2.1. The epistemic value of RecSys as knowledge systems for consumers’ food strategy decision

Given the multi-faceted expression of RecSys, their definition is dynamic reflecting platforms, users, developers activities, and processes (including evaluation standards) in responding to ever-acute needs to overcome information overload (Tran et al. Citation2018; Park et al. Citation2012; Roy and Dutta Citation2022). At the broadest level, RecSys are noted to offer information that is relevant to facilitate decision-making and provide satisfaction to users (Tran et al. Citation2021). Two main types of algorithms, collaborative filtering (CF) and content-based filtering (CBF) are leveraging diffusion of innovation and simulation theories along consumer choice theory to provide estimated rating for products and services (Alyari and Navimipour Citation2018). CF leverages users’ previous history and preferences. These algorithms are mainly based on finding out users with similar taste or type of products (e.g. songs that are fast, slow, etc.). They leverage what users have tagged and rated along viewing/listening/purchasing/browsing history. Three main limitations are noted in the literature namely sparsity (when not enough data points are available), scalability (when a large amount of training data are required), and privacy (associated with the gathering and processing of personal data). On the other hand, CBF relies on the rating of items by users, producers or/and policy makers to infer a profile that is then used to recommend new or additional items. Filtering is traditionally based on content categorisation to recommend products most similar to original items, product rating (e.g. in our case nutritional quality, additives) and product availability. CBF approaches are based on objective information on items which are automatically or manually extracted from different sources. CBF core problem lies in overspecialised recommendation patterns.

Other approaches have been developed, including data mining techniques, demographic RecSys, context-aware RecSys, and knowledge-based RecSys (KBRS) helping to improve the overall performance of RSs through the production of personalised recommendations. Nevertheless, the literature notes that algorithms’ accuracy of prediction is still not adequate for many frequently non-commercial activities ranging from social innovation-based policies, welfare, group decisions, to risk administration in public projects (e.g. education-learning design, medical care, smart cities). Algorithms are not adapted to derive policy consensus when the required pragmatism of contextual situations associated to multiple possible optimisation behaviours is present (Yago, Clemente, and Rodriguez Citation2018; Zhitomirsky-Geffet and Zadok Citation2018). As such, shortcomings are underlined where it may be hard to provide recommendations when users have unusual requirements, when they are not buying for themselves, when preferences evolve rapidly, or when they cannot test proposed items (Alyari and Navimipour Citation2018). These shortcoming challenge RecSys’ developers on how to engineer the next generation of systems whose value is recognised over time by RecSys’ users as consumers. This encourages to think deeper on how RecSys can rely on their users as reflective agents who are not trapped by technologies (Seaver Citation2019; Choi et al. Citation2014; Sun et al. Citation2015).

In food RecSys, a wide range of solutions is available (Pecune, Callebert, and Marsella Citation2022). Health aware food RecSys consider the increasing set of ailments and lifestyles issues faced by consumers to offer recommendation ranging from improved recipes, adaptive diets, cooking methods, best habits, labelling to food waste reduction (Lawo et al. Citation2021). CF systems traditionally recommend healthier food based on similarities from past items, while CBF recommend items that have been identified beforehand through a variety of filters by users, or care providers. Many solutions are proposed to change consumers’ behaviour, when foodtech is used. These include ‘food swaps’ in which a specific product is chosen in a system that retrieves products that are similar but healthier (Jansen, van Kleef, and Van Loo Citation2021) and more generally ‘hybrid systems’ that aim to achieve a balance between what users want to eat and what they need to eat to stay healthy (Tran et al. Citation2018; Starke, Asotic, and Trattner Citation2021). For users, the value of these solutions is not only based on the recommendations provided but also on the potential amount of valuable knowledge encapsulated in the discretionary yet socially shaped usage of food RecSys.

As such, two main clusters have been identified namely building complex information models and nutritional information processing (Toledo, Alzahrani, and Martinez Citation2019). The first cluster relies heavily on both dynamic information voluntarily provided by users along context and knowledge filters that are increasingly available through smartphone and IoT objects. While both the technological measurement (e.g. continuous vs. discrete data) along privacy and ethical-related issues are witnessing progress, it is noted that the amount of possible data available is also growing exponentially, making recommendations challenging (Min, Jiang, and Jain Citation2019). The second cluster leverages principally secondary data already available by law on food products investigating the social nature of RecSys. Food recommendation is approached from an optimisation perspective with new variants including both genetic algorithms, colony optimisation, and bacterial foraging (Syahputra et al. Citation2017; Rehman et al. Citation2017; Hernández-Ocaña et al. Citation2018). Food analysis includes effective representation of single singularities (product per product), the multimodality and the heterogeneity of food data (Yang et al. Citation2017; Bossard, Guillaumin, and Van Gool Citation2014; Cordeiro et al. Citation2015). The general architecture of food RecSys generally encompasses four main areas: nutritional information awareness, preference, semantic-based and optimisation-based models (see Toledo, Alzahrani, and Martinez Citation2019 Table 1). It is underlined by Trattner and Elsweiler (Citation2017) that RecSys research challenges are thus related to both the collection, compilation and modelling of more objective nutritional information and users’ preference data but most of all, the abilities of RecSys to track and change collective eating behaviour in the long run. This leads to question more deeply the role played by food RecSys in consumers’ daily lives, which in essence questions the perceived value of food RecSys’ recommendation.

