Mining consumer mindset metrics with user-generated content

Kübler, Raoul; Adler, Susanne J.; Welke, Lina; Sarstedt, Marko; Pauwels, Koen Article Mining consumer mindset metrics with user-generated content Schmalenbach Journal of Business Research (SBUR) Provided in Cooperation with: Schmalenbach-Gesellschaft für Betriebswirtschaft e.V. Suggested Citation: Kübler, Raoul; Adler, Susanne J.; Welke, Lina; Sarstedt, Marko; Pauwels, Koen (2025) : Mining consumer mindset metrics with user-generated content, Schmalenbach Journal of Business Research (SBUR), ISSN 2366-6153, Springer, Heidelberg, Vol. 77, Iss. 3, pp. 497-525, https://doi.org/10.1007/s41471-025-00219-4 This Version is available at: https://hdl.handle.net/10419/331934 Standard-Nutzungsbedingungen: Die Dokumente auf EconStor dürfen zu eigenen wissenschaftlichen Zwecken und zum Privatgebrauch gespeichert und kopiert werden. Sie dürfen die Dokumente nicht für öffentliche oder kommerzielle Zwecke vervielfältigen, öffentlich ausstellen, öffentlich zugänglich machen, vertreiben oder anderweitig nutzen. Sofern die Verfasser die Dokumente unter Open-Content-Lizenzen (insbesondere CC-Lizenzen) zur Verfügung gestellt haben sollten, gelten abweichend von diesen Nutzungsbedingungen die in der dort genannten Lizenz gewährten Nutzungsrechte. Terms of use: Documents in EconStor may be saved and copied for your personal and scholarly purposes. You are not to copy documents for public or commercial purposes, to exhibit the documents publicly, to make them publicly available on the internet, or to distribute or otherwise use the documents in public. If the documents have been made available under an Open Content Licence (especially Creative Commons Licences), you may exercise further usage rights as specified in the indicated licence. https://creativecommons.org/licenses/by/4.0/ ORIGINAL ARTICLE https://doi.org/10.1007/s41471-025-00219-4 Schmalenbach Journal of Business Research (2025) 77:497–525 Mining Consumer Mindset Metrics With User-Generated Content Raoul V. Kübler · Susanne J. Adler · Lina Welke · Marko Sarstedt · Koen Pauwels Received: 8 October 2024 / Accepted: 16 July 2025 / Published online: 4 September 2025 © The Author(s) 2025 Abstract In the wake of digital transformation, marketers gained access to large amounts of user-generated content and data in which consumers specifically mention and discuss brands, products, and services. This data offers rich information potential and may ultimately provide marketers with the ability to use this data pool to approximate survey-based consumer mindset metrics that mirror consumer attitudes alongside the different levels of the decision-making process. We argue that leveraging this potential may ultimately help marketers overcome common limitations of survey-based metrics and enable companies to observe and track mindset metrics that have been so far inaccessible due to financial and other constraints. To this end, we propose a four-step process that first identifies the key aspects of a mindset metric based on the existing body of developed constructs, then pinpoints potential data sources, and subsequently chooses an adequate data transformation tool. Keywords Mindset metrics · User-generated content · Construct validity · Machine learning · Customer decision journey Raoul V. Kübler · Lina Welke ESSEC Business School, 3 Avenue Hirsch, 95800 Cergy, France E-Mail: [email protected] Susanne J. Adler · Marko Sarstedt Institute for Marketing, LMU Munich School of Management, Ludwig Maximilian University of Munich, Ludwigstr. 28RG, 80539 Munich, Germany Marko Sarstedt Faculty of Economics and Business Administration, Babe, s-Bolyai-University, Strada Teodor Mihali 58–60, 400591 Cluj-Napoca, Romania Koen Pauwels D’Amore-McKim School of Business, Northeastern University Boston, 450 Dodge Hall, 360 Huntington Avenue, Boston, Massachusetts 02115, USA K 498 Schmalenbach Journal of Business Research (2025) 77:497–525 1 Introduction The digital transformation of our world led to substantial changes in how much data is available to decision-makers within the last two decades. With millions of digital devices observing and tracking our everyday lives, and consumers actively creating data points whenever they browse, shop, or interact with each other, we have entered the data-richest period in human history. Today, humans and the devices they use willingly and unwillingly create more than 402.74 million terabytes of data each day (Duarte 2024), surpassing the cumulated amount of data humankind has created since the dawn of civilization. A large proportion of this data is actively created by consumers through user-generated content (UGC) such as (but not exclusively) social media posts, reviews, videos, messages, online search patterns, online service interactions, website visits, or podcasts. Users provide valuable and informationrich data with a specific focus on consumption patterns, product preferences, satisfaction information, or company and product evaluations. In light of the growing availability of data, new methods have emerged to extract information and insights from large datasets available to decision-makers (Kübler et al. 2017). Within the last decade, substantial developments in machine learning techniques, coupled with the availability of more powerful hardware, enabled decision-makers to accurately describe, predict, explain, and re-generate human behavior. Deep learning algorithms have cumulated in large language models, which, through their chat interfaces or simple application descriptions, now allow access to powerful algorithms for generating consumer insights (e.g., through the generation of synthetic datasets; Sarstedt et al. 2024a) without the need for sophisticated coding expertise (e.g., Hartmann et al. 2023). The combination of large amounts of data, machine learning methods, and powerful hardware has also enabled managers with the ability of accurately measuring, tracking, or predicting consumer behavior such as brand and product consideration and perception (Ringel and Skiera 2016), as well as ad clicks (Wang et al. 2018), purchase (Chevalier and Mayzlin 2006), or churn behavior (Khodabandehlou and Rahman 2017). While machine learning applications have been shown to be able to ultimately predict final decision outcomes with the help of UGC, the prediction or approximation of latent variables such as consumer mindset metrics (CMMs) with UGC, has not yet been fully explored (Hair and Sarstedt 2021). Instead, managers commonly still rely on primary data to measure and track pre-purchase consumer attitudes such as brand awareness, brand consideration, or purchase intention, as well as postpurchase attitudes such as brand satisfaction, or recommendation behavior. This is especially surprising as primary data collected through surveys with consumer panels has been shown to face several restrictions such as high costs, reduced timeliness, and proneness to sampling, as well as response errors (Hulland et al. 2018). In addition, managers often rely on latent CMMs as key performance indicators in marketing, which are often included in corporate dashboards and used to evaluate the shortand long-term impact of marketing actions. Similar to other tasks in consumer (Ringel and Skiera 2016) and market observations (Matthe et al. 2023), UGC may be an alternative data source to approximate, measure, and track