Systematic searches were conducted on January 19, 2024, in Web of Science core collection, Embase, and Scopus, with an updated search run on September 23, 2025, to retrieve newly published papers. An experienced academic librarian was consulted to optimize the search strategy and to ensure that database-specific operators were used correctly. The final search string was limited to peer-reviewed papers published in English and contained keywords related to digital media, interventions, and healthy and sustainable eating behavior (Multimedia Appendix 1). No date restrictions or other published search filters were applied. The reference lists of all included papers were manually searched, and a citation search using Web of Science was conducted. Authors of all included studies were contacted for additional relevant papers. No other online resources or study registries were searched.

Studies were included if they met the following PICO-based criteria. Reviews, dissertations, books, and conference papers were excluded.

KV and BS independently extracted relevant data using the preregistered extraction book, covering study characteristics (eg, authors, year of publication, and country of origin), study design (eg, comparator and research question), population (eg, age, gender, and SES indicators), intervention characteristics (eg, digital medium, duration, BCTs, and approach), outcomes (eg, behavioral outcomes, metrics, and covariates), and results (eg, analysis technique and statistical results). The 3 reviewers collaborated closely to resolve any discrepancies during the data extraction process.

BCTs that were not explicitly reported in the included studies were coded by 2 reviewers (KV and BS) using the Behavior Change Technique Taxonomy version 1 (BCTTv1) [24]. One reviewer (KV) coded all studies, and a second reviewer (BS) randomly double-coded 5, achieving 90% agreement. Both reviewers were certified coders, as they completed the online taxonomy training (“BCTTv1 Online Training”). Discrepancies were resolved via discussion between the 2 reviewers. For studies with multiple intervention groups (IG), BCTs were coded separately for each active arm. When insufficient detail was provided to code the BCTs, related documents (eg, protocols) were consulted.

Included studies were appraised for study quality using the Cochrane risk-of-bias tool [44]. For RCTs, the Revised RoB 2 tool was used; for cluster RCTs, an adapted version of RoB 2; and for nonrandomized studies (NRS), ROBINS-I [45-47]. The risk of bias assessment was conducted with consideration of best practices in food and communication research. While self-reported measures typically carry a moderate risk of bias, this was not considered problematic, as self-reporting is often the most appropriate or even the only feasible method for measuring eating behavior.

Risk of bias assessments were independently conducted by 2 reviewers (KV and BS), and discrepancies were resolved through consultation with a third reviewer (TS). One paper included 2 distinct experimental studies [48], resulting in a total of 53 separate studies evaluated for risk of bias. The majority of RCTs, including all 3 cluster RCTs, had a moderate risk of bias, primarily due to the use of self-report measures or lack of prespecified analysis plan (31/39, 79%; Figure 1) [49]. In total, 5 RCTs were rated as high risk and 3 as low risk. NRS were rated as moderate (8/14, 57%) or high (6/14, 43%) risk (Figure 2), commonly due to baseline confounding and issues similar to those in the RCTs (see Multimedia Appendix 2 [48,50-99] for a detailed overview of the risk of bias assessments).

To narratively summarize the findings, IGs with differing content were treated as distinct, and identical interventions across studies were grouped [ie, 50-53]. A study was considered to have a consistent effect if it demonstrated a significant improvement in food consumption (ie, within-subjects effect) or if this change was significant compared to a control group (ie, between-subjects effect). For quantitative synthesis, standardized mean differences (Cohen d) were calculated. For pre- and postdesigns, Cohen dav was estimated, which is based on the mean difference and average SD of both sets of observations [100,101]. For studies that included both an intervention and control group, we calculated the effect size based on the mean pre- and postchange in the treatment group minus the mean pre- and postchange in the control group, divided by the pooled pretest SD [102]. This estimation has proven to be the most recommended effect size for repeated-measure designs in terms of bias and precision [103]. The variance of this effect size was calculated in R (R Foundation for Statistical Computing) with equation 25 of Morris [103], assuming a standard pre- to postintervention correlation of r=0.50. As an estimate of ρ (ie, the correlation of effect sizes) is required to develop efficient weights, several sensitivity analyses were conducted with correlation values of r=0.10 and r=0.90, which led to the standard size of r=0.5 [104]. Missing data were requested from authors and obtained for 2 papers.

