A Hybrid Multi-Agent Conversational Recommender System with LLM and Search Engine in E-commerce

Field and quasi-field designs demonstrate that performance enhancements can lead to increased option acceptance and GMV per user.

“Field and quasi-field designs demonstrate that enhancements in performance can lead to quantifiable business results, such as increased option acceptance and GMV per user [10]”

Behavioral outcomes such as conversion, click-through, repurchase, and GMV are less common but provide stronger external validity.

“Behavioral outcomes (like conversion, click-through, repurchase, and GMV) are less common but provide stronger external validity (e.g., [10,41,55]).”

Where behavioral or platform-embedded evidence is present, the same rationale can lead to quantifiable performance enhancement, such as increased acceptance and revenue/GMV through improved conversational recommendation, instead of being solely intention-driven.

“In instances where behavioral or platform-embedded evidence is present, the same rationale can lead to quantifiable performance enhancement (e.g., increased acceptance and revenue/GMV through improved conversational recommendation) instead of being solely intention-driven [10,55].”

Field data support that performance enhancements can result in increased option acceptance and revenue metrics.

“It is further supported by field data demonstrating that performance enhancements can result in increased option acceptance and revenue metrics [10].”

The review includes a small number of field or quasi-field studies that show the strongest link to real outcomes, but these studies are often hard to replicate across platforms and agent architectures because they fail to report enough on assignment, exposure windows, and guardrails.

“The review includes a small number of field or quasi-field studies that show the strongest link to real outcomes (like acceptance/conversion, spend/GMV), but these studies are often hard to replicate across platforms and agent architectures because they fail to report enough on assignment, exposure windows, and guardrails [10,55].”

Field evidence is often reported with little methodological detail, such as short A/B reports with unclear sampling or randomization.

“Field evidence is not uncommon, and when it is, it is often reported with little methodological detail (for example, short A/B reports with unclear sampling/randomization) [10].”

Genuinely agentic execution (cart/checkout/payment) and negotiation are still uncommon, and the empirical base is still focused on low-autonomy, text-based assistants with outcomes dominated by intentions and satisfaction constructs.

“First, genuinely agentic execution (cart/checkout/payment) and negotiation are still uncommon, and the empirical base is still focused on low-autonomy, text-based assistants with outcomes dominated by intentions and satisfaction constructs [10,40].”

Field and quasi-field contributions enhance inferences regarding market outcomes, but methodological limitations constrain generalizability and the testing of mechanisms.

“Field and quasi-field contributions enhance inferences regarding market outcomes; however, methodological limitations-including ambiguous assignment, platform specificity, abbreviated observation periods, and analytical simplifications-constrain generalizability and the testing of mechanisms [10,55].”

Spend/GMV/WTP is positive in field and some survey settings.

“Positive in field and tings some survey set-”

Behavioral lift appears when systems reduce friction (latency/steps) or improve option fit (recommendation relevance).

“Behavioral lift appears when systems reduce fric-havioral intention rather than verified execution. tion (latency/steps) or improve option fit (recom-mendation relevance).”

Observed conversion/purchase behavior is positive when friction drops or recommendations fit, but not universal.

“Positive when fric-tion drops or rec-ommendations fit, but not universal”

In many studies, conversion is proxied via self-reported ordering/behavioral intention rather than verified execution.

“In many studies,”