When Positive Service Logistics Encounter Enhanced Purchase Intention: The Reverse Moderating Effect of Image–Text Similarity

Abstract

E-commerce platforms offering regional fresh produce often face a trade-off between logistics costs and product quality. Due to limited use of cold chain logistics, consumers frequently receive damaged goods, resulting in negative post-purchase experiences. This study examines how logistics service encounters, as reflected in consumer reviews, influence subsequent purchase behaviour, and how the alignment between review images and text moderates this relationship. We analyse sales and review data from 694 fruit products on Tmall between February and April 2024. Latent Dirichlet Allocation (LDA) is applied to extract logistics-related review content. At the same time, image–text similarity is assessed using the Chinese-CLIP model. Regression analysis reveals that positive logistics service encounters significantly enhance purchase intention. However, high image–text similarity weakens this positive effect, suggesting that overly repetitive content may reduce informational value for prospective buyers. These findings advance understanding of consumer behaviour in online fresh produce markets by highlighting the interactive effects of logistics experiences and user-generated content. The results offer practical implications for improving logistics services, enhancing content diversity in review systems, and increasing consumer trust in e-commerce environments.

1. Introduction

The online sale of fresh fruits has gained considerable traction in the e-commerce sector, driven by increasing consumer demand for convenience [1] and accessibility [2]. However, to minimise costs, most online retailers of fresh fruits rely on standard courier services, which are ill-suited for transporting perishable goods like fruits. This conventional logistics approach often leads to spoilage and damage during transport [3]. For instance, temperature-sensitive fruits may ripen or rot prematurely if proper temperature control measures are not in place. Additionally, inadequate packaging or careless handling can lead to physical damage, compromising the value of the goods and the consumer experience. Consequently, ordinary courier services jeopardise the fruit’s quality and undermine customer satisfaction and trust in online purchases [4]. As the online fresh produce market grows, retailers must address key aspects of consumer logistics service encounters, including delivery speed, shipment timeliness, packaging condition, doorstep delivery, information transparency, and claims assurance. Additionally, merchants should understand potential consumers’ perceptions to evaluate and enhance their logistics systems, ensuring safe and efficient fruit delivery that meets consumer expectations.

Fruits are highly susceptible to spoilage and damage during transportation, prompting consumers to rely increasingly on visual and textual information to inform their purchasing decisions [5]. According to Zinko et al., the absence of images and reliance solely on textual descriptions often leads to conservative purchasing attitudes [6], as text alone cannot accurately convey product details, diminishing trust and affecting purchasing intentions [7]. Conversely, providing only images without accompanying text makes it challenging for consumers to access specific product details, such as origin, variety, and quality, leaving doubts unaddressed [8]. A combination of visuals and text offers a comprehensive approach, visually showcasing the product while providing detailed descriptions, thus enabling consumers to make more informed decisions.

While the impact of logistics services on consumer purchasing behaviour has been extensively studied, most existing research focuses on the relationship between logistics service quality and consumer satisfaction, loyalty, or purchase intention. However, the integration of logistics service with service encounter theory remains underexplored, particularly in fresh products. Specifically, the influence of key contact points in logistics services on consumer decision-making has not been thoroughly examined. Furthermore, the role of graphic similarity in logistics service evaluation has received minimal attention. Existing studies predominantly address graphic consistency in advertising or product presentation, neglecting its potential relevance in logistics service contexts. This study addresses these gaps by investigating the combined effects of logistics service encounters and graphic similarity on consumer purchasing behaviour.

This study raises two key research questions to address this gap: (1) How do logistics service encounters mentioned in online reviews affect potential consumer purchasing behaviour? (2) Does the similarity between review text and images moderate this relationship? Accordingly, the primary objective of this study is to investigate the impact of logistics service encounters on consumer purchase behaviour within the context of online fresh produce reviews. A secondary aim is to examine how graphic similarity between textual and visual content in reviews moderates this impact. The study utilises review data from fruit products on the Tmall platform. The LDA method is employed to identify logistics service encounter themes and quantify their frequency by leveraging the suitability of Latent Dirichlet Allocation (LDA) for extracting themes from datasets. The Contrastive Language-Image Pretraining (CLIP) model is also applied for multimodal computation, normalising review text and images into vectors and calculating graphic similarity using cosine similarity. Finally, regression analysis using SPSS Statistics 27 modelling is conducted to reveal the influence of these factors on consumer decision-making. Ultimately, it was found that the more logistics service encounters, the more consumer purchasing behaviour; the lower the graphic similarity, the higher the consumer purchasing behaviour.

This paper holds significance both theoretically and practically. Theoretically, it first integrates service encounter theory to examine how different dimensions of logistics services systematically influence consumers’ purchase intentions. Second, it explores the impact of graphical inconsistency on consumer decision-making, grounded in information richness theory. Finally, by combining service encounter theory, information richness theory, and consumer behaviour theory, this paper introduces a novel perspective for understanding the combined influence of logistics services and graphical information, thereby expanding the theoretical framework for addressing complex shopping situations and information asymmetry. Practically, the study highlights the critical role of logistics service optimisation in driving sales on e-commerce platforms. It recommends that e-commerce enterprises focus on enhancing logistics processes by selecting reliable logistics partners and implementing intelligent logistics systems to improve customer experience. Additionally, high-quality graphic reviews are essential in shaping potential buyers’ decisions. E-commerce platforms can incentivise consumers to provide more in-depth and valuable reviews, thus supporting informed purchasing decisions.