CMMs. Doing so would offer multiple advantages: First, secK Schmalenbach Journal of Business Research (2025) 77:497–525 499 ondary data such as UGC is continuously available and can thus also be used on a daily base to monitor brand health or measure and track the impact of marketing activities. Second, UGC is freely available and is commonly not prone to traditional biases known from primary data research such as strategic answering behavior, survey fatigue, and non-response bias. Third, UGC is commonly tied to true customer experiences (i.e., product reviews or customer complaints on X) and can thus be understood as a form of implicit, revealed preferences, whereas CMMs are commonly understood as stated preferences. Finally, being implicit feedback, UGC can be considered as a more natural form of feedback than survey-based CMMs, which are—being an explicit form of feedback—known to be more prone to, for example, social desirability bias. Despite the potential of UGC for consumer mindset tracking, so far little insights exist into the suitability and how to process, use, and leverage UGC to track latent CMMs. Addressing this gap in research, we set out to investigate and discuss the potential pros and cons of such an approach, as well as the necessary steps to facilitate the use of UGC in the context of CMMs. Our goal is to identify, systematize, and map the necessary tasks, develop guidelines for the usage of UGC to measure CMMs, and identify future research needs to ultimately develop a structured research agenda in this domain. The remainder of this paper is structured as follows: First, we highlight and discuss the current applications of CMMs in marketing research and marketing practice to determine the strengths and weaknesses of the current approaches as well as to determine the needs and requirements that need to be met by a UGC-based measurement approach. Subsequently, we highlight current applications of UGC in modern marketing research to determine the potential of UGC for CMM measurement and potential boundary conditions. Linking insights from both sections we then develop aframeworkfor combining both worlds in which we highlight necessary steps that need to be addressed to leverage the potential of UGC for CMM measurement. Finally, we develop a research agenda to stimulate further work in the field that is necessary to ultimately reach the goal of making UGC available to managers for measuring and tracking CMMs. 2 Consumer Mindset Metrics 2.1 Concepts of Customer Mindset Metrics Both psychology and marketing literature identify consumer attitudes as constructs that indicate how consumers think about (cognition), feel about (affect), and act toward (conation) a brand (Vakratsas and Ambler 1999) and other objects of interest. Starting in the early 1960s (Colley 1961; Lavidge and Steiner 1961), marketing developed measures of consumer attitudes to evaluate the impact of marketing campaigns and to predict their sales effect. In their theory of buying behavior, Howard and Sheth (1969, p. 14) noted, “Attitude is an input into executive decisions because many marketing decisions, including advertising, can be more adequately evaluated or measured in terms of attitude than of purchase behavior.” Corresponding attitudeK 500 Schmalenbach Journal of Business Research (2025) 77:497–525 related measures are also referred to as CMMs in that they seek to capture the whyquestion necessary to interpret observed consumer behavior. For instance, consumers may visit a brand’s website because they are considering buying its products, are browsing for fun, have questions about products already bought, or want to rationalize their choice for a competing brand (Dotson et al. 2017). Pauwels and van Ewijk (2020) show that attitude-related metrics offer a superior prediction of brand sales over several months, as compared with aggregate online behavioral metrics such as the amount of weekly clicks and visits. The literature has brought forward a multitude of CMMs. The most prominent CMMs are arguably brand awareness, brand consideration, purchase intention, brand satisfaction, brand recommendation, and brand equity. Brand awareness captures whether a brand is top-of-mind and can be measured as a consumer’s ability to recall a brand’s name (unaided) or as a consumer’s ability to recognize a brand among other brands (aided; see, e.g., Mecredy et al. 2022; Pauwels and van Ewijk 2020). Brand consideration describes which brands consumers would regard as suitable options for usage or purchase while purchase intention captures which brands consumers would actually be willing to buy (e.g., Kübler et al. 2020; Mecredy et al. 2022). CMMs that can only be formed after having purchased and used a product or service include brand satisfaction and brand recommendation. Brand satisfaction captures whether a brand has met a consumer’s expectation while brand recommendation measures a consumer’s willingness to suggest the brand to others, respectively to tell others to avoid a brand (e.g., Anselmsson and Bondesson 2015; Kübler et al. 2020). Lastly, brand equity captures a consumer’s preference for one brand over other brands (Washburn and Plank 2002). 2.2 The Psychometrics of Customer Mindset Metrics As the concepts underlying CMMs are inherently unobservable, their measurement typically relies on consumers’ answers to sets of survey items. These items are supposed to capture the respondents’ assessment of specific traits that underlie the concept under consideration (e.g., Anselmsson and Bondesson 2015; Pauwels and van Ewijk 2020). For example, to measure customer satisfaction, researchers routinely rely on three survey items that gauge the respondents’ overall satisfaction, expectancy confirmation, and perceived performance versus the customer’s ideal product or service in the category (Fornell et al. 1996). From a measurement-theoretic perspective, these items are assumed to be reflections or consequences of customer satisfaction (hence, reflective measurement), thereby acting as empirical realizations of the unobserved concept; their correlation is assumed to be “caused” by the underlying concept (Sarstedt et al. 2016a). A formative measurement of customer satisfaction, on the other hand, would capture contributing traits such as satisfaction with the service, price, or product; assuming that a composite of such individual traits captures the target construct. The logic underlying formative measurement acknowledges the approximative nature of measurement whereas reflective measures implicitly assume that any measurement can, in principle, capture a concept in full (Rigdon and Sarstedt 2022). The manifold uncertainties that come with any measurement, for example, with reK Schmalenbach Journal of Business Research (2025) 77:497–525 501 gard to the conceptualization, item wordings, measurement scales, data collection, and validation, however, call the assumption of perfect measurement into question (Rigdon et al. 2019,2020; Sarstedt et al. 2024b). Considering the practical challenges that come with collecting data on multiitem measures (e.g., with regard to careless response behavior or non-response), researchers and particularly practitioners frequently rely on single-item measures of a concept. These single items can be seen as a global assessment of the concept, such as when asking a respondent: “Overall, how do you rate