Indeed, over the year, a third cluster investigating the social function of RecSys is slowly gaining ground. Related works aim at helping users to change, nudging them to adopt better diets to manage their health (Tran et al. Citation2018). Indeed, while it is noted that numerous food-related RecSys are recognised to be effective, they have to find ways to integrate the high level of variability often present within a given population. As technology evolves, new agile modelling to implement such tailored intervention needs to be designed, going beyond legal information accuracy and regulatory compliance (Liang Citation2019; Seo et al. Citation2017). In order to promote well-being whether through private initiatives or public policy, the gain of new knowledge via Food RecSys, including in academic literature, needs to be supported and to draw from greater empirical evidence by learning from active users of such technologies (Prost et al. Citation2019). It is relevant here to approach food RecSys based on their social function, considering that the recommendations and the attempt they make to change behaviours of users as consumers are significant in the long term towards a better society. The emerging normative social function of food RecSys is the point of departure of our study. We underline the relevance of blending social norms and RecSys’ potential to help users as consumers attain their food choice targets. This goes along with the fact that even though food RecSys are valuable, they have currently limited adaptability to users’ broader goals (Trattner and Jannach Citation2020; Trattner and Elsweiler Citation2017). Food RecSys, beyond profiler, filter, and ranker steps, ought to be more driven by collective values about food choice within rapidly shifting users’ preferences (Ekstrand and Willemsen Citation2016; Elsweiler, Trattner, and Harvey Citation2017). As such, Yuka exemplifies the increasing existence in RecSys of a negotiation protocol which is underlined as closer to users’ daily reality, encouraging to reveal constraints while allowing for proactive feedback, thus improving system transparency and innovation (Buzcu et al. Citation2022). We are thus drawing from a case that illustrates food RecSys as closer to a nutritional virtual coaching system (NVC) and social Internet of Things (Atzori et al. Citation2012).

Indeed, food RecSys are found to assist decisions for a variety of consumers’ segments from allergies sufferers, elderly, sport competitors to individuals who are just willing to buy the right thing in retail (Lawo et al. Citation2021; Tran et al. Citation2021; Gunawardena and Sarathchandra Citation2020). In such conditions, exploring the epistemic value of RecSys ought to provide potential answers to the problem of scalability when considering long-term usage and needs for further external (e.g. social) validation. This paradigm is influenced by the social construction of technology that underlines, in our case, that one key value of RecSys is how the acquisition of knowledge, by considering the effect of social structures that form around them, imposes novel epistemic responsibilities. Recommendations are indeed generated automatically, which raises questions on where/how to correct mistakes, how to know if these corrections are considered relevant, and on how to deal with the question of representativeness in addressing the issue of amplificative effect for example (Miller and Record Citation2017).

Connecting with the above, in consumer research, epistemic value is defined as a value that stems from a product or service that arouses curiosity, provides novelty, and/or satisfies a desire for knowledge (Greco and De Sa Citation2018). It is considered as a form of consumption value among other values, such as functional, social, or emotional values (Wang et al. Citation2018; Whelan and Clohessy Citation2021). The development of RecSys for knowledge management capacity necessitates widening the concept of epistemic value to capture the ways RecSys’ technology consumption drivers include (a) the effects of long-term reciprocal value-laden exchanges between RecSys and consumers (Itani, Kassar, and Loureiro Citation2019), (b) the impact of such consumption on shopping strategy both in the short and long term (see Zwick and Dholakia Citation2006) and (c) the potential mediating role of RecSys’ developers, retailers, and policy maker.