The paper is structured as follows: Section 2 reviews the literature and develops hypotheses; Section 3 outlines the methodology; Section 4 presents the research findings; Section 5 discusses the implications; and Section 6 concludes. By examining the roles of image–text similarity and logistics service encounters in influencing consumer purchasing behaviour through online reviews, this study aims to reveal how e-commerce platforms can enhance consumer trust and purchase intent through better information transmission, offering new directions for practical application and theoretical exploration.

2. Literature Review and Hypotheses

2.1. Impact of Online Reviews

In the digital age, online reviews have become a primary channel for consumers to express opinions and share experiences, significantly influencing various aspects of the e-commerce sector [9]. Reviews not only drive sales but also provide essential insights for businesses to optimise products [10], improve service quality [11], and enhance market competitiveness [12]. Numerous studies have demonstrated that the quantity and length of reviews can positively impact consumer purchasing decisions [13]. Additionally, reviews with strong emotional content, particularly positive sentiment, significantly increase purchase intentions, boosting sales [14]. However, the prevalence of fake reviews and “water army” tactics has led consumers to question the credibility of reviews. Research indicates that credible reviews are more likely to influence purchasing decisions [15]. Moreover, reviews that are logically structured and relevant to the product provide consumers with more valuable information [16]. Thus, online reviews play an increasingly crucial role in consumer decision-making and offer valuable opportunities for businesses to improve, making them a critical factor in e-commerce.

Online reviews are a primary source of product information for consumers. From the perspective of information richness theory, effective information delivery depends on detail, complexity, and variety, with different formats (e.g., text and images) providing distinct types of cognitive support [17]. Text-based reviews convey detailed qualitative and quantitative information, such as product functions, features, and usage experiences [18]. In contrast, images help consumers quickly grasp a product’s appearance, quality, and usage effects through visual representation [19]. Thus, relying solely on textual reviews does not fully capture consumers’ overall perceptions, as it lacks the intuitive appeal and emotional resonance that images provide [20]. Empirical studies confirm that reviews with images are more effective than text-only reviews. For instance, Vazquez [21] found that reviews featuring pictures are more likely to drive purchasing behaviour, with high-quality images drawing more attention and enhancing product appeal [22].

Additionally, combining text and images significantly enhances information richness, helping consumers better understand product features and usage effects [23] and increasing their trust and satisfaction [24]. Existing research has extensively examined the effects of textual reviews, visual reviews, and their combination through the lens of information richness theory. Most research shows that integrating text and images enhances information transmission and influences consumer decision-making. However, there is a lack of comprehensive analysis regarding the moderating role of image–text similarity in sales. To be specific, limited attention has been paid to how image–text similarity affects consumers’ perceptions of logistics services and, in turn, influences sales performance, particularly in online transactions involving perishable goods. As logistics play a critical role in shaping the online shopping experience, directly impacting the delivery condition of perishable items and consumer satisfaction, the interactive effects of image–text information in this context warrant further investigation. Addressing these gaps, this study focuses on the logistics service context to examine the moderating effect of image–text similarity. By systematically pairing and analysing images and accompanying text in consumer reviews, we aim to assess their combined influence on sales performance and contribute to closing this gap in the existing literature.

2.2. Logistics Service Encounter and Consumer Decision

The service encounter theory, introduced by Shostack [25], has played a foundational role in understanding service quality. Initially framed as a binary interaction between service providers and customers [26], the concept has since evolved into a more complex triadic interaction [27], reflecting the multifaceted nature of modern service systems. This evolution has fostered significant advancements in theoretical and practical applications, driving innovations in the service industry and enhancing customer experiences [28]. As the theory has matured, it has highlighted the importance of intangible service components, particularly the touchpoints that shape the overall customer experience [29].

Service encounter theory has been further enriched in the digital era by integrating big data and artificial intelligence [30]. These technologies have redefined service interaction frequency, nature, and patterns, allowing the theory to adapt to modern technological developments and introducing new service paradigms [31]. This transformation emphasises the complexity and continuity of service encounters, as various service interfaces form a chain of interactions contributing to the overall service experience [32].

In e-commerce, logistics services are essential touchpoints within the service encounter chain, acting as the backbone of logistics systems [33]. They are critical in optimising service quality, shaping customer perceptions, and influencing subsequent behavioural intentions [34]. Logistics service quality, a key metric for assessing both operational efficiency and customer satisfaction in e-commerce, is evaluated across nine dimensions: personnel contact quality, order fulfilment, information quality, ordering procedures, order accuracy, order status, order quality, handling of discrepancies, and timeliness [35]. These dimensions provide a comprehensive framework for assessing logistics services, contributing to enhanced customer experiences [36] and increased business competitiveness [37].

As e-commerce platforms expand, the scope of logistics service encounters continues to grow; its importance in driving consumer purchasing decisions has been well-documented in empirical studies [38]. From a consumer’s perspective, the price and quality of products are not the only factors influencing purchase decisions [39]. Increasingly, consumers place significant value on the speed and reliability of logistics and after-sales service in their shopping experience [40]. Efficient and timely delivery enhances consumer trust, often leading to repeat purchases and positive recommendations [41]. Conversely, delays or damaged goods, even when the product is of high quality, can lead to disappointment and drive consumers to switch to competitors offering better logistics service encounters [42]. Thus, a well-executed logistics service encounter fosters customer loyalty and boosts repurchase rates [43].

Furthermore, logistics service encounters also influence consumers’ perceptions of convenience and cost [44]. Modern consumers increasingly seek instant gratification [45]. They are particularly drawn to fast delivery options that reduce waiting times and improve shopping efficiency [46].