your satisfaction with the company’s service?” The practical advantages, however, come at the expense of reduced reliability and lower predictive validity, which may compromise their usefulness in research settings, potentially triggering flawed managerial recommendations (e.g., Diamantopoulos et al. 2012; Sarstedt et al. 2016b, c). In their studies, managers typically customize the exact measurements to their brand or make changes in response to observed relationships, for example, pruning metrics that were too highly correlated amongst each other or did not predict behavior over time (e.g., Pauwels and Joshi 2016). This measurement flexibility also applies to measures for CMMs that are not fully standardized. Research has witnessed a proliferation of all sorts of metrics that claim to measure essentially the same concept, although often with little chance to convert one instrument’s measures into any other instrument’s measures (Salzberger et al. 2016). For example, research and practice have proposed a multitude of measurement instruments for corporate reputation, which rest on the same definition of the concept but differ fundamentally in terms of their underlying conceptualizations and measurement items (Sarstedt et al. 2013). Similarly, Bergkvist and Langner (2017,2019) find considerable heterogeneity in the operationalizations of common advertising constructs, such as attitude toward the ad, attitude toward the brand, ad credibility, ad irritation, and brand purchase intention. In addition, construct conceptualizations and operationalizations change over time (Bergkvist and Eisend 2021), while the theoretical entity of interest (i.e., the conceptual variable) generally remains the same. These findings suggest there is no set way to perfectly measure a concept, even though some standardized measurement approaches can be observed. For example, brand awareness is measured as the unaided recall and aided recognition of a brand (e.g., Mecredy et al. 2022; Pauwels and van Ewijk 2020). Other CMM measures exhibit a lower consensus. To measure brand consideration, purchase intention, brand satisfaction, or recommendation intentions, consumers may select relevant brands from a list of multiple brands (e.g., Colicev and de Bruyn 2023; Kübler et al. 2020; Mecredy et al. 2022) or indicate their assessment on a Likert-scale (e.g., Anselmsson and Bondesson 2015;Cain2022; Pauwels and van Ewijk 2020). Researchers have also suggested the use of a Net Promoter Score for which respondents indicate their willingness to recommend a brand on a scale from 0 to 10 and are sorted into endorsement categories because of their rating (Baehre et al. 2022). Measures of satisfaction range from asking participants to select brands they are satisfied, respectively not satisfied, with (Kübler et al. 2020) to using multi-item scales (Anselmsson and Bondesson 2015; Petersen et al. 2018). Moreover, the relationship between survey-measured CMMs and consumer behavior is not strong pointing toward potentially limited predictive validity (Pauwels and van Ewijk 2020). K 502 Schmalenbach Journal of Business Research (2025) 77:497–525 Table 1 Popular UGC Channel Types and Examples Channel Type Examples Exemplary studies Social media platforms Facebook, Instagram, X, TikTok, LinkedIn, Sina Weibo, Snapchat Liu et al. (2016): Twitter (now X) Tweets to predict TV show demand Online review platforms Yelp, TripAdvisor, Trustpilot Rocklage et al. (2021): Emotionality of Yelp reviews to predict restaurant table reservations E-commerce platforms Amazon, eBay, Etsy Chevalier and Mayzlin (2006)usereviewsto predict sales Search engine logs Google, Bing, Yahoo Ringel and Skiera (2016) use Google search term volume to map consumer considerations Blogs and microblogs Medium, Substack, Blogger, Tumblr Onishi and Manchanda (2012) examine impact of blog posts on brand awareness and sales Q&A forums Quora, Stack Overflow, Reddit Burtch et al. (2022) examine the impact of badges on UGC creativity Streaming platforms Twitch, YouTube Live, Douyin, Huya Live, RedNote Lin et al. (2021) examine impact of emotions in live streams on tips and user activity Podcasts Spotify, Apple Podcasts, SoundCloud Kozinets et al. (2010) investigate how podcastbased seeding influences spread of UGC and eWoM 3 User-Generated Content UGC has been in the spotlight of marketing research now for over two decades. Initially, often generally referred to as (online or electronic) word of mouth (Godes and Mayzlin 2004; Hennig-Thurau et al. 2004), the understanding and definition of UGC have become more granular over time. While the traditional digital or electronic word of mouth literature limits itself to UGC that specifically focuses on consumption-related content (Babi´ c Rosario et al. 2020), the classic UGC literature goes beyond and includes any sort of digital information created by a private (in contrast to a paid) user’s communication and online behavior and that is shared with the general public via any sort of digital (web) channel (Daugherty et al. 2008). UGC data sources investigated in the literature include a broad range of different channels. Table 1provides an overview of popular channel types, along with examples. With the constant emergence of new channels, the volume and variety of digital traces left by consumers continue to expand. A key characteristic of UGC is that it rarely occurs in a structured (i.e., numeric) data format (de Haan et al. 2024). If so, often simple count measures are used as approximations, such as popularity (e.g., the number of followers), brand liking (e.g., through the count of daily likes of brand posts), and brand-consumer relations (e.g., through the count of daily brand post shares). These simple approaches, however, neglect the richness of information within UGC that arises from the fact that users do not only create information by engaging with brand content, but also actively create posts, comments, videos, chats, etc. More importantly, as evidenced in the list of data sources above, the vast majority of UGC comes in unstructured (i.e., non-numeric) data such as images, text, and videos. Several studies have made use of such data to generate consumer insights. For example, the majority of research relying on UGC to approximate common brand equity metrics such as brand knowledge or brand K Schmalenbach Journal of Business Research (2025) 77:497–525 503 strength have commonly drawn on textual data obtained from various online and social media sources (see, e.g., Colicev et al. 2018). Fewer studies, however, have gone one step further and used non-numeric UGC, such as images (Dzyabura et al. 2023; Hartmann et al. 2021), audio (Wang et al. 2021), or video (Zhou et al. 2021) to capture consumer or brand insights. Unstructured data brings the advantage of richer information and thus bares more potential for generating unique insights (de Haan et al. 2024). At the same time, however, UGC data needs to be transformed into structured data in order to be further processed and analyzed. Thus, the need for data transformation must be considered as another key characteristic of UGC. De Haan et al. (2024) provide a detailed and structured framework for how to turn unstructured data into a suitable structured format, taking various application and company-specific contingencies into consideration. The most commonly used tools for data transformation—at