: Diagnostic agent implementation · p. 6

Stage 5: System implementation and architecture

  • The implementation supports uniform platform clients, pluggable entity resolution, structured run logs, and snapshots of leaderboards, recommendations, resolutions, and configuration.The controlled intervention experiment needed to establish causal effects is specified as future work.Source: METHODOLOGY · p. 3, Stage 5: System implementation and… · pp. 6–7

EMPIRICAL RESULTS

Across 100 synthetic signal-ablation trials, the diagnostic agent recovered the ablated signal in 39.0% of trials, with higher precision among top-correlated signals. Supporting evaluations cover a diagnostic case, resolver robustness, a live-platform pilot, and run provenance.

EMPIRICAL RESULTS

  • The running-shoes evaluation used 23 SKUs across 7 retailers, 34 queries, and three platforms, producing 306 raw responses and 1,524 extracted mentions.The evaluation included three repeats per query-platform pair.Source: EMPIRICAL RESULTS · p. 7

Diagnostic-agent reliability via signal ablation: a feasibility evaluation

  • The 100-trial protocol perturbed one of 14 signals and scored whether the agent’s top-1 recommendation named it; this tests signal recovery, not independent business-effect ground truth.The agent prioritizes signals associated with visibility, so weakly correlated ablations may be de-prioritized; the evaluation is a feasibility result for the prototype.Source: Diagnostic-agent reliability via… · p. 7
  • 39.0% raw Precision@1 (39/100; Wilson 95% CI: 30.0%–48.8%) was a 5.5× lift over the 7.1% expected from uniform random selection.Among trials whose ablated signal was among the top three by absolute correlation, Precision@1 was 63.9% (39 of 61).Source: Diagnostic-agent reliability via… · p. 7

Qualitative case study of the diagnostic chain

  • For adidas in the feature-query cluster, ASoS was 7.24% versus a peer mean of 14.29%, and the agent recommended increasing has_qa with low effort.The recommendation received SOP validation across the retailer’s full SKU portfolio.Source: Qualitative case study of the… · p. 8

Pipeline robustness and live-platform validation

  • The embedding resolver achieved 99.7% per-mention resolution at 8.2 seconds, while QLoRA achieved 100.0% at 125.5 seconds; both yielded the same retailer leaderboard.In the Gemini pilot, the extractor found a product mention with non-null brand attribution in 21 of 30 responses (70.0%); the leaderboard showed a sharper top-1 share than mock data, reflecting marketplace attribution.Source: Pipeline robustness and live-platform… · pp. 8–9

Provenance and reproducibility

  • Each pipeline invocation logged configuration and outputs, while artifact hashes enabled byte-level reconciliation across runs.For the resolver comparison, extraction hashes matched while resolution hashes differed.Source: Provenance and reproducibility · p. 9

DISCUSSION

The ablation results offer feasibility evidence for agent-supported merchandising, while synthetic data, limited live testing, and confounded associations constrain the claims. The discussion identifies operational fairness concerns and proposes controlled, longitudinal intervention testing as a stronger next step.

Feasibility evidence for AI-supported merchandising decisions

  • 39/100 trials recovered the ablated signal overall; recovery was 63.9% when the signal ranked among the top three correlates and 100% for the dominant correlate.The conditional result characterizes performance on stronger correlates, not overall recovery; recommendations prioritize estimated business impact, and the prototype warrants human review.Source: Feasibility evidence for AI-supported… · p. 9

Boundary conditions and limitations

  • Most analysis uses synthetic responses, while the Gemini pilot covers 10 queries and 30 calls; live extraction success falls to 70% and retailer attribution differs from mock data.The ablation cannot establish causal ground truth, the diagnostic associations are confounded, extraction and attribution lack human-coding validation, and the main study covers one category, platform, and period.Source: Boundary conditions and limitations · pp. 9–10

Safety, ethics, and competitive fairness

  • Scaling catalog-enrichment recommendations may widen the gap between well-resourced and smaller sellers, so the paper proposes monitoring long-tail ASoS share and marketplace-versus-DTC leaderboard divergence.The paper also flags adversarial generative-engine optimization as a risk to monitor in production.Source: Safety, ethics, and competitive fairness · p. 10

Future work: proposed experimental design for intervention closure

  • A matched treatment-control study could measure weekly ASoS for four weeks after selected catalog changes, using competitor ASoS as a covariate to account for platform drift.This intervention-level test is proposed to supply ground truth that the ablation cannot; the broader agenda includes multi-platform live measurement and real web-scraped signals.Source: Future work: proposed experimental… · p. 10

CONCLUSION

The paper presents a multi-agent system for competitive visibility measurement, diagnosis, and ranked merchandising recommendations. Its ablation evaluation supports feasibility, while controlled interventions, human-coding audits, and broader longitudinal measurement remain necessary before treating it as a reliable proxy for managerial judgment.

CONCLUSION

  • The system measures retailer visibility, correlates gaps with per-SKU catalog-enrichment signals, and uses a ReAct-based agent to produce ranked, interpretable interventions.The paper frames these results as an association-based decision-support prototype for human merchandising teams.Source: CONCLUSION · pp. 10–11
  • 39.0% of 100 trials recovered the ablated signal overall, a 5.5× lift over random selection; recovery reached 63.9% for top-three correlates and 100% for the dominant correlate.A live Gemini 2.5-flash pilot confirms that the data substrate operates on real LLM output, while final intervention selection remains for human review.Source: CONCLUSION · p. 11

DECLARATION ON GENAI USAGE

  • The author reports using Anthropic Claude for code scaffolding, manuscript drafting, and editorial review, and taking responsibility for the publication.Source: DECLARATION ON GENAI USAGE · p. 11
Loading 2609.11190v1…