Wu and Dong [47] studied logistics service encounters in the pharmaceutical e-commerce sector. They demonstrated their direct impact on customer satisfaction. Their findings offer valuable insights into improving logistics service quality and managing customer relationships, particularly within this industry. Despite the ongoing development of service encounter theory and growing recognition of the importance of logistics service encounters in e-commerce, significant research gaps remain. First, extant research has not adequately explored logistics service encounters, particularly in the agricultural products sector. This field has unique characteristics, such as perishability, high logistics timeliness, and preservation requirements. Relevant studies are incredibly scarce, failing to adequately reveal the unique influence mechanisms of logistics service encounters in agricultural products. Secondly, extant research exhibits a paucity of practical methodological applications and systematic discussions on precisely analysing the specific impact pathways of logistics service encounters on sales through user-generated content, such as online reviews. To address this gap, this paper employs Latent Dirichlet Allocation (LDA) modelling to analyse the impact of logistics service encounters on sales, as reflected in online reviews. Based on this, the following hypothesis is proposed:

H1. 

Logistics service encounters positively influence consumer purchasing behaviour.

2.3. Image–Text Similarity in Comments Reduces Consumer Uncertainty

Information richness theory plays a critical role in information dissemination. It posits that different media vary in information richness, with image–text similarity being a key factor influencing this richness. As a fundamental concept in information communication, image–text similarity ensures complete transmission and effective reception of information, enhancing user experience and promoting consumption. This concept is particularly significant in communication and marketing in the digital era, as it supports diverse information presentation and deepens users’ understanding and perception. This concept has gained particular relevance in communication and marketing in the digital age, where it enhances the multidimensional presentation of information, deepening users’ understanding and perception.

Image and text similarity improve the effectiveness of information delivery. It enhances users’ cognitive experience by providing a clearer understanding of the content [48]. When images and text are combined, consumers can more intuitively grasp the characteristics of products or services, making information communication more comprehensive and impactful [49]. Cognitive psychology suggests that people use multiple senses when processing information, and the integration of visuals with text satisfies this need for diverse sensory input, enhancing comprehension and retention [50]. As a result, comments with pictures allow users to develop a deeper understanding of the information, leading to a more accurate perception of the product or service.

When users face incomplete or unclear information during purchase decisions, image–text similarity helps bridge information gaps. It reduces uncertainty about products or services [51]. By presenting intuitive images alongside clear text, consumers gain a more comprehensive understanding of product features and benefits, which increases decision-making confidence and lowers perceived risk [52]. Reviews combining images and text often include other users’ experiences and feedback, serving as valuable social references that guide consumers toward more confident choices [53]. For example, seeing photos of products in use and positive reviews from other buyers provides reassurance. It reduces hesitation in purchasing [54].

Li et al. [55] and other scholars have found that reviews combining images and text are more useful than text-only reviews, particularly regarding emotional resonance and hedonic value. Zhang and Choi’s [56] research highlights the synergy between images and text in reviews, showing that their combined effect significantly improves the communication and understanding of information. Chen’s [57] study further explores how including images adds depth and breadth to reviews, offering consumers a more comprehensive perspective that aids in decision-making. Similarly, Li et al.’s [58] research demonstrates that adding images significantly enriches the information in reviews, boosting consumers’ willingness to purchase. A comprehensive review of the existing literature highlights the positive impact of integrating visual media with written text in enhancing information dissemination. This multimodal approach has been shown to reduce consumer uncertainty and shape purchasing intentions, emphasising the importance of their synergistic effects. Nonetheless, notable research gaps remain. First, few studies have specifically examined image–text similarity in consumer reviews, and standardised methodologies for measuring this variable are lacking. Second, within the context of logistics service experiences, the moderating role of image–text similarity on the relationship between logistics encounters and consumer purchasing decisions has not been adequately investigated, leaving the underlying mechanisms poorly understood. To address these gaps, this study proposes to quantify image–text similarity in product reviews and examine its moderating effect on the influence of logistics service encounters on purchase decisions, thereby contributing to the advancement of research in this area. Based on this foundation, the following hypothesis is proposed:

H2. 

Image–text similarity moderates the impact of logistics service encounters on consumer purchase decisions.

Considering that many factors influence consumer decision-making, the most studied are Review Timestamp, Sentiment Score, and Semantic Ambiguity, which are used as control variables in this paper. The variable model diagram illustrating this relationship is presented below (Figure 1).

3. Methodology

3.1. Data Description

Selecting an online shopping platform with many sellers and buyers is essential to investigate the influence of logistic service encounters and image–text similarity on consumer purchase decisions. These allow for a more comprehensive consumer behaviour analysis across different product categories with varying characteristics. Fruit products, in particular, present unique challenges due to their perishable nature. Maintaining freshness, dealing with seasonal availability, managing quality fluctuations, and meeting consumer demands for healthy, convenient consumption are critical factors influencing purchasing decisions. Given these complexities, Tmall serves as an ideal platform for this study. Its well-established e-commerce ecosystem has attracted many high-quality sellers and active buyers. Tmall’s extensive and diverse selection of fruit products across various categories offers a robust dataset for analysis. Additionally, the platform’s reliable comment management system helps ensure the authenticity and longevity of customer reviews. This minimises the risk of review deletion, allowing for more accurate and trustworthy data collection, which is crucial for the integrity of this research.