least in the context of brand measurement (Berger et al. 2020)—are sentiment measures. For textual data, sentiment measurement tools can be split into top-down and bottom-up approaches (Humphreys and Wang 2018). Top-down approaches such as the often-used LIWC or VADER commonly rely on word lists curated by linguistic research to capture sentiment or valence information. By contrast, bottom-up approaches use machine learning tools and pre-coded training data to predict the sentiment of a document by looking at the embedded words or word combinations within the document (Kübler et al. 2020). In the case of audio, either transcriptions of the spoken content are used to rely on standard textual sentiment classifiers (Zhou et al. 2021), or tools to detect loudness or tone within voice to capture sentiment (Wang et al. 2021). For images, research so far most often relies on facial expressions to capture sentiment and valence (Toisoul et al. 2021), which, however, requires images to display faces. Similarly, in the case of videos, the video material is cut into individual frames, which are then subsequently treated like images (e.g., Li et al. 2019). Besides the binary measurement of sentiment (positive and negative), an increasing number of studies now attempt to capture not only the valence, but subdimensions of emotions such as anger, fear, anxiety, sadness, joy, surprise, and anticipation (e.g., Holiday et al. 2023; Schwenzow et al. 2021; Zhou et al. 2021). 4 Connecting Conceptual Facets With UGC Insights CMMs are typically measured with sets of items that cover multiple facets and aspects of the concept under consideration. The latent concept of satisfaction may, for example, be measured formatively by asking consumers to rate various aspects of a service (e.g., cleanliness, proximity, calmness, taste), which jointly form satisfaction. In case of a reflective approach, the same concept would be measured by the attempt to capture the outcome of the construct such as, “How satisfied are you with the brand?” and “How well does the brand match your expectations?” (Fornell et al. 1996). Traditional CMMs measurement therefore captures stated preferences, collected through explicit feedback. In contrast, UGC inherently reflects aspects of revealed preferences embedded in implicit feedback. This content often contains K 504 Schmalenbach Journal of Business Research (2025) 77:497–525 Table 2 Advantages and Disadvantages of Traditional Data versus UGC for CMM Measurements Advantages Disadvantages Traditional data Sampling: Potential to ensure and check for the representativeness of a sample for a specific population Targeted sampling of, for example, specific groups Sampling: Potentially high data collection costs for recruiting and incentivizing study participants Reduced timeliness Prone to sampling errors including non-response biases Potential careless response behavior Scale use: Direct measurements, for example, concerning attitudes Validated, standardized measures available Scale use: Scale-specific disadvantages (e.g., reduced reliability and lower predictive validity in singleitem scales) Many implicit assumptions in scale design increase uncertainty (e.g., scale types, wording of items and answer categories) Use of unstandardized, ad-hoc measures Scale modifications may change the nature of measurement; necessary scope often unclear Emergence of response biases (e.g., social desirability bias) UGC Captures real-world consumer behavior Available from multiple sources, continuously available, and often freely available Rich information from combining multiple data sources available Less prone to traditional biases known from primary data research (e.g., social desirability, survey fatigue, non-response bias) Uncertain data composition in terms of target groups who provide UGC Self-selection bias (e.g., dominance of negative responses and data from very active users) Risk of incomplete construct assessment Measures are mostly indirect and approximative Object of measurement may be ambiguous (e.g., product-related vs. brand-related assessments) Combination of data from various sources can be prohibitively complex and expensive Data is mostly unstructured, and transformation is often necessary K Schmalenbach Journal of Business Research (2025) 77:497–525 511 Table 3 (Continued) Metric Characteristics Sample Items Text-based sources Image-based sources Video-based sources Brand recommendation Consumers’ willingness to encourage others to buy or use a brand Which of the following brands would you recommend to a friend or colleague? Which of the following brands would you tell a friend or colleague to avoid? How likely is it that you would recommend brand X to a friend or colleague? Use word lists/dictionaries to track recommendation-specific posts. Extract Net Promotor Score-like elements (e.g., 10 out of 10). Use topic models to identify and track testimonial statements. Extract and collect recommendation scores from, e.g., reviews. Identify UGC where users tag others recommending the brand. Apply NLP tools to captions, tags, or comments. Use computer vision tools to identify interactive elements on visual social media such as interactive stories, polls, or recommender links. Use computer vision tools to identify the context in which testimonials present the product. Apply NLP tools to audio transcripts, captions, tags, or comments. Apply computer vision tools to video frames. Brand equity The ability of a brand to affect buying decisions based on objective or subjective characteristics consumers value more for the focal brand than for its competitors It makes sense to buy X instead of any other brand, even if they are the same. Even if another brand has the same features as X, I would prefer to buy X. If there is another brand as good as X, I prefer to buy X. If another brand is not different from X in any way, it seems smarter to purchase X. Count the number of statements where users express a rationale for choosing a focal brand over others. Rely on word embedding models to identify the number of instances in which UGC acknowledges that competitors might be “as good” but still express preference for the focal brand. Rely on comparison mining methods (use dependency parsing or other syntactic tools to extract comparative statements involving focal brands and competitors). Use ABSA to conduct reasoning extraction, by employing a logic classifier that identifies phrases where users rationalize or justify their choices. Apply NLP tools to captions, search for hashtags, and other identifiers. Rely on object classifiers to extract logos and understand which brands are more often or more prominently featured. Rely on object classifiers to grasp setting and apply word vector models to identify objects to understand the fit between content and brand. Count the number of repeated brand features/occurrences in a specific setting or for specific keywords or search terms. Apply NLP tools to captions, search for hashtags, and other identifiers. Create transcripts from voiceover andapplyNLP tools to obtained text. Rely on vector-space models of brand mentions to identify comparison across brands, rely on distance from model to understand strength. K 512 Schmalenbach Journal of Business Research (2025) 77:497–525 unstructured data, which limits its effectiveness in measuring consumer attitudes like brand satisfaction or recommendation. Podcast platforms primarily provide largely unstructured data through transcripts rather than numeric engagement metrics such as likes or stream counts. While these transcripts can offer rich insights into consumer evaluations and discussions about a brand, they are less useful for gauging brand awareness at the beginning of the marketing funnel. 