before launch. automatically extracts structured data, which is then evaluated by a user-configurable population of reviewer agents. These agents generate simulated reviews, which an analyst agent subsequently processes (sentiment, key aspects) and visualizes on a dashboard. The goal is to predict product reception and identify strengths/weaknesses Review simulation, multi-agent systems, web scraping, product evaluation The launch of new products into the market carries significant risks, with consumer acceptance being a critical factor for success. Understanding how a product will be perceived by diferent segments of the target audience before investing considerable resources in its production and marketing is of vital strategic importance. Traditionally, this understanding is obtained through market studies, surveys, or focus groups-methods that can be costly in terms of time and resources, and do not always dynamically capture the diversity of opinions.
CEUR Workshop
ISSN1613-0073
Benefits”) explores the practical uses of the platform as a preliminary market research tool and the advantages it ofers. Finally, Section 6 (”Conclusions and Future Work”) summarizes the contributions of the work and outlines lines for future improvement and development of the system.
Deep networks are also defining the state-of-the-art in opinion mining, like in aspect-based sentiment
analysis [
In recent years, the capabilities of Large Language Models (LLMs) to emulate diferent human profiles
has open a new topic focused on the use of personas as artificial profiles for many diferent tasks,
mainly in the creation of synthetic data for model tranining [
The application of artificial populations has not been yet explored in opinion mining, to the best of our knowledge, as a means to evaluate commercial products, ofering a prospective information of potential weaknesses and strengths through artificial opinions. This approach could benefit from the maturity of opinion mining techniques to evaluate products from personas’ opinions.
The proposed system follows a modular workflow implemented through a multi-agent architecture. Each type of agent has a specific responsibility within the overall simulation and analysis process, as illustrated in Figure 1.
The process begins with user intervention and unfolds through the following main stages: (1) Data input and information extraction, where product information is obtained; (2) Specification of parameters and generation of reviewer agents, where the simulated population is configured; (3) Generation of simulated reviews, where agents evaluate the product; and (4) Aggregated analysis and visualization, where results are processed and presented.
The user provides the URL of a product webpage of interest (Figure 2). The page is not required to belong to a specific domain; the system is designed to be generic. A first agent, the Extractor Agent (represented with a pickaxe in Figure 1), is responsible for performing web scraping on the provided URL.
This agent navigates the page, identifies, and extracts relevant product information such as name,
description, technical features, price, and possibly images or key specifications [
A crucial step in the simulation is the detailed configuration of the Population of Reviewer Agents. To facilitate this process, the system provides dedicated graphical interfaces that allow the user to define multiple facets of these virtual agents and their behavior when generating reviews. Initially, through the interface shown in Figure 4, the user establishes the general parameters of the simulated population and the desired characteristics for the reviews. The system enables the configuration of various aspects, including: – The total size of the population. – Demographic attributes such as age range, education level (allowing for specific or mixed profiles), and gender distribution. • Textual properties of the resulting reviews: – Positivity bias. – Verbosity (length and elaboration).
– The product-specific level of detail they should include.
For deeper modeling, the user can refine the Personality Profile of the agents using the interface in Figure 5. This allows adjusting values across various trait spectrums, defining tendencies on axes such as: • Introverted/Extroverted • Independent/Cooperative • Analytical/Creative
The combination of these demographic, review style, and personality parameters gives the user the ability to model with great granularity the target audience whose reaction is to be simulated. These traits are expected to influence each agent’s perspective, their points of focus or criticism, and the overall tone of the reviews they generate. Finally, once the user completes and confirms the configuration (e.g., using the ”Generate bot profiles” button visible in Figure 5), the Profile Generation Agent uses the complete set of specifications to generate the reviewers’ profiles Figure 6.
Upon user review of the generated profiles in the profile panel (Figure 6), the information is sent to the Reviewer Generator component (see Figure 1). Each profile is then used to instantiate a Reviewer Agent, which is ready to evaluate the product from its unique perspective.
Once the structured product information (JSON) is available and the population of reviewer agents has
been created, the product information is distributed to each agent in the population. Each individual
Reviewer Agent (represented by the robot thinking about the product in Figure 1) processes the
product information through the prism of its own defined parameters and personality. Using LLM
models, each agent generates a simulated textual review (the prompt used is included in Appendix
A). This review will reflect its particular evaluation of the product, highlighting positive or negative
aspects based on its configured criteria [
The set of simulated reviews generated by the population is sent to an Analyst Agent (represented by
the robot with a magnifying glass over a web page in Figure 1). This agent is tasked with processing
and analyzing all the reviews. Analysis techniques performed automatically by the LLM can include:
• Sentiment analysis: Determine the overall polarity (positive, negative, neutral) of the reviews
[
The results of this analysis are consolidated and presented to the end-user through a Dashboard
(represented by the JSON with graphs at the end). This panel (Figure 8) visualizes key findings, allowing
the user to quickly and efectively understand the simulated product reception, identify patterns,
and detect perceived strengths and weaknesses (Figure 9) by the simulated target audience. Evolved
sentiment analysis techniques allow for a more nuanced understanding of the opinions expressed in
reviews [
The implementation of the Product Review Simulator prototype is based on a set of technologies selected for their suitability for developing agent-based systems and interactive applications. The key components of the technology stack are as follows: 1. Agent Orchestration Framework: CrewAI. The core of the simulation, i.e., the management and operation of the reviewer agent population, was implemented using CrewAI1. This Python framework is specifically designed to facilitate the creation of autonomous multi-agent systems. Its capability to define: • Agents: Represent the agents mentioned in the architecture description. • Tasks: Define the tasks assigned to the agents.