This study collected sales data from 31 fruit categories, including 694 items, covering sales volume, product review text, and associated images from February to April 2024. The data were then cleaned and organised. First, out-of-range data and missing image information were removed. Next, each review’s text, images, and item names were aligned. The text data were cleaned by removing punctuation marks, stop words, numbers, special characters, and non-textual elements, ensuring consistency in measurement units across variables. Using lexical processing techniques, the text was decomposed into smaller semantic units, individual words or phrases, each representing a specific meaning from the original text. After this cleaning and processing, 10,956 valid data points were retained, each consisting of a review and the corresponding images.

3.2. Logistics Service Encounters Are Extracted Through LDA

This study employs Latent Dirichlet Allocation (LDA) to analyse consumers’ logistics service encounters in customer review texts. LDA, a generative model, efficiently identifies underlying themes in large-scale text data, allowing for a clearer understanding of the distribution of multiple topics within reviews. Unlike methods such as Term Frequency–Inverse Document Frequency (TF-IDF), which identifies keywords but lacks thematic depth [59], or Non-Negative Matrix Factorisation (NMF), which decomposes text matrices but is less interpretable [60], LDA captures both word frequency and co-occurrence to deliver a more accurate thematic analysis. While the Variable Auto-Encoder Variational Autoencoder (VAE) offers complex generative modelling, it requires more training data and remains less interpretable [61]. LDA’s balance of accuracy and interpretability makes it ideal for this study.

To improve LDA’s recognition of logistics-related themes, key terms like “logistics speed”, “service attitude”, “door-to-door service”, and “individual packaging” were incorporated into a specialised lexicon. This approach refines the model’s precision, ensuring that LDA can better calculate the probability distribution of logistics topics within customer reviews, thus providing more accurate insights into consumer perceptions of logistics services.

As a core algorithm in text mining, LDA primarily simplifies large-scale text data into a series of distinguishable dominant topics, summarising core themes through a combination of representative terms [62]. The Latent Dirichlet Allocation (LDA) model is optimal for extracting text themes for large volumes of review text. The LDA model comprises three layers: the document layer (D), the word layer (W), and the topic layer (K) [63], aiming to uncover hidden thematic structures within a document collection. This study defines the dictionary size in the LDA framework as L, where each L-dimensional vector (1, 0, 0, …, 0, 0) represents a specific term. Suppose a review of N words is denoted as d = (w₁, w₂, …, w_N). Further, if a product’s review collection D comprises M reviews, it is represented as D = (d₁, d₂, …, d_M). In these M reviews, assume there are K topics, labelled as Z₁, Z₂, …, Z_K.

In the LDA model, $\alpha$ and $\beta$ are the hyperparameters of the Dirichlet function. $\theta$ represents the parameters of the multinomial distribution of topics within documents, following a Dirichlet prior distribution with hyperparameter $\alpha$, while $\mathrm{\varnothing }$ denotes the parameters of the multinomial distribution of terms within topics, following a Dirichlet prior distribution with hyperparameter $\beta$. The LDA model diagram is shown in Appendix A. The implementation steps are as follows:

1.

For document d, the LDA model first draws a topic distribution ${\theta }_d$ from a Dirichlet distribution, controlled by the parameter α.

2.

Based on the topic distribution ${\theta }_d$ for document d, a topic $z_{d,n}$ is sampled from the corresponding topic multinomial distribution for document d.

3.

A word multinomial distribution ${\mathrm{\varnothing }}_{z_{d,n}}$ associated with the topic $z_{d,n}$ is then drawn from a Dirichlet distribution, controlled by the parameter $\beta$.

4.

A word $w_{d,n}$ is selected from the multinomial distribution ${\mathrm{\varnothing }}_{z_{d,n}}$ corresponding to the chosen topic $z_{d,n}$.

3.3. Use the CLIP Model to Obtain Image and Text Similarity

In the digital age, graphical similarity is crucial in linking textual descriptions with visual content, particularly in e-commerce, which directly influences consumers’ purchasing decisions. In 2021, OpenAI introduced the CLIP model, which, through its innovative comparative learning framework, efficiently aligns image and textual features in a unified vector space, paving the way for advanced crossmodal information processing. CLIP captures subtle correlations between images and text. It demonstrates strong generalisation abilities for previously unseen categories, significantly enhancing the accuracy and flexibility of multimodal data processing [64]. The model excels in graphical similarity computation, with results that closely match manual assessments. It is highly suitable for accurately measuring graphical similarity in reviews.

The CLIP model is known for its high accuracy for assessing image–text similarity, particularly in English contexts. However, the Chinese-CLIP model is more effective when dealing with complex textual and visual data, especially in Chinese comments. Therefore, this study uses the Chinese-CLIP model to evaluate image–text similarity [65] quantitatively.

The processing flow of the CLIP model consists of four key steps. First, data preprocessing is performed, where images are resized (e.g., 224 × 224 pixels), converted to RGB colour space, and normalised. Text is converted into a sequence of numeric IDs using a disambiguation technique, then padded or truncated to a fixed length. Next, feature extraction occurs, with images transformed into high-dimensional feature vectors by an image encoder (e.g., Vision Transformer or ResNet). Simultaneously, text is converted into representations of the same dimensionality as the image features using a Transformer-based text encoder for crossmodal alignment. Subsequently, image and text feature vectors are normalised to unit vectors, ensuring that similarity calculations focus on angular differences rather than vector length. Finally, similarity is measured by calculating the inner product (cosine similarity) between the normalised image and text feature vectors, where smaller angles indicate higher similarity.