5.3 Information Extraction In the third step, researchers need to extract the aspect information from the UGC with the help of adequate tools. As shown in Table 3, various approaches and methods are suitable for this task. The choice of a method depends on the data format (structured vs. unstructured), the complexity of the data, the type of data (text vs. images vs. videos), as well as budget and time constraints. In the case of top and bottom of the funnel metrics, such as brand awareness and satisfaction, where numeric information is available and easily accessible, simpler methods that rely on count data may be sufficient to provide initial insights (see, e.g., Kübler and Seggie 2024). Tracking the number of followers of social media platforms or counting the number of review ratings can validly depict brand awareness or satisfaction, as demonstrated by numerous studies in the field using such data (see, e.g., Hewett et al. 2016; Colicev et al. 2018; or Kübler et al. 2018)to approximate or explain brand awareness or satisfaction. However, understanding the various facets of these constructs requires a more nuanced approach that draws on unstructured data and applies more complex tools. For CMMs, where the identification of brand mentions to capture brand awareness or brand consideration, is essential, researchers may rely on tools developed for object identification in unstructured data. This may involve brand name lists (or dictionaries) to tag social media posts or calculate share of voice scores by comparing how many times the own brand is mentioned compared to others (e.g., Ringel and Skiera 2016). For non-textual data, object classifiers or object detector models may be used to recognize brand logos or key products to achieve a similar outcome. Besides using one of the various commercial off-the-shelf solutions built into AWS and MS Azure or open-access tools hosted on platforms such as HuggingFace (see, e.g., https://huggingface.co/spaces/nathanjc/Logo_detection_ YoloV7), researchers may build their own detector models. Training and fine-tuning such classifiers have become easier with various packages in R or Python, offering the application of state-of-the-art detection models—see, for example, Yildirim and Kübler (2023) for a logo identifier markdown approach. The availability of pretrained models on platforms such as HuggingFace that only require a “few shots” of training data (commonly less than 100 images per object) to deliver reliable results, has also significantly contributed to the dissemination of classifiers—see, for example, Carion et al. (2020) for a transformer-based object detection model. In situations where not only brand awareness needs to be tracked, but also brandrelated elements such as dimensions of brand equity (e.g., Yoo and Donthu 2001), other machine learning-based techniques may be helpful. For textual data, topic K Schmalenbach Journal of Business Research (2025) 77:497–525 513 models, which allow the identification of latent aspects within text data (Büschken and Allenby 2016), may help extract such insights. The key idea of these models is that specific words are tied to topics within documents and that documents are composed of different topics (Blei et al. 2003). Different forms of topic models are available with Latent Dirichlet Allocation (LDA) models being the most widely applied ones (e.g., Adler and Sarstedt 2021; Adler et al. 2024). However, researchers increasingly apply structural topic models (STM), which can account for evolving trends over time (Roberts et al. 2019), or models that account for word embeddings such as BERTopic (Grootendorst 2022). For more details and a step-by-step guide on how to estimate topic models, see Yildirim and Kübler (2023). In the case of image (and video) data, object classifiers can similarly be used to capture contextual information and derive insights about various aspects or the meaning of a brand. By counting or measuring the types of objects commonly shown alongside a brand or product, researchers may obtain better information about how consumers perceive brands. In some cases, this may require a combination of tools, where an object detector is used to identify all objects occurring with a brand, and word-vector models are subsequently used to automatically derive meaning from the obtained brand-object-embeddings—see, for example, the approach of context fit measurement that Kübler et al. (2024)present. As shown in Table 3, the ability to capture and integrate contextual information becomes increasingly relevant for mid-funnel CMMs, such as brand consideration and purchase intention. For the former construct, when relying on textual data, precurated word lists, topic models or vector models may be used to identify which brands are commonly mentioned together in a consumption context. Similarly, researchers may employ these methods to determine which other brands are frequently mentioned or searched alongside their focal brand to depict the consumers’ consideration set (e.g., Ringel and Skiera 2016). Subsequently, the share of voice or mentions can be used to understand the position within the consideration set as a measure for the strength of consideration. The same can be achieved with the help of object classifiers in the context of visual information, such as images or videos. While share of voice may suffice to capture brand consideration, purchase intention may require a more detailed understanding of how strongly a UGC element suggests that the content creator prefers a mentioned brand or product over its alternatives (see, e.g., Karniouchina et al. 2022). To address this concern, researchers may rely on textual data using either word dictionaries (Pennebaker 2001)thatdetect futureand goal-oriented language (like LIWC) or part-of-speech taggers that find instances where users indicate clear buying intentions (e.g., through the use of words like “can’t wait to get it” or “going to buy”) or preferences (e.g., through words such as “better than” or “would buy X instead of Y”) as, e.g., demonstrated by Sepehri et al. (2023). Similarly, sentiment scores can be applied to UGC that discusses or compares alternatives. In particular, aspect-based sentiment analysis (ABSA) can be leveraged to assign a sentiment score to each brand mentioned in a UGC piece (Do et al. 2019). ABSA can be understood as a combination of topic modeling and sentiment analysis. Pre-trained state-of-the-art models are available on all common platforms and can be fine-tuned with few shots (see, e.g., SetFitABSA on https://huggingface.co/blog/setfit-absa). Fine-tuning requires researchers to proK 514 Schmalenbach Journal of Business Research (2025) 77:497–525 vide the model with training data that features aspects (or in our case, brands) and examples of positive and negative brand-specific comments (Liu et al. 2020). For images and videos, one may apply the same techniques to textual data coming with an image or video, such as captions, comments, subtitles, or transcripts of voiceovers (see, e.g., Ghosh and Deb 2022). Meanwhile, the identification of purchase-specific UGC may be achievable by first trying to identify images or videos that are tied to a purchase context (e.g., point of sale content, unboxing posts, or product haul videos) to subsequently identify which brands are featured more often, as demonstratedbyOrtietal.