• Tools: ScrapeWebsiteTool2 was used to extract information from the product URL. CrewAI allowed abstracting much of the complexity associated with coordinating and executing tasks among multiple agents. 2. Backend Framework: Flask. Flask was used to build a lightweight and eficient API, facilitating communication between the user interface, the simulator logic, and the agents. 3. User Interface: React. React enabled the development of a dynamic and interactive interface for visualizing the reviews generated by the agents and managing simulations. 1https://www.crewai.com/open-source 2https://docs.crewai.com/concepts/tools
4. Large Language Models (LLMs). The natural language generation capability of the agents is based on integration with large language models. Specifically, Gemini 2.0 Flash was used, with its interaction managed through CrewAI.
Together, this technological framework provided the necessary tools to eficiently develop a prototype that credibly simulates the generation of product reviews by configurable agents.
The main application of this prototype lies in its ability to function as a tool for preliminary market research and product concept evaluation. Before making significant investments in development, production, or large-scale marketing campaigns, companies can use this simulation to: • Identify Perceived Strengths and Weaknesses: The simulation reveals which product features are likely to be best valued and which could generate criticism or dissatisfaction in a specific market segment (defined by agent parameters). • Test the Reaction of Niche Markets: By configuring the agent population to represent a particular demographic or psychographic group (e.g., ”young university students interested in sustainability,” ”middle-aged professionals sensitive to price”), it’s possible to predict how that specific niche might react to the product. • Compare Product Variants: Simulations could be run for diferent hypothetical versions of a product (e.g., with diferent features or prices) to assess which would have a better simulated reception. • Refine Marketing Strategies: Understanding which aspects resonate most (positively or negatively) with the simulated audience can help focus marketing messages on the most relevant strengths and proactively address potential weaknesses. • Reduce Risks and Costs: By obtaining early simulated feedback, more informed decisions can be made, potentially avoiding costly design or market positioning errors.
The automated nature of scraping and review generation/analysis allows these insights to be obtained
much faster and potentially at a lower cost than traditional market research methods, especially in the
initial ideation and design phases. Aspect-based sentiment analysis (ABSA) provides a more granular
understanding of opinions on specific product features, which is crucial for informed decision-making
[
We have presented a prototype of a multi-agent system for product review simulation. The proposed architecture integrates specialized agents for data extraction, configurable generation of reviews based on simulated consumer profiles, and aggregated analysis of these reviews, culminating in a visual dashboard. This tool ofers a novel (To the best of our knowledge there are no similar work in generating artificial opinions in e-commerce platforms) and automated approach to gaining early insight into the possible reception of a product in a target market.
Potential benefits include the early identification of strengths and weaknesses, the ability to test scenarios with diferent target audiences, and the reduction of risks associated with launching new products.
As future work, we plan to refine the internal models of the reviewer agents, possibly
incorporating advanced language models (LLMs) to generate more realistic and nuanced reviews [
Additionally, a future research line will systematically investigate how agent parameters such as demographics and personality traits concretely influence the prompts provided to the language models and, consequently, the review generation process (Section 3.3).
Regarding the data extraction component (Section 3.1), the robustness of the web scraping process across diferent website structures remains to be thoroughly evaluated. Future work will include extensive testing and refinement of scraping methods.
This work has been partially supported by projects CONSENSO (PID2021-122263OB-C21), MODERATES (TED2021-130145B-I00), SocialTOX (PDC2022-133146-C21) funded by Plan Nacional I+D+i from the Spanish Government, and by the scholarship (FPI-PRE2022-105603) from the Ministry of Science, Innovation and Universities of the Spanish Government. Also, this work has been funded by the Ministerio para la Transformación Digital y de la Función Pública and Plan de Recuperación, Transformación y Resiliencia - Funded by EU – NextGenerationEU within the framework of the project Desarrollo Modelos ALIA.
During the preparation of this work, the authors used GPT-4 and Grammarly in order to: Grammar and spelling check. After using these tools, the authors reviewed and edited the content as needed and take full responsibility for the publication’s content. Prompt used to generate opinions (translated from Spanish): 1. Review the following product information: <JSON Description of the product>.