LLM-based web agents have the potential to automate long-running web tasks, such as searching for products in multiple e-shops and subsequently ordering the cheapest products that meet the user's needs. Benchmarks for evaluating web agents either require agents to perform tasks online using the live Web or offline using simulated environments, the latter allowing for the exact reproduction of the experimental setup. While DeepShop and ShoppingComp provide online benchmarks that require agents to perform challenging shopping tasks, existing offline benchmarks such as WebShop, WebArena, and Mind2Web cover only comparatively simple e-commerce tasks performed against a single shop containing product data from a single source. What is missing is an e-commerce benchmark that simulates multiple shops containing heterogeneous product data and requires agents to perform complex retrieval tasks. We fill this gap by introducing WebMall, the first offline multi-shop benchmark for evaluating web agents on challenging comparison shopping tasks. WebMall consists of four simulated shops populated with product data extracted from the Common Crawl. The WebMall tasks range from specific product searches and price comparisons to advanced searches for complementary or substitute products, as well as checkout processes. We validate WebMall using eight agents that differ in observation space, availability of short-term memory, and the employed LLM. The validation highlights the difficulty of the benchmark, with the best-performing agents achieving task completion rates below 65% in the task categories cheapest product search and vague product search.

LLM-based agents are increasingly used to automate web tasks such as product search, offer comparison, and order placement. Current research explores different interfaces through which these agents interact with websites, including traditional HTML browsing, retrieval-augmented generation (RAG) over pre-crawled content, communication via Web APIs using the Model Context Protocol (MCP), and natural-language querying through the NLWeb interface. Yet no systematic comparison of the effectiveness and efficiency of these interfaces on identical challenging task sets exists. To address this gap, we introduce a testbed consisting of four simulated e-shops, each offering its products via HTML, MCP, and NLWeb interfaces. For each interface (HTML, RAG, MCP, and NLWeb), we develop specialized agents that perform the same sets of tasks, ranging from simple product searches and price comparisons to complex queries for complementary or substitute products and checkout processes. We evaluate the agents using GPT-5 and GPT-5-mini. Our evaluation shows that RAG, MCP, and NLWeb agents outperform HTML browsing agents by 11 percentage points in task completion while requiring 2–5 times fewer tokens on search-oriented tasks. The GPT-5 RAG agent achieves the highest task completion rate (0.79) while maintaining moderate token consumption.

An increasing number of data providers have adopted shared numbering schemes such as GTIN, ISBN, DUNS, or ORCID numbers for identifying entities in the respective domain. This means for data integration that shared identifiers are often available for a subset of the entity descriptions to be integrated while such identifiers are not available for others. The challenge in these settings is to learn a matcher for entity descriptions without identifiers using the entity descriptions containing identifiers as training data. The task can be approached by learning a binary classifier which distinguishes pairs of entity descriptions for the same real-world entity from descriptions of different entities. The task can also be modeled as a multi-class classification problem by learning classifiers for identifying descriptions of individual entities. We present a dual-objective training method for BERT, called JointBERT, which combines binary matching and multi-class classification, forcing the model to predict the entity identifier for each entity description in a training pair in addition to the match/non-match decision. Our evaluation across five entity matching benchmark datasets shows that dual-objective training can increase the matching performance for seen products by 1% to 5% F1 compared to single-objective Transformer-based methods, given that enough training data is available for both objectives. In order to gain a deeper understanding of the strengths and weaknesses of the proposed method, we compare JointBERT to several other BERT-based matching methods as well as baseline systems along a set of specific matching challenges. This evaluation shows that JointBERT, given enough training data for both objectives, outperforms the other methods on tasks involving seen products, while it underperforms for unseen products. Using a combination of LIME explanations and domain-specific word classes, we analyze the matching decisions of the different deep learning models and conclude that BERT-based models are better at focusing on relevant word classes compared to RNN-based models.