(2019). The importance of capturing both context and content within UGC to accurately understand the relevant aspects becomes even more important for the post-purchase CMMs. In the case of satisfaction, one may directly assess satisfaction by the source information (i.e., by examining reviews or review-specific content, such as review images or product review videos) and rely on the emotions expressed within this content to capture the direction and the strength of satisfaction. Emotions can be captured in different ways. For textual data, emotion dictionaries such as EmoLex (often also referred to as NRC; Mohammad and Turney 2013) allow a quick and easy measurement of various emotional dimensions such as joy, surprise, anticipation, anger, disgust, and fear. However, EmoLex suffers from the common dictionary-based shortcoming as it does not capture negations and is only available for a few major languages (Berger et al. 2020). Hartmann et al. (2023) provide a pretrained sentiment analysis transformer-based model that captures the same emotional dimensions, offers greater accuracy, and can be fine-tuned to brand-specific contexts. The resulting information can then be used to create a satisfaction index, as demonstrated by the online video game distribution platform Steam, which also uses an emotion classifier—accessible through Steam’s main API—to derive a satisfaction score from its review texts (Guzsvinecz and Sz˝ucs 2023). Hotz-Behofsits et al. (2025) furthermore present a new tool that first “emojifies” textual information to then infer emotions from the emojified content. Similarly, computer vision as a research discipline that attempts to transform unstructured visual content into a structured format (Voulodimos et al. 2018) provides tools for the identification of facial emotions, allowing to capture an emotion score from product review video frames. Both approaches can also be applied to measure brand recommendation, a CMM at the bottom of the funnel. Here, one may actively use word lists or topic models to identify whether a piece of UGC contains a specific recommendation, before applying the previously discussed NLP and computer vision tools to determine which brands are recommended and for what reasons. In addition, referrer and affiliate links, poll mechanisms, and scores that resemble classic ratings or net promoter score voting elements (such as “I would give 9 out of 10” or “for me a clear five-star hotel”) may be used to identify recommendation-specific content. In this case, the numeric information can also be leveraged to understand how strongly a piece recommends a brand or product. K Schmalenbach Journal of Business Research (2025) 77:497–525 515 5.4 Validation Once one has collected and transformed the data with the help of the methods described above, the question arises, how well the obtained measures depict the CMM and if the obtained measures can be used to track customer mindset along the decision-making funnel. The validation of UGC measures may, in principle, follow standard approaches well-known from the psychometrics literature, while accounting for the specificities of the data type. This process starts with evaluating how well the UGC measure captures the concept domain identified in Step 1 (Fig. 1). Such an evaluation is subjective, but systematic in nature, typically drawing on researchers with expert knowledge in the domain. This qualitative assessment should be followed by a quantitative analysis that examines the degree to which the UCG measure correlates with an alternative measure of the same concept (convergent validity) and its ability to predict relevant outcome variables (predictive power). The straightforward approach to assess convergent validity is to use survey-based CMMs and estimate the correlation between the survey and UGC data (e.g., Kübler et al. 2020). Such an analysis, however, requires researchers to have control data at hand, which is often not the case due to financial constraints or because such data has not been collected before and is thus not available. In some cases, researchers may draw on secondary data proxies as surrogates for primary survey data. For example, research institutions have used customer complaints as a proxy for (dis)satisfaction (Hunt 1991). Such secondary data proxies rarely cover the concept’s domain in full, but may still act as reasonable standards of comparison, provided that the literature offers support for their use (Houston 2002). Without such theoretical support, examining the proxy’s correlation with other UGC measures within the study can offer clues regarding discriminant validity (i.e., the degree to which measures of different concepts are sufficiently empirically distinct). The primary concern in the validation process should be establishing the UCG measure’s predictive validity to support its relevance for managerial decision-making (e.g., Sarstedt and Danks 2022). Predictive validity assessment may also draw on survey-based control data, but in the absence of such data, researchers may test the obtained metrics by relying on the general concept of the decision-making funnel. According to this concept, investments at each level of the funnel (e.g., awareness) will, over time, move down the funnel and affect lower levels of the decision-making process (e.g., Colicev et al. 2018). Knowing that each CMM level is linked to (or predicts) a lower-level CMM and, in turn, to observable outcomes, one can develop a traditional KPI-type model. Recognizing that there is a causal chain between awareness, interest, consideration, purchase intention, satisfaction, and recommendation, researchers may attempt to tie the metrics obtained from UGC to performance variables that are observable at each stage following the lead performance indicator approach that Pauwels (2014) describes. By understanding how well each UGC-based CMM measurement predicts metrics further down the funnel, alongside performance variables available within the company (such as leads, purchases, or repeat purchases), researchers may be able to confirm the validity of the K 516 Schmalenbach Journal of Business Research (2025) 77:497–525 obtained UGC measures. For a guide on how to conceptually build and empirically test such a model, we refer to Hanssens et al. (2014) and Yildirim and Kübler (2023). While establishing convergent and predictive validity should be relevant for all types of UGC measures, researchers may consider further validity types, depending on the nature of the data and its source. For example, Berger et al. (2020) provide a comprehensive framework for validating measures derived from text analysis. Rust et al. (2021) propose a real-time brand reputation tracker based on social media mining using a customer equity framework. Their work highlights the value of theory-driven UGC analysis and provides robust validation techniques by linking brand drivers to financial outcomes. Finally, Houston (2002) introduced a validation guideline for measurements of marketing constructs that draw on secondary data. This procedure comprises a three-step process, which partly overlaps with our framework (e.g., theoretical specification), but considers additional elements such as one-dimensionality and nomological validity assessment. 