Figure 2 illustrates the text–image encoding and matching framework. Here, $T_n$ denotes the text features generated by the Text Encoder; each $T_n$ corresponds to a semantic segment of the input text (e.g., “The apples are fresh” is decomposed into multiple text feature vectors). $I_n$ represents the image features output by the Image Encoder, capturing distinct visual patterns of the input image (e.g., colour, shape of the apple). These features are then matched across modalities (e.g., $I_1$ $T_1$ denotes the similarity between the first image and the first text features).

We conducted descriptive statistics on sales volume, topic frequency, and similarity scores to ensure a comprehensive and in-depth analysis. The relevant findings are presented in

Table 1. Descriptive statistics of sales, topic frequency, and similarity scores.

	Number	Min (M)	Max (X)	Average (E)	Standard Deviation
Sales amount	10,956	3.00	10,778.00	1236.23	1529.57
topic1	10,956	0.00	0.93	0.15	0.19
topic2	10,956	0.00	0.96	0.20	0.23
topic3	10,956	0.00	0.95	0.13	0.17
topic4	10,956	0.00	0.96	0.14	0.18
topic5	10,956	0.00	0.97	0.14	0.18
topic6	10,956	0.00	0.94	0.25	0.25
Similarity score	10,956	0.00	0.58	0.45	0.04
effective N (sample size)	10,956

. This table outlines the range and average levels of each variable. The sales volume statistics highlight the extremes and overall performance of sales. The data for topics (topic 1 to topic 6) illustrate their characteristics and the extent of changes within each area, aiding in identifying trending topics and evaluation patterns. Additionally, the similarity scores measure the degree of association between the compared items, clarifying the overall level of agreement and individual consensus.

3.4. Model of Product Sales

This paper uses a linear regression model to examine the relationship between logistics service encounters and merchandise sales. This model was selected because it allows us to determine whether a linear relationship exists between the two variables and to quantify their strength. Specifically, it helps assess how delays and damages in logistics services affect final sales performance, providing insights for improving service and boosting sales. Linear regression is a widely used method due to its simplicity, intuitiveness, and ease of interpretation, making it ideal for analysing causal relationships. This paper analyses individual merchandise categories, with sales volume of commodity i in month t as the dependent variable. The independent variable, logistics service encounter, is derived from user reviews using the LDA model. The moderator variable, graphical similarity, is also based on user reviews of commodity i in month t, with calculations performed using the CLIP model. The linear regression model employed in this study is expressed as follows:

$${Sales}_{i,t}={\alpha }_1{LogisticsService}_i+{\mu }_{1i}$$

(1)

where

${Sales}_{i,t}$: Sales (total number of transactions) in month t of commodity i to which the comment belongs;

${LogisticsService}_i$: Probability of each logistic theme for the commodity i to which the comment belongs;

${\mu }_{1i}$: Normally distributed error terms for the dependent variable.

Research indicates that image–text similarity has a certain impact on product sales volume [66]. In this paper, the CLIP model is utilised to compute the similarity of the graphics in the reviews. Eventually, each review under each product corresponds to a similarity value. So, after adding the moderating variables, the following is obtained:

$${Sales}_{i,t}={\alpha }_1{LogisticsService}_i+{\alpha }_2{GraphicSimilarity}_i+{\mu }_{2i}$$

(2)

where

${GraphicSimilarity}_i$: Image text similarity for each product;

${\mu }_{2i}$: The normally distributed error term of the dependent variable.

4. Results

4.1. Results of LDA

Various methods can be used to implement LDA models. For the dataset in this paper, the gensim library is the most suitable platform. Gensim is specifically designed for natural language processing tasks and supports topic modelling, enabling the automatic extraction of topics from large text corpora. It also utilises streaming processing, allowing for computations without loading the entire dataset into memory. This feature makes gensim more efficient for handling large-scale document sets. Additionally, its optimised algorithms and parameter tuning options enhance the speed of model training and inference. Compared to the sklearn library, gensim is better suited for rapid processing and real-time analysis of complex text tasks.

LDA models are commonly used for text clustering, but the optimal number of topics varies for each dataset. This study employed a combined strategy using perplexity and coherence to determine this parameter. Perplexity is an important metric for evaluating the effectiveness of LDA models; it generally decreases as the number of topics increases, indicating better classification accuracy [67]. However, an excessive number of topics may lead to ambiguous topic meanings and reduce interpretability. In contrast, the coherence index directly reflects classification quality, with higher values indicating stronger accuracy and reliability. Thus, this study seeks the optimal balance between perplexity and coherence to achieve effective classification and interpretability.

The trend graphs of perplexity and coherence, shown in Figure 3 and Figure 4, indicate that k = 6 is the optimal number of topics for this dataset. This configuration balances classification efficacy and topic interpretability, forming the basis for a detailed analysis of logistics-related topics.

Because the logistics aspect was targeted in the data processing stage, most results were related to logistics.

Table 2. Top 10 factors for topics.