2.2. Food RecSys epistemic value through the lens of perceived value

RecSys epistemic value is considered as driving enablers and processes (as sufficiently intelligent programmes acting along algorithms) through which collective knowledge can be leveraged by multiple market actors (Yago, Clemente, and Rodriguez Citation2018). In this way, RecSys facilitate food choice resilience and sustainability, but they also improve innovative behaviours, and responsiveness to political and environmental changes. For retailers, manufacturers, and foodtech, a refined understanding of RecSys’ epistemic value will lead to a better appraisal of the different forms of value by ‘truly supporting consumers throughout their day and serving as a trusted companion in capacities that are still being imagined’ (Wang, Liao, and Yang Citation2013, 19). Indeed, food RecSys are creating social interactive engagement and a technologised experience with consumers while shopping; they lead to complex valuation negotiations and tweaking of value perception when facing multiple options that are often part of a basket purchase. At stake here is the possibility to endow RecSys with capabilities to shape a legitimate strategy-oriented view through technology-mediated epistemic value creation processes (Miller and Record Citation2017).

Value (in the singular) characterises the consequence of an evaluative decision, and values (in the plural) are described as ‘the determinants of any social behavior including attitude, ideology, beliefs and justifications’ (Boksberger and Melsen Citation2011, 230; see also Rokeach Citation1973). Perceived value as a multi-faceted concept is traditionally analysed using three main approaches: (1) value/cost trade-offs (i.e. value proposition and value network in a market), (2) design-driven value/innovation change where it ‘aims [to] radically change the emotional and symbolic content of products (i.e. their meanings and languages)’ (Verganti Citation2008, 436), and (3) user-driven, following customer-centric models of innovation diffusion including co-creation, co-production, and interactive value creation (Rivière and Mencarelli Citation2012; Sweeney and Soutar Citation2001). The third approach encourages an understanding of the way consumers perceive food RecSys epistemic value within service encounters.

At this point, the dynamics that underlie consumers, RecSys’ developers, retailers, and policy makers’ embeddedness in socio-technical socialities need to be better apprehended by identifying the drivers of perceived value for digitally enabled consumers (Seaborn et al. Citation2021). Yet, research analysing either acceptability of or resistance to engagement with RecSys is noted as not ‘sufficiently deal with the antecedents of being [socially] engaged in an overall service’ (de Kervenoael et al. Citation2020, 3); thus, it disregards the consequences related to the fact that engagement considering perceived value is more than just related to specific performance. With so many inflections, beyond their technological dimensions, food RecSys perceived value represents users’ internal negotiations and acknowledges users’ learning processes within the shopping contexts in which they are embedded (Yago, Clemente, and Rodriguez Citation2018; Dascalu et al. Citation2016). Perceived value in service has become a reflective instrument to recognise and protect oneself from the technologisation of society (Foucault Citation1988). Retailers and food manufacturers should inform a greater part of that discussion. Food RecSys encapsulate the implicit but often problematic relationship between individuals, regulatory requirements, and technology in an era of increasing questioning of the role and place of technology in society (Van Belleghem Citation2015). In a networked, co-produced information ecosystem (reflecting the experiences of choice and shopping by users of RecSys) creates new spaces leading, or not, to retail strategy acceptance or resistance/rejection (Kozinets Citation2019).

3.1. Compatibility

‘Incompatibility with values poses a factor that can hinder an innovation’s diffusion' (Rogers Citation1995, 224). Compatibility represents a psychological dimension that originates in DoI theory recognised as important in contexts such as e-book or IoT objects but more importantly regarding food policy in issues related to health and diet (Pecune, Callebert, and Marsella Citation2022). Compatibility is defined as ‘the degree to which an innovation is perceived as being consistent with existing values, needs, and past experiences of potential adopters’ (Rogers Citation1995, 224). As such, it is important to consider how consumers’ perceptions of compatibility with habitual shopping behaviour that reflect current food choices (e.g. Nutriscore and carbone scores in France) affect the epistemic value of food RecSys. Furthermore, manipulating a smartphone within the grocery retail environment may not be practical, and may require a new approach (e.g. a smartphone stand on the trolley, that may be imposed by policy makers) as it constitutes a whole different shopping experience. It is thus important to emphasise that the disciplinary potentials exerted by RecSys are more likely to prevail when the latter are ‘compatible’ with consumers expected behaviour and by extension retail connected strategies. Thus, hypothesis 1 is proposed:

3.2. Memory vividness

Qualitative and quantitative information displayed on RSs, such as ratings, reviews, or, in the case of Yuka, photographs of product packaging represent precision and bias that relate to subjective memory vividness (Wilmer, Sherman, and Chein Citation2017). These salient factors are often overlooked in the diffusion of innovation theory and RecSys literature. They have clear policy implications, which requires intentional design that reflects both multitasking usage and for example, the needs of vulnerable groups all affecting the recollection of recommendations (Uncapher, Thieu, and Wagner Citation2016). Memory vividness represents an emotional experience’s intensity, which may cause negative memories to often be recalled more easily than positive ones (Richter et al. Citation2016). Memory biases in remembering precise information in subsequent tasks are also linked to memory vividness (Cooper, Kensinger, and Ritchey Citation2019). Therefore, considering that specific RecSys’ atmospherics require multitasking and their overall settings moderate memory vividness, it appears that memory vividness is part of a consumer’s successful problematisation related to RecSys’ epistemic value. As such, it is expected to drive perceived value. Memory is indeed often taken for granted in terms of food choice impact (Higgs Citation2008). Hypothesis 2 is thus formulated as:

3.3. Self-confidence when considering recommendation

Self-confidence is broadly defined as the ability to meet targets, face challenges, and achieve complex tasks (Erasmus, Donoghue, and Fletcher Citation2015). For consumers, self-confidence as an assured orientation beyond RecSys accuracy is essentially tied to positive enthusiasm about having an impact on successful achievement. Self-confidence is important for individuals when choosing to proceed before all the requirements and outcomes of an action are fully known, thus implying a willingness to take risks. There are two possible valences related to self-confidence: specific and general (Rosenberg Citation2015). Specific self-confidence occurs when a consumer has already experienced a service or product, and it influences how they continue using similar specific services. General self-confidence is broader and represents a generic experience of technology, food choice debate, or, in our case, smartphones usage for food choice (Chuang et al. Citation2013; Henke, Joeckel, and Dogruel Citation2018). In this study, specific self-confidence was tested because all the respondents have self-declared their regular use of the Yuka RecSy. Arguably, self-confidence when considering a recommendation ought to have a disciplinary effect that generates greater perceived value (D’Souza et al. Citation2021). Hence, we propose hypothesis 3:

3.4. Personal innovativeness

Rogers (Citation1995) defined personal innovativeness as ‘the willingness of an individual to try out any new information technology’ (206). It encapsulates the forces required to advance experimental and pioneering behaviours (Roehrich Citation2004). When consuming food RSs’ recommendations, personal innovativeness has been recognised as influential (Li et al. Citation2021; Payini et al. Citation2019; Pecune, Callebert, and Marsella Citation2022). It has been considered mainly based on prior experiences and changes in function of frequency and usage, along with regulatory communication leading to a predisposition to adopt and innovate (Agarwal and Prasad Citation1998; Yi, Fiedler, and Park Citation2006). It reflects complex innate characteristics and facets of individual identity that are disciplinary in their effects. Indeed, most consumers may (perhaps naively) assume that food RecSys respect the law as well as the latest recommendation from policy makers. Contrary to Ajzen’s (Citation1991) theory of planned behaviour whereby innovative individuals display more positive attitudes towards change, we anticipate that personal innovativeness is negatively associated with perceived value because many RecSys are risk-free (e.g. freemium model), do not require specific IT skills, and are easily accessible. Therefore, a significant but negative relationship between personal innovativeness and perceived value that reflects the broad market targeted by this type of audience i.e. internet native, post-1990s population. This led to hypothesis 4:

3.5. Learning

Because of the need for recommendation processing, learning shapes and changes a consumer’s perception of the choices offered (Pecune, Callebert, and Marsella Citation2020; Behera et al. Citation2019). Learning is noted as an important dimension in the consumer journey from pre-purchasing, purchasing (the ‘zero moment of truth’), to long-term loyalty and repurchasing. Learning, as information gathering and reflection, is thus a standard part of the shopping process and has led to detailed literature in optimal searches (Doval Citation2018; Renner et al. Citation2020). In this context, RecSys provide an interactive and vivid environment that encourages information acquisition and comparison (Yago, Clemente, and Rodriguez Citation2018). In food-related markets, because of the abundance of external information on numerous products available globally, learning is required to improve adoption and moderate negative experiences (Gupta Citation2019). As such, manufacturers and retailers should influence minimum standards in the learning process offered by RecSys to approach holistically how the interlinkages between food RSs, food production, distribution and consumption contribute (or not) to a better food system approach (Tran et al. Citation2018; Rockström et al. Citation2020). Arguably, learning ought to be a driver of RecSys’ users problematisation when relying on recommendation technology, which would enhance their perceived value. Thus hypothesis 5 proposes:

4.1. Participants and survey procedures

This study used a convenience, non-probability sampling technique drawing from that part of the population that has self-revealed interests in the subject (Iacobucci and Churchill Citation2009). Respondents had to have been using the RecSys Yuka in France for at least 6 months. The sampling strategy allowed for varied profiles along key dimensions, including size of household, income, employment, and age (self-declared by the respondents). To access users, we posted advertising asking for volunteers on different cooking communities: Cookeo recettes entre nous & vous (>72 K members), Facebook group on food and nutrition Regime & Santé & nutrition, (>100 K members), on Instagram Foodwatch.fr (>7 K members) and Herbalife nutrition (>1.5 M members). Administrators of each community were contacted to explain the non-commercial aspect of the project so that a message could be placed on the community platform. A link to a Google doc survey was made available in the message summarising our study goals and our institution ethical guidelines guaranteeing anonymity. We received 270 responses within the 2-week period the survey was open. Of those, 17 incomplete responses failed the attention check and were disregarded. The final sample size was thus 253 responses; this number exceeds the recommended level for structural equation modelling (Hair, Ringle, and Sarstedt Citation2013). We used G*Power software to calculate the minimum required sample size for the proposed model, which was n = 92 (statistical power: 0.80; Faul et al. Citation2009). The first section of the survey was designed to collect generic demographic and contextual information (see Table 1).

From information available publicly, in France the balance of Yuka application users is 40% male, 60% female and 18% of the French population have downloaded the app (Huot Citation2019). Yuka was ranked the 6th most popular app in France in 2018, with around 3 million barcodes scanned in the app every day. We found that in our sample 90% of respondents were female underlying a potential gap in the literature in food RecSys gender divide. As such, recent studies have noted that the original gender gap in technology and digital engagement is reducing in developed economies. It is underlined that men and women choice/motivation/time spent and utilisation of technologies varies (Tanczer et al. Citation2018). Examples of these variations include distinctions in social media participation, influences by subjective norms, and risk positions. The consideration of these differences allows to reflect on the often taken for granted gender (in)equalities while shopping and cooking at home (Taillie Citation2018). Indeed, this leads to ponder the often-polarised academic debate connecting technology usage to gender issues toward developing more inclusive technology (Goswami and Dutta Citation2015; Fortunati and Edwards Citation2022; Albert et al. Citation2019). While there is evidence that gender aspects are rarely central to the design and usage of technology systems (Wolfson et al. Citation2021), our sample allows to counteract technology determinism and alleged gender-neutral use of technology towards reducing misleading managerial recommendations. The second section included questions related to how respondents feel about the different constructs (Table 2).

4.2. Measures

A five-point Likert scale was used to measure the constructs. We adapted most of the items of this study from well-established sources (see Table 2). Items for compatibility and perceived value were adapted from Wang et al. (Citation2018), items for memory vividness were adapted from Wilmer, Sherman, and Chein (Citation2017), items for self-confidence when considering recommendations were adapted from Erasmus, Donoghue, and Fletcher (Citation2015), items for learning were derived from Froehle and Roth (Citation2004), and items for consumer innovativeness were adapted from Zhang, Lu, and Kizildag (Citation2017).

4.3. Statistical analysis

Partial Least Squares-based (PLS-based) Structural Equation Modelling (SEM) was used to analyse the structural model. PLS-SEM was chosen because the phenomenon analysed was at an early stage of development (Fornell and Bookstein Citation1982), and its elevated statistical power (vs. covariance-based SEM; Hair et al. Citation2017). PLS is also appropriate in exploratory studies because it does not need a multivariate normal distribution (Albert and Merunka Citation2013).

5.1. Measurement model

Assessments were run using Smart PLS 3.0 to test for each construct’s convergent validity, discriminant validity, and reliability. To assess the convergent validity, a confirmatory factor analysis (CFA) was conducted to ensure the reliability and validity of the model. Table 3 shows that almost all items loaded properly within their theoretical constructs. Each construct’s Cronbach's alpha and composite reliability was assessed using PLS-SEM (Lowry and Gaskin Citation2014). As shown in Table 3, each construct except for two items, CP_1 and SC_1, loaded to the required threshold level of 0.70 (Chin Citation1998).

Table 4 shows the discriminant validity assessment of the measurement model, in which the square roots of average variance extracted (AVE) are displayed by the numbers on the diagonal, and the interconstruct correlations are below them. Table 4 demonstrates appropriate discriminant validity because the square roots of AVE are greater than the interconstruct correlations (Lowry and Gaskin Citation2014).

5.2. Structural model and analysis

The structural model was tested by using a bootstrapping sample of 5000. The results of the test are displayed in Figure 2, which shows that the strength of the relationship between dependent and independent constructs is measured by path coefficients with R-squared values that demonstrate how much the variance is explained by independent constructs.

Figure 2. Structural model and analysis: standardised path coefficients.