A random-effects meta-analysis was performed, as we aimed to estimate the mean of a distribution of effects, and we anticipated heterogeneity in true effect sizes across studies due to variations in study characteristics [105]. Positive effect sizes reflect improvements in healthy or sustainable food intake (ie, increased plant-based consumption or reduced meat consumption), whereas negative values reflect declines. Effect sizes were dependent, as studies often contained multiple IGs compared to a common control, multiple types of food outcomes, or multiple follow-ups. Including effect sizes from the same study in a single model creates complications due to statistical dependence, which violates the assumption of independent sampling errors. To account for this dependency, robust variance estimation was performed in R with the packages metafor, meta, and ClubSandwich [106,107]. This method offers a robust solution to handle dependency, even when the nature of the dependence structure is unknown, by grouping effect sizes based on commonalities (ie, hierarchical clustering) and accounting for the correlation of sampling errors [107].

In total, 6 outliers and 1 intervention with inconsistent results were excluded, leaving 41 papers containing 57 interventions, with 82 effect sizes. Given that NRS represent an important part of the evidence base in digital intervention research and the lack of clear consensus on how to best integrate different types of study designs into meta-analyses [108,109], both RCTs and NRS were included in the meta-analyses. To assess the robustness of the findings, we conducted sensitivity analyses excluding all NRS. The modified method of Hartung-Knapp-Sidik-Jonkman was applied for greater accuracy [110]. Heterogeneity was evaluated using the I² statistic and prediction intervals (PIs). Although I² is a commonly reported metric, researchers have noted that it may not be the most informative indicator of heterogeneity; PIs are often preferred, as they reflect the distribution of true effects [111,112]. Funnel plots and Egger test were applied to examine small-study effects; however, the interpretation of funnel plots should be approached with caution due to their known limitations [113]. To identify the conditions under which digital interventions differ in effectiveness, we conducted a series of exploratory moderator (ie, subgroup) analyses. More specifically, subgroup analyses were conducted to examine (1) the stability of effects (ie, postintervention vs follow-up), as well as variations in effectiveness based on (2) behavioral goal orientation, (3) participant age, (4) digital medium, and (5) the presence of each distinct BCT cluster. Each moderator was examined in a separate model to avoid overfitting. Moderator analyses were also conducted for individual BCTs, but results are reported only in Multimedia Appendix 3 due to the limited number of studies per BCT and their similarity to the findings from BCT cluster analyses. More information on the meta-analysis can be found in Multimedia Appendix 4 [50,51,78,96,100-107,114].

The database searches identified 9318 records in total. Following the removal of duplicates through the SR Accelerator software (Bond University) [115], 4765 papers were uploaded to Zotero for eligibility screening. Titles and abstracts were screened independently by the first author (KV) and 2 graduate students, followed by full-text screening. Papers that were commonly included by all screeners received a definitive inclusion. Discrepancies were resolved through discussion with 2 researchers (BS and TS) who had not been involved in the initial screening process. The final database of included papers was checked again by 1 reviewer (BS) to ensure that each paper fulfilled eligibility criteria. The updated search conducted in September 2025 led to the identification of 2283 additional records. After deduplication, 1603 records remained and underwent the same screening procedure, resulting in 6 additional papers being included. This process resulted in 42 papers, with 10 additional studies identified through searching methods, totaling 52 papers. Figure 3 shows the study selection process and provides a more detailed summary of each screening stage, and Figure 4 provides an overview of the updated search.

Of the 52 papers that met the inclusion criteria, 1 paper contained 2 separate experiments [48], while 2 pairs of papers reported findings from the same study conducted among the same sample [54-57]. Therefore, this review contained 51 unique studies, of which most were RCTs (n=38). The included studies were published across a wide range of disciplines, including public health, medical, communication, and psychological journals, reflecting the interdisciplinary nature of research on digital interventions and eating behavior. Most studies were conducted in developed countries, with the majority in the United States (n=10), followed by Italy (n=9), the United Kingdom (n=6), Australia (n=5), and other Western countries such as Denmark (n=2) and Belgium (n=2). Only 4 studies took place in a developing country, specifically in Mexico (n=2), China (n=1), and Saudi Arabia (n=1). All studies were published between 2004 and 2025 and were mostly conducted among adults (n=24), young adults (n=16), or adolescents (n=6). Some focused on children (n=1) or child-parent dyads (n=3), while 1 study did not disclose on population type or age [58]. Sample sizes varied considerably from 28 to 5062 participants at baseline and 24 to 1788 participants after the intervention. Altogether, the studies included 24,652 participants at baseline. More information on the study characteristics can be found in Multimedia Appendix 5 [48,50-99].