Number	Topic	Top Words
1	Logistics Speed	Excellent; Fresh; Exceptionally Fresh; Outstanding; Highly Recommended; Valuable; Absence of Damaged Fruit; Efficient Logistics; Highly Satisfied; Logistics are Efficient
2	Praise and Recommendation	Satisfactory; Very Fresh; Quite Good; Very Good; Fresh; Positive Feedback; Worthwhile; Recommended; Individually Packaged; Highly Satisfied
3	Compensation for Logistics Speed	Positive Feedback; Five-Star Rating; Not Fresh; No Damage; Fast Delivery; Particularly Good; Damaged Fruit; Especially Liked; SF Express; Overall
4	Quality Defects	Unsatisfactory; Negative Feedback; Fresh; Positive Feedback; Compensation; Cost-Effective; Not Fresh; Door-to-Door Delivery; Receipt of Goods; No Damage
5	Packaging Integrity	Extremely Poor; Negative Feedback; Well-Packaged; Repeat Purchase; Fairly Fresh; Generally; Super Fresh; Extremely Unsatisfactory; Particularly Fast; Partial
6	Cost-Effectiveness	Exceptionally Delicious; Very High Quality; Recommended; Very Poor; Effective; Excellent Value for Money; Worth the Price; Excessively Sweet; Easy; Refreshing

presents the top ten probability factors in these six topics, and the topic names are profiled according to the top ten significant feature words in each topic. For example, the words “Fresh”, “Highly Satisfied”, and “Efficient Logistics” appear in topic1, indicating that topic1 concerns fast logistics where consumers are satisfied when fresh fruit is delivered; therefore, topic1 is named “logistics speed”, and these six topics are named according to this logic, respectively, which are designated as follows: logistics speed, praise and recommendation, compensation for logistics speed, quality defects, packaging integrity, and cost-effectiveness.

This study investigates key concepts influencing consumer purchase decisions in logistics service encounters.

Logistics speed refers to the time taken for product delivery, emphasising its role in consumer satisfaction. Faster delivery aligns with consumer expectations and often leads to higher satisfaction levels. Praise and recommendations capture consumer evaluations and endorsements of products. Positive reviews significantly enhance a product’s attractiveness and encourage purchases. Compensation for logistics issues includes safeguards against lost or damaged items during shipping. This factor highlights consumers’ preference for risk management, as they are more inclined to buy from companies that offer these assurances. Quality defects focus on whether products meet expected standards. Consumers are likely to avoid items perceived as defective, which can deter their purchase decisions. Packaging integrity pertains to the condition of product packaging upon delivery. Intact packaging affects first impressions and overall consumer experience. Finally, cost-effectiveness reflects the balance between price and quality. Consumers seek assurance that the quality justifies the price paid.

Figure 5 shows the average proportion of each topic in the comments, through examining these themes, this research aims to enhance understanding of the factors that shape consumer concerns and decision-making processes in purchasing.

4.2. Regression Model Results

This paper will construct a linear regression model to regress the probability of logistics service encounters with the theme and the corresponding sales. The results are shown in Model 1 and Model 2 in

Table 3. Results of the linear regression model.

Variables	Model 1	Model 2	Model 3	Model 4
Sentiment Score	−0.069 ***	−0.067 ***	−0.065 ***	−0.065 ***
Time	0.169 ***	0.169 ***	0.170 ***	0.170 ***
Semantic Ambiguity	0.010	0.007	0.006	0.005
Logistics Speed		1.472 *	1.368 *	1.645 **
Praise and Recommendation		1.875 *	1.745 *	1.948 **
Compensation for Logistics Speed		1.366 *	1.271 *	1.488 **
Quality Defects		1.494 *	1.392 *	1.557 **
Packaging Integrity		1.450 *	1.350 *	1.671 **
Cost-Effectiveness		2.020 *	1.880 *	1.891 *
Similarity Score			−0.042 ***	0.042
Logistics Speed × Similarity Score				−0.270 *
Praise and Recommendation × Similarity Score				−0.195
Compensation for Logistics Speed × Similarity Score				−0.211
Quality Defects × Similarity Score				−0.158
Packaging Integrity × Similarity Score				−0.315 *

*** p < 0.001.; ** p < 0.01; * p < 0.05.

below.

According to Model 1 in Table 3, semantic ambiguity among the three control variables does not significantly impact the dependent variable. However, the time and sentiment score significantly affect the dependent variable, with p-values below 0.01. In Model 2, all six extracted themes positively and significantly influence consumer purchasing behaviour (p < 0.05). Specifically, when the Compensation for logistics speed increases by 1 unit, sales increase by 1.366 units; for every 1 unit increase in packaging integrity, sales increase by 1.45 units; when the logistics speed increases by 1 unit, the sales increase by 1.472 units; when quality defects increase by 1 unit, sales increase by 1.494 units; for every 1 unit increase in praise and recommendation, sales increase by 1.875 units; for every 1 unit increase in cost-effectiveness, sales increase by 2.02 units. The supports Hypothesis H1: Higher frequency of logistics service encounters correlates with greater certainty in consumer purchasing behaviour.

Increased frequency of logistics service encounters enhances consumers’ sense of security and trust. Consumers tend to feel more confident in brands that offer timely and reliable services, which affects their purchasing decisions and fosters brand loyalty. Over time, reliance on frequent logistics interactions can deepen, influencing their willingness to purchase additional products.

Furthermore, the notable effects of review timing and sentiment scores indicate that real-time information and emotional resonance are essential in today’s shopping landscape. Consumers often turn to others’ opinions and feedback when faced with numerous options. Timely reviews reflect a product’s current status and evoke emotional connections, boosting consumer confidence in purchasing decisions. Consequently, merchants should prioritise developing a robust review system to ensure transparency and the timely dissemination of information.