Display full size

Figure 2 shows that there are five independent variables contributing to the variance of the perceived value of food RSs. From the figure, compatibility has a significant positive influence on perceived value (β = 0.137, p < 0.05), which supports hypothesis 1. The second construct, memory vividness, is found to have the greatest and significant positive influence on perceived value (β = 0.308, p < 0.05), which supports hypothesis 2. In addition, the influence of memory vividness is found to be moderated by the frequency of usage, suggesting that the influence is greater for consumers who use the RS less often (β = −0.154, p < 0.05). Self-confidence when considering recommendation is found to have a significant positive influence on perceived value (ß = 0.155, p < 0.05), which supports hypothesis 3. The impact of consumer innovativeness on perceived value is found to be significant and negative (β = 0.093, p < 0.05), which supports hypothesis 4. Lastly, learning was found to have a significant positive influence on perceived value (β = 0.208, p < 0.05), which supports hypothesis 5.

Overall, Figure 2 shows that memory vividness has the most significant positive influence on perceived value. Frequency of usage was found as a significant moderator to memory vividness. This is followed by learning, self-confidence when considering recommendations, compatibility, and personal innovativeness. It can also be seen that the R-square value of perceived value is 64.60%, which means that 64.60% of the variance of perceived value is explained by the five constructs. Finally, the proposed model’s overall goodness-of-model fit was assessed by evaluating the standardised root mean square residual (SRMR), unweighted least squares (ULS) discrepancy (dULS), and geodesic discrepancy (dG) (see Table 5) (Henseler et al. Citation2014). The value of SRMR is 0.0384, which is less than the recommended threshold value of 0.08 (Benitez et al. Citation2019). All the discrepancies in Table 5 were less than the 95% quantile of the bootstrap discrepancies. This suggests that our proposed model should not be rejected based on the alpha level of 0.05, which provides a good explanation of the key antecedents leading to the perceived value of using the Yuka RS with a probability of 5% (Benitez et al. Citation2019).

6.1. Problematising drivers

Problematisation prevails in the epistemic usage of RecSys. It is influencing what is acceptable usage within, in our case, food choice and probably healthier food choice (de Kervenoael et al. Citation2021). From this perspective, the problematising driver of learning illustrates not only the knowledge management derived (or not) from the recommendations provided by RecSys but also the posture and gaze adopted when using food RecSys. Memory vividness and learning relate to the manifestation of critical consumption in everyday life when using Yuka as a RecSys. These variables reveal that consumers mobilise RecSys to reflect on social trends toward informed decisions (Lawo et al. Citation2021). Attention to official positions but also scientific experts and other spheres of influence are considered in the final score, allowing a proactive approach to risk (e.g. controversial additives, interest in legislation). For example, overall food RecSys usage is noted to have an impact on shopping flow as they are used by over 55% of the French population (Huot Citation2019; Pizay and de Guenyveau Citation2020) but, more importantly, on the directionality and reflexivity supporting or not by RecSys’ developers, retailers, and policy makers’ strategy. As a matter of fact, the retail supermarket Intermarché changed 900 receipts for their own brand by removing 142 potentially harmful additives (Yago, Clemente, and Rodriguez Citation2018; Brancato Citation2019). In short, consumers who use food RecSys create new behaviours including blockages in the supermarket aisles because interactions with RecSys are a nonplanned activity within offline shopping environment (e.g. stopping, constantly having an object in one hand, retracing steps to find specific items, etc.) but more importantly problematise in real time what food they are ready to buy. Other side effects may include rejecting initially chosen articles when accepting RSs’ recommendations, which creates a clutter of discarded items across stores that may also have policy impacts in term of cold chain regulation, forcing once again retailer to re-think their strategy in regard to food RecSys as a long-term tendency.

In this regard, the normalisation that is occurring through regular usage is captured by memory, which reflects the social perception that everybody while knowing how to use RecSys need help to make better food choices. This has implications on RecSys’ perceived value, and future developments that should go beyond accuracy features towards including their social function which has the power to contribute to the establishment of more durable forms of RecSys. RecSys’ developers ought not to simply consider epistemic value as exogenous to the usage of RecSys in the determination of algorithm but rather as a variable that is endogenous and ought to serve as a constitutive element of RecSys’ design. RecSys’ epistemic value yields the conditions for the overcoming of RecSys current limitations. Here, RecSys’ perceived value is portrayed as an essential element and an ineradicable counterpoint to filtering systems accuracy. RecSys ought to be considered as actants that have an impact in building future learning partners whereby norms, culture, and interpersonal goals such as building good relationships are incorporated in the next generation designs (Pecune, Callebert, and Marsella Citation2022; Jugovac and Jannach Citation2017; Rese et al. Citation2017). More pragmatically, without importance given to the problematising drivers that concur to food RecSys’ epistemic value, users as consumers may not only become targets of curiosity but also perceived as undesirable, slow, problematic, or disruptive.