6 Discussion Understanding how UGC translates into consumer perceptions, attitudes, and intentions has become crucial for managerial decision-making. While marketing researchers and practitioners often rely on numeric data (e.g., like counts) to approximate consumer assessments, UGC offers much richer information that can be reaped with today’s machine learning tools. Addressing the challenges that come with such an approach, we developed a four-step process that uses item contents identified in psychometric assessments of CMMs as a blueprint for source identification as well as information extraction and control. To facilitate the adoption of our framework, we classify UGC data sources according to their suitability for capturing prominent CMMs and identify tools to do so. Applying our framework, marketing researchers and practitioners can efficiently leverage UGC data to measure CMMs. This approach offers further benefits to the field: The combination of different UCG measures in terms of, for example, numeric data (e.g., likes on social media) and text data (e.g., reviews on retailer websites) can help deepen the understanding of a CMM through triangulation. Insights from UGC data can also be combined with scale measurements from customer surveys to bring the best of the old and the new worlds together. Additionally, due to its timeliness and lower level of theoretical input, UGC can also help improve CMM’s survey measurements. Specifically, comparing results from UGC to survey items indicates which scale dimensions or items are represented well in UGC but also which aspects are missing. Indeed, consumers’ online reviews could help identify which product or service features are likely linked to customer satisfaction. Such assessments can inform formative measurement approaches to ensure that the items capture the concept in full. Similarly, the absence of themes in UGC could help refine scales. Respondents in surveys must often answer all items to proceed in the questionnaire—irrespective of whether a measurement is actually useful (Avis et al. 2014). If a repeated assessment on the grounds of UGC does not identify a specific measurement aspect, it may not be relevant and could be considered for removal in future psychometric assessments. UGC can therefore K Schmalenbach Journal of Business Research (2025) 77:497–525 517 Table 4 Agenda of Future Research Research Agenda Specific Research Questions Conceptualization Which aspects are suitable for product-specific CMM characteristics? Which aspects are suitable for brand-specific CMM characteristics? How strongly can UGC based CMM measures be standardized vs. need of customization? Which metric/UGC data combination produce which level of uncertainty? Source Identification Which data source is suitable for which CMM aspect? How to identify brandand product relevant UGC for CMM measurement? How to sample UGC data according to the target group? Which mix of UGC sources is necessary to ensure representativeness? How to avoid J-shape trap of valence? How to best avoid data dependence and ensure permanent data access? How much is UGC data prone to manipulations and how to avoid fake insights? Information Extraction Which content measure tool for which CMM aspect and data source? Which CMM aspects generally require which extraction tool traits? What are costs vs. benefits trade-offs of measurement approaches in terms of CMM measure improvement? What are costs and benefits of multimodal vs. singular model measurement approaches Validation What are suitable validation guidelines for UGC-based CMMs? Do we need different guidelines for different types of CMMs? Which reliability and validity types should be considered in the validation of UGC-based CMMs? How can fake reviews and AI-generated content be identified? K 518 Schmalenbach Journal of Business Research (2025) 77:497–525 serve as an outside validity check for established scales that would be specifically useful if the construct is context-dependent or time-variant. Due to the variety of channels that produce UGC, researchers may also draw on a wide range of data from diverse populations (e.g., teens and younger adults on TikTok and business professionals on LinkedIn). With their specific socio-demographics, experiences, and attitudes, using UGC for scale refinement may aid or replace exploratory steps in scale development. 7 Future Research While our framework sets the stage for using UGC for CMM measurement, it also offers room for follow-up research, which we summarize in Table 4. As for the identification of key CMM construct aspects, we first focused on extending measurements of CMMs with UGC. Compared to the unstructured data that UGC provides, CMMs come in the form of standardized measurements. However, measures validated on psychometric grounds provide an even higher level of standardization and can be accompanied by norms that identify which values on a scale are high or low compared to a specific population (Rigdon and Sarstedt 2022). This level of standardization is, as of now, not achieved by measures relying on UGC. Future research should thus prioritize the delivery of an understanding whether UGC CMM measures always need to be customized or if they can also be standardized, and if so, what is required for this process. Relatedly, in applying our framework, researchers need to explicitly acknowledge that UGC data only allows for approximating the concepts of interest. Recent research highlights the role of measurement uncertainty in psychometric assessments of concepts in this context, which induces a validity gap between the theoretical concept and the construct (e.g., Rigdon et al. 2019). Using UGC data is likely to widen this gap due to the idiosyncrasies of the data and the researcher’s degrees of freedom in the processing and analysis of the data. The use of UGC in CMM measurement therefore emphasizes Rigdon et al. (2019) call to put greater effort into quantifying measurement uncertainty (see also Rigdon and Sarstedt 2022). Future studies should also develop validation guidelines for metrics derived from UGC. While content, convergent, and predictive validity, as highlighted above, are important elements, other aspects of validity are also relevant. Corresponding studies may build on the conceptual works of Berger et al. (2020) and Rust et al. (2021), or draw on extant guidelines proposed in related contexts (e.g., secondary data validation; Houston 2002). The goal should be to develop a comprehensive approach that enables researchers to assess the validity of a given metric based on UGC traits, CMM aspects, and external validation data characteristics. This approach should help determine how well a metric aligns with existing external data sources and evaluate its validity holistically. In doing so, future research should thus also target the question of which aspects drive or minimize metrological uncertainty, which has been identified as a major contributor to low replication rates (Rigdon et al. 2019). Regarding data source management, future research should first establish a clearer understanding of which sources—and combinations of sources—are