The moderating effect in this paper was tested by including image–text similarity scores in the regression analysis of the moderating variables in the discussion of logistics service encounters, topic frequency, and sales. The results are displayed in Model 3 and Model 4 of Table 3. The findings indicate that the similarity between images and text significantly affects only logistics speed and packaging integrity. Specifically, when the image–text similarity increases by 0.1, the positive impact of packaging integrity on sales decreases by 0.0315 units. When the similarity between images and texts increases by 0.1, the positive impact of logistics speed on sales decreases by 0.027 units. These suggest that these two are the only ones influenced by image–text similarity among the explored themes. The insignificance of praise and recommendation may be attributed to the high homogeneity of positive feedback within the sample or the presence of irrelevant or artificial comments, which could dilute their correlation with sales performance. The non-significant impact of logistics speed compensation might result from the measurement’s failure to capture nuanced differences, such as the amount and method of compensation, thereby limiting its ability to reflect actual effects. The lack of significance in quality defects could be due to the absence of a severity-based differentiation of defects, coupled with consumers’ higher tolerance for minor issues, given the perishable nature of fruit products. Lastly, the insignificant relationship between cost-effectiveness and sales may stem from the fact that fruit is typically considered a low-involvement product, where consumers place less emphasis on rational cost–benefit evaluations, or where individual variations in price sensitivity obscure broader trends.

Furthermore, the effect of image–text similarity is negative and significant; lower similarity leads to more supplementary information for consumers. Low image–text similarity can lead to consumer uncertainty regarding their perception of a product, prompting them to seek additional information to fill this gap. In such situations, enhanced logistics speed allows consumers to receive their products more quickly, alleviating the anxiety associated with waiting times and improving the overall purchasing experience. Additionally, strong packaging integrity safeguards the product during transit and reinforces the consumer’s connection to the brand.

4.3. Robustness Analysis

To verify the stability of the model, this paper employs robustness analysis, utilising various methods to establish reliability. In this paper, we employ a method of substituting the dependent variable for robustness analysis. We keep the independent variables constant by transforming sales to log(0.00001 × sales + 1) to establish a regression model. The results, presented in

Table 4. Robustness test results.

Variables	Model 1	Model 2	Model 3	Model 4
Semantic Ambiguity	0.010	0.008	0.006	0.006
Time	0.171 ***	0.171 ***	0.172 ***	0.172 ***
Sentiment Score	−0.069 ***	−0.067 ***	−0.065 ***	−0.065 ***
Logistics Speed		1.481 *	1.377 *	1.653 **
Praise and Recommendation		1.887 *	1.757 *	1.958 **
Compensation for Logistics Speed		1.375 *	1.279 *	1.498 **
Quality Defects		1.504 *	1.401 *	1.565 **
Packaging Integrity		1.460 *	1.359 *	1.676 **
Cost-Effectiveness		2.033 *	1.892 *	1.904 *
Similarity Score			−0.042 ***	0.041
Logistics Speed × Similarity Score				−0.269 *
Praise and Recommendation × Similarity Score				−0.192
Compensation for Logistics Speed × Similarity Score				−0.213
Quality Defects × Similarity Score				−0.157
Packaging Integrity × Similarity Score				−0.312 *

*** p < 0.001.; ** p < 0.01; * p < 0.05.

, indicate that among the control variables, comment time and sentiment score remain highly significant (p < 0.01). Before introducing the moderating variables, all six themes significantly affected sales (p < 0.05). After incorporating the moderating variables, however, only logistics speed and packaging integrity retain their significance (p < 0.05). These findings align with the earlier regression results, suggesting that the model exhibits robustness. These findings have significant implications for e-commerce platforms and merchants, suggesting that optimising logistics services can enhance consumer satisfaction and drive sales growth, creating a beneficial cycle.

5. Discussion

Existing research indicates that various factors influence consumer purchasing behaviour, including logistics service encounters and image–text similarity. High-quality logistics service encounters can significantly enhance consumers’ intention to purchase. At the same time, reviews with substantial image–text similarity tend to foster consumer trust. Through a detailed analysis of review data for fruit products on the Tmall platform, this study finds a positive correlation between the frequency of logistics service encounters in reviews and consumers’ purchasing behaviour. Furthermore, greater inconsistency in image–text similarity provides additional information, promoting more effective purchase decisions. These findings reveal the complexity of consumer purchasing behaviour and underscore the critical roles logistics service encounters and image–text similarity play in decision-making. This research carries important theoretical and practical implications.

5.1. Theoretical Contributions

This paper applies service encounter theory to fruit-based e-commerce, offering insights into the impact of logistics service encounters on online sales of fruit-related products. A logistics service encounter encompasses consumers’ interactive experiences with logistics services during the purchasing process, including key aspects such as service quality, information transparency, and delivery efficiency [68]. Research indicates that improving logistics service encounters can significantly boost sales, providing fresh perspectives on consumer behaviour in fruit-based e-commerce. Unlike prior studies that focused on isolated dimensions, such as service quality or customer satisfaction [69], this research highlights the integrative role of logistics service encounters in shaping consumer purchase decisions. Its theoretical contribution lies in its comprehensive analysis of logistics service dimensions, revealing the systematic influence of logistics service encounters on purchase intentions. This expands the application of service encounter theory to e-commerce settings and offers novel insights and directions for future theoretical research.

Given that the freshness of fruit products is often conveyed through images [70], this paper examines the role of image–text similarity in consumer decision-making through the lens of information richness theory. Information richness theory emphasises the quantity and quality of information, encompassing details, contextual integrity, and multidimensional features [71]. Images and text offer distinct information sources in reviews, enhancing overall information richness [72]. The study found that inconsistencies in image–text similarity increase the information available to consumers, boosting purchase intention. This finding offers a novel perspective on consumer behaviour under conditions of information asymmetry, distinguishing it from traditional research focusing on consistent information presentation. This study’s theoretical contribution is to demonstrate how image–text inconsistency influences the consumer decision-making process. It also promotes the further application of information richness theory in e-commerce environments, thereby expanding the theory’s explanatory power in understanding consumer behaviour.