6.2. Disciplinary drivers

Given the previous discussion, further investigation needs to be carried out to answer questions such as ‘who are the direct and indirect stakeholders affected by the design at hand?’ and ‘what values are implicated?’ (Friedman, Kahn, and Borning Citation2006, 72). Personal innovativeness, compatibility, along with the self-confidence to consider information in real-time underline key aspects of the dilemma consumers face when using the Yuka RecSy. Being perceived as caring for what is in the food one buys represents how consumers positively respond to issues such as global food safety and resources, environmental concerns, or health concerns (Havinga and Verbruggen Citation2017; Taylor et al. Citation2019). In today’s sociocultural context, self-determination, investigation into alternative consumption and production model possibilities, sustainability, and resistance are becoming central to consumers’ identities. Through closer attention paid to RecSys’ social function, these aspects can be better integrated into retail strategy (Castellini, Savarese, and Graffigna Citation2021). With increasing control, consumers are re-embedding traditional retail offerings with reflective RecSys-related behaviours as reinforcing agents that confront normative shopping conventions by bypassing retailers’ promotions and convenience offers (Castellini, Savarese, and Graffigna Citation2021; Renner et al. Citation2020; Vasconcelos et al. Citation2021). As such, our framing of perceived value allows policy makers to send retailers, producers, and consumers signals by co-analysing, with RecSys developers, decision-making patterns in a proactive manner. Perceived value of food RecSys, such as Yuka, is thus associated with better access to data as resources that were once kept hidden. Exhaustion over food crises (e.g. mad cow disease, bird flu, salmonella, food waste, product safety, contamination, coronavirus, etc.), the unpredictability of ingredient changes (e.g. ready meals and cosmetic), and the intractability of the global production chain have led to uncertainties that need to be better addressed by accessing the power of big data not only in supply chain (Rejeb, Rejeb, and Zailani Citation2021) but also to support consumers as responsible individuals (Marvin et al. Citation2017).

7.1. Theoretical contributions

Today, RecSys’ developers, retailers, and policy makers need to be more transparent in justifying their choices in listing, qualifying or authorising food items (Forster Citation2017; Marvin et al. Citation2017). Beyond functional priorities, such as safety, declining resources, worker rights, or shelf-life, they must consider how, when mediated by technology, consumption choices are operationalised in real-time (de Kervenoael et al. Citation2020). They must recognise and better understand that their own and their stakeholders’ responsibilities (including users) are crystallised in the attitude towards and consumption of food RecSys technology. Considering food RecSys’ perceived value in particular, health as a legitimate concern is reverberated. This implies that beyond the accuracy or the relevancy of predictions of algorithms, more and more sophisticated food RecSys make it possible to understand social and societal implications of technologies that have a great appeal to an especially large set of actors (e.g. consumers, RecSys providers, retailers, media, policy makers) (Sumner Citation2016; Siegrist and Hartmann Citation2020; Vasconcelos et al. Citation2021). Not only food RecSys provide a large amount of data (easily linked to key strategic goals and operational efficiency), but while relying on shared criteria and more subjective preferences, they also facilitate the understanding of the elements that drive behavioural change testifying the social function of RecSys. As a result, food RecSys have become highly valued by users. Consumers benefit indeed from personalised and improved communication relying proactively on informational features and concrete components such as push notifications (Behera et al. Citation2019).

In this study, through the case of Yuka, we further address the ways consumers make sense of their daily usage of RecSys by disentangling the drivers of their perceived value. Our first theoretical contribution is to show that RecSys’ perceived value is convincingly captured by consumers’ repeated usage leading to socialised epistemic value of RecSys. For algorithms and their developers, the problematisation and disciplinary drivers that are found to constitute this value imply to understand better how queries become the ingredient for a questioned yet socially supported recommendation. RecSys’ problematisation and disciplinary drivers need to be recognised as central and increasingly unavoidable aspects of consumers’ practices. They are found to constitute the underlying premise of RecSys scalability issues whereby daily usage is required to make recommendations viable (leading to loyalty to RecSys as learning partners). As supporting RecSys, algorithms are defined both by what they are and are not (yet); they question the RecSys epistemic responsibility in how and when to address long-term consumers relationships management. We show that the value of RecSys is mainly based on socialised affects that consumers act upon by leveraging their memory, learning, compatibility, self-confidence, and innovativeness. These insights shed light on the importance of consumers’ a posteriori thinking, i.e. subsequent to RecSys usage. While going beyond their prediction accuracy or dominant algorithm logics (e.g. CF, CBF), we add to the so far largely unquestioned RecSys current usage leading to understand how users consume RecSys as technological objects and what the social function of such consumption is about (Competition Market Authorities Citation2022). This leads to reflections on how to appraise RecSys not solely through their effectiveness but also through the social processes through which users accomplish their agenda. In doing so, this study outlines relevant evidence on RecSys’ practices and a research agenda that dialogues with the ‘dark side’ usage of recommendations, including coercing, steering or deceiving users into making likely detrimental decisions (Mathur, Kshirsagar, and Mayer Citation2021). Change in consumer welfare that results from dark recommendation patterns, including RecSys’ lack of capacity to correct mistakes, the likely amplificative effect in terms of representativeness, or RecSys’ impact in reinforcing and spreading negative habits ought to also take into account the cognitive burden hence the epistemic responsibilities attached to RecSys (Miller and Record Citation2017),