most suitable K Schmalenbach Journal of Business Research (2025) 77:497–525 519 for UGC-based CMM measurement. This involves assessing the representativeness of data sources for a given target group and, on the other, developing a more robust approach to constructing a valid sample from diverse UGC sources. A key challenge lies in the inherently polarized nature of UGC, where users tend to either strongly praise or harshly criticize a brand or product. To mitigate the well-known J-shaped distribution issue observed in online review ratings (Hydock et al. 2020), researchers must explore effective strategies for capturing neutral UGC as well. Finally, it is important to recognize that UGC is susceptible to manipulation, such as fake reviews. Future research should therefore explore methods for detecting and removing invalid or manipulative content from samples. Relatedly, the quality of UGC-based CMM measures can be compromised by the presence of non-organic content, such as fake reviews or AI-generated texts (e.g., Kovács 2024). Fake reviews, often generated for promotional or malicious purposes, can introduce significant biases in CMM measures by distorting sentiment distributions, misleading consumer perception analyses, and reducing the overall credibility of insights derived from UGC. Despite efforts by platforms to detect and filter fraudulent content, evolving techniques in automated review generation and coordinated manipulation campaigns continue to pose challenges (Wu et al. 2020). Similarly, the rise of AI-generated texts introduces another layer of complexity. With advancements in natural language generation, AI systems can produce highly convincing UGC, ranging from product reviews to social media posts. While AIgenerated content can enhance engagement and streamline communication, it also risks undermining the validity of CMM measures if such content lacks genuine consumer intent or introduces artificial sentiment trends, particularly if these are being reinforced by the reuse of AI-generated data (Xing et al. 2025). To mitigate these risks, future research and practice should focus on developing robust filtering mechanisms, integrating AI-powered detection tools, and refining validity controls for UGC-based CMM measures. While UGC is widely available online, accessibility remains a critical challenge for researchers. Many platforms increasingly restrict automated data retrieval through their terms of service, advanced technical measures, or limited API access (Boegershausen et al. 2022), thereby seeking to maintain competitive advantages. For instance, platforms like Amazon or Google employ advanced measures to prevent web scraping. Public APIs, by contrast, offer a more efficient alternative but are often expensive, rate-limited, or subject to lengthy application processes. For example, X charges $ 5000 per month for its Pro API to retrieve 1,000,000 posts (X, 2025), while TikTok’s free academic API is restricted to select regions and institutions (TikTok 2024) Such restrictions may intensify in the future, posing significant barriers to researchers and disproportionately impacting smaller or mid-sized companies that lack the resources to obtain proprietary datasets or access expensive APIs. Reflecting on these dynamics, future research should address the implications of limited data accessibility and explore alternative ways to gather and utilize UGC data. As de Haan et al. (2024) and Kübler (2023) highlight, research has provided a rich and expanding array of UGC information extraction tools. New methods and tools often lead to classic “benchmarking” studies, which compare the suitability K 520 Schmalenbach Journal of Business Research (2025) 77:497–525 of a new tool against existing tools. This chain of research becomes increasingly difficult to follow due to the rapid development of methods. We suggest that future research should instead focus more on the aspects of the different CMMs and assess methods on a meta-level rather than solely on their performance increases. We further encourage future research to shift its focus toward identifying which tools are best suited for specific CMM-related tasks. This includes not only assessing the cost-benefit trade-offs of available tools but also determining whether a singletool approach (singular measurement) is sufficient or if a combination of tools (i.e., a multimodal approach) would yield better results. Finally, while our framework considers important CMMs, future studies should extend it to further metrics that feature prominently in the literature. These include, but are not limited to measures of corporate reputation (Sarstedt et al. 2013)and brand image (Driesner and Romaniuk 2006). Despite its limitations, our framework offers a first step into leveraging UGC. Researchers and practitioners can draw on the various tools described in our article to put the framework into practice. Funding The authors did not receive support from any organization for the submitted work. Conflict of interest R.V. Kübler, S. Adler, L. Welke, M. Sarstedt and K. Pauwels declare that they have no competing interests. Open Access This article is licensed under a Creative Commons Attribution 4.0 International License, which permits use, sharing, adaptation, distribution and reproduction in any medium or format, as long as you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons licence, and indicate if changes were made. The images or other third party material in this article are included in the article’s Creative Commons licence, unless indicated otherwise in a credit line to the material. If material is not included in the article’s Creative Commons licence and your intended use is not permitted by statutory regulation or exceeds the permitted use, you will need to obtain permission directly from the copyright holder. To view a copy of this licence, visit http://creativecommons.org/licenses/by/4. 0/. References Adler, S., and M. Sarstedt. 2021. Mapping the jungle: A bibliometric analysis of research into construal level theory. Psychology & Marketing 38(9):1367–1383. Adler, S.J., M.K. Schöniger, M. Lichters, and M. Sarstedt. 2024. Forty years of context effect research in marketing: A bibliometric analysis. Journal of Business Economics 94(3):437–466. Albers, S. 2010. PLS and success factor studies in marketing. In Handbook of partial least squares: Concepts, methods and applications in marketing and related fields Springer Handbooks of Computational Statistics Series, Vol. II, ed. V. Esposito Vinzi, W.W. Chin, J. Henseler, and H. Wang, 409–425. Berlin: Springer. Anselmsson, J., and N. Bondesson. 2015. Brand value chain in practise; the relationship between mindset and market performance metrics: A study of the Swedish market for FMCG. Journal of Retailing and Consumer Services 25:58–70. Avis, M., S. Forbes, and S. Ferguson. 2014. The brand personality of rocks. Marketing Theory 14(4):451– 475. Babi´ c Rosario, A., K. De Valck, and F. Sotgiu. 2020. Conceptualizing the electronic word-of-mouth process: What we know and need to know about eWOM creation, exposure, and evaluation. Journal of the Academy of Marketing Science 48(3):422–448. Baehre, S., M. O’Dwyer, L. O’Malley, and V.M. Story. 2022. Customer mindset metrics: A systematic evaluation of the net promoter score (NPS) vs. alternative calculation methods. Journal of Business Research 149:353–362. K