Drawing from consumer psychology and information processing motivation theory, we identify clear boundary conditions: the reverse effect of image–text similarity emerges primarily in reviews concerning logistics speed and packaging integrity, domains characterised by subjective judgement and service variability, heightening perceived uncertainty. In such contexts, consumers are more likely to treat inconsistencies as diagnostic cues and use compensatory information processing to reduce ambiguity. Conversely, image–text inconsistencies do not significantly influence purchase intent in themes such as product quality, cost-effectiveness, or general praise. These themes are often emotionally one-dimensional or rely on factual descriptions that demand minimal cognitive effort, thereby failing to activate deeper processing. Our findings suggest that consumer engagement is conditional; only when motivation or perceived risk is sufficiently high will inconsistent cues prompt more profound cognitive elaboration. By embedding these insights within the framework of information processing motivation, this study refines the theoretical understanding of when and how information inconsistency affects consumer decision-making.

This paper examines the combined impact of logistics services and graphic information, exploring how these factors shape the consumer shopping experience. By integrating service encounter theory, information richness theory, and consumer behaviour theory, this study offers a novel perspective that deepens our understanding of consumer behaviour in online shopping environments. Specifically, it highlights how logistics service encounters and graphical information inconsistency influence consumer decision-making and purchase intentions. This theoretical integration extends the application of service encounter and information richness theories in e-commerce. It provides new insights into consumer behaviour theory, particularly in addressing complex shopping situations and information asymmetry. Furthermore, the research opens avenues for interdisciplinary exploration. It provides a solid theoretical foundation for future academic inquiry in this field.

5.2. Practical Implications

First, logistics services are crucial in e-commerce by connecting sellers and buyers. The research shows that consumers’ interactions with logistics services significantly impact their purchasing decisions. Consumers may discontinue their purchasing behaviour when more negative information about logistics service encounters occurs in the reviews. Conversely, if more positive information occurs, it may both reduce the risk of consumers’ purchasing and promote purchasing. To address this, e-commerce companies should focus on optimising their logistics processes. This can be achieved by partnering with reliable logistics providers, adopting intelligent systems to shorten delivery times, and offering real-time tracking services. These improvements can reduce negative experiences with logistics and increase purchase rates among potential customers.

Second, high-quality user reviews strongly influence potential buyers’ decisions, particularly those featuring detailed text and images. Such reviews provide specific product information and showcase its effectiveness, making them valuable references for other consumers. E-commerce platforms can encourage the generation of these reviews by offering incentives, such as reward points for high-quality content or prioritising the display of detailed reviews. This approach can motivate consumers to share more in-depth and comprehensive feedback, enriching the shopping experience for others.

Finally, the findings of this study offer valuable insights into competitive strategies within the industry. In today’s market, high-quality logistics services have become a key differentiator. Companies that excel in providing reliable logistics can gain a competitive edge. Additionally, the moderating effect of image–text similarity indicates that businesses should innovate in marketing and explore new ways to communicate information that aligns with evolving consumer preferences. By prioritising these factors, companies can drive sales, elevate industry-wide service standards, and foster innovation, contributing to long-term sustainable development.

5.3. Limitations and Future Research

Although this study offers a novel perspective on understanding consumer buying behaviour, several limitations remain. It is important to note that, given the perishable nature of fresh produce, its stringent requirements for logistics efficiency and preservation, and consumers’ strong preference for directly assessing its appearance and quality, the impact of image–text similarity in logistics service encounters and reviews may be particularly pronounced in this sector. As such, the findings of this study are most applicable to the fresh produce context, and caution should be exercised when generalising these conclusions to other product categories. Additional limitations include the following: First, the short time horizon of the data does not account for the long-term effects of temporal factors, such as seasonality or promotional activities, on consumer decision-making. Future research could explore the influence of these temporal variables on purchasing behaviour.

Additionally, the study does not fully address the impact of individual consumer differences (e.g., age, gender, educational background) on perceptions of logistics services and image–text similarity. Future research could examine how these individual differences influence the decision-making process. Finally, this study centres on the effects of user-generated content, and future research could compare the influence of merchant-generated versus user-generated content on consumer purchase decisions, particularly in terms of trust and interactivity. Future research could expand on these areas, providing a more comprehensive theoretical framework for understanding the multidimensional nature of consumer behaviour.

6. Conclusions

We believe the findings of this research will foster further interest and inquiry into strategy development for the logistics industry and e-commerce platforms in optimising services and enhancing user experience. The study reveals that logistics service encounters and image–text similarity significantly influence consumer purchasing behaviour. Specifically, when consumers have more logistics service encounters mentioned in the reviews, they obtain more logistics-related information, allowing them to have a more comprehensive grasp of the logistics status of the goods. At the same time, the study also indicates that consumers can acquire more product details quickly when there is a considerable disparity between the picture and text information in the reviews. This wealth of information promotes a deeper understanding of the product and consequently boosts the purchase intention. This discovery emphasises the significance of logistics service encounters and image–text similarity in online purchasing. It offers multidimensional information about the product and expedites consumers’ information acquisition and decision-making, positively impacting purchasing behaviour.