Second, the results of this study show the importance of considering RecSys as multi-faceted beyond the privacy issue of data collection and private firm intents that have led many to describe algorithmic filtering as a Skinner box (Leslie Citation2016). As such, food policy makers, in our case, may have a reasonable cause to start intervening via regulation to counterbalance such engineering (Thorseng and Grisot Citation2017). The real time usage of RecSys along literature on general social acceleration needs also to be better appreciated to tackle RecSys intentional goal and directed processes (Colin, Shruthi, and Ahreum Citation2020; Mathur, Kshirsagar, and Mayer Citation2021).

Third, we contribute to the critical debate on how the usage and attitude toward technology depend on gender effects. Rather than concentrating on a holistic set of users in which gender issues are amalgamated when analysing technology (Tanczer et al. Citation2018), this study with a sample of 90% women provides evidence on the significant role of women in understanding and impacting RecSys effects and consumption. This encourages RecSys developers to focus on gender differences when developing algorithms and to avoid being trapped in over generic behavioural constructs.

7.2. Managerial implications

Worldwide, food manufacturers, retailers, and RecSys’ developers need to be more transparent in justifying the choices they make in authorising or listing food items (Macready et al. Citation2020). The long-term value of RecSys, such as Yuka, lies indeed in information beyond what is already displayed in stores or on products (Behera et al. Citation2019). Dieting RecSys, for example, often exacerbate a culture of tracking, controlling, and negative micro-management in the long term. Although the educational opportunities and prospects to connect like-minded communities are clear (confirmed by the significance of the learning driver in our model), manipulating food intake by following hypes and trends can lead to more eating disorders. This reflects the difference between quality information and propaganda (Lupton Citation2018). It is a key challenge to encourage common sense eating despite ‘normal food’ (e.g. butter) appearing as red on most food RecSys list (Herz Citation2017). Food RecSys value thus represents attributes that include the need to support a long-term positive relationship with shopping and eating food (Tran et al. Citation2018; Thorseng and Grisot Citation2017). As seen in our model, positive memory of a relationship with food will need to be further explored. Finally, we note that a traditional food manufacturer sued Yuka, and they were condemned for disloyal commercial practices, deception, and acts of denigration over the classification of nitrite additive (Ouest France Citation2021). This has critical implications on the way the different RecSys’ developers, retailers, and regulators ought to interact and to decide about issues such as the authorisation to include (or not) brands in rankings, approval (or not) of rating, scope, and rules to sanction abuses.

7.3. Limitations and future research

In assessing this study’s results, it is important to point out certain limitations. A first limitation, familiar to many surveys, was the number of dimensions respondents could answer in the time they had to complete the survey. Second, the study is based on a one-off study, and it should be advanced by others over time to see the evolution of the food RecSys phenomenon while also considering aspects that are country specific (Choi et al. Citation2014; de Edelenyi et al. Citation2019). Further studies could explore in more detail the correlation between perceived value and particular aspects of retail strategy, including certification and quality insurance. A focus on specific groups (e.g. first-time parents, allergy sufferers, consumers with specific health conditions, such as diabetes or high cholesterol) could provide further insights into food RecSys’ perceived value (Depper and Howe Citation2017; Taylor et al. Citation2019). Disciplinary and problematising drivers should be further elucidated by including new dimensions, such as capacity building, playfulness, or excessive use, potentially leading to delineate epistemic value against what is called choice paralysis. In doing so, there would be additional opportunities to understand the boundaries between disciplinary and problematising drivers of food RecSys perceived value and to integrate the underlying mechanisms of these drivers to understand better-evolving consumer food choices.