Socioeconomic characteristics were inconsistently reported, with only 56% (29/52) of the included studies providing relevant data. The most commonly used indicators were educational level (24/29, 83%), income (8/29, 28%), and occupation or working status (5/29, 17%). One study used food assistance as a proxy for income [54], and 3 incorporated area-based indicators [59-61]. Nearly all studies that included indicators were conducted among adults, and only 2 studies reported socioeconomic characteristics for adolescents. The majority demonstrated a predominance of participants with higher educational attainment, with some reporting up to 95% having tertiary education or International Standard Classification of Education levels 3 to 8. Similarly, among studies reporting income data, the proportion of participants in the lowest income category was small, ranging from 3.5% to 21%, with the exception of 1 study that reported an equal distribution between low- and middle-income groups [62]. While some studies controlled for socioeconomic indicators, few included them as moderators or created subgroups. Only 4 studies explicitly examined the role of SES in intervention effectiveness, of which 3 found significant effects on food intake among lower SES participants [56,57,62]. Additionally, Lim et al [63] found that income correlated with increased legume intake and decreased intention to consume animal-based foods. However, due to the limited and inconsistent reporting of SES as well as the underrepresentation of individuals from lower socioeconomic backgrounds, the planned moderator analysis based on SES was not feasible.

Across the 52 papers, a total of 69 digital interventions (k) were assessed. Intervention duration and intensity varied widely, spanning from 1 week to 6 months. While some were self-paced, most delivered content at intervals ranging from once a week to twice daily. The majority focused on promoting fruit and vegetable intake (FVI; 61%) as an indicator of healthy and sustainable eating behavior. A smaller proportion addressed the reduction of meat consumption (19%) or a broader healthy and sustainable diet (20%; eg, Mediterranean diet). Most interventions used SMS text messaging (41%) or websites (25%), followed by social media (13%), mobile apps (9%), games (7%), and emails (4%).

In total, 53 unique BCTs were identified belonging to 15 of the 16 hierarchical clusters of the BCTTv1. The most frequently used BCT clusters were “goals and planning” (65%), “natural consequences” (64%), “feedback and monitoring” (52%), and “comparison of behavior” (41%). The number of BCTs per intervention ranged from 1 to 21, with an average of 6 (SD 4.5). More information on the prevalence of each BCT can be found in Multimedia Appendix 6.

While the meta-analysis found a moderate and statistically significant overall effect of digital interventions on healthy and sustainable food consumption, the substantial heterogeneity observed across studies suggests that these interventions do not have a consistent effect across all populations and contexts. Additionally, included studies mostly had a moderate risk of bias, and together with the inconsistency in effects, this may reduce the certainty of the evidence regarding the overall effectiveness of digital interventions. Both the narrative review and the moderator analyses showed that the effectiveness of digital interventions varied by multiple characteristics and settings.

Regarding the food outcomes, interventions preventing animal-based food consumption and those promoting plant-based food intake both yielded a significant, moderate effect. While interventions with a prevention focus yielded a slightly larger pooled effect size, moderator analyses showed no significant difference with interventions promoting plant-based food intake. This contrasts with prior research, suggesting that a promotion focus is more effective in encouraging healthy dietary patterns [26,117]. However, this might be explained by a difference in the food consumption studied. For example, the studies reviewed by Zlatevska et al [117] primarily targeted discretionary items such as sugary drinks and snacks, not meat consumption. Only 1 of the included interventions in this review combined a prevention and promotion strategy and found sustained effects at follow-up [64], highlighting a need for future research to test more integrative approaches that align closely to the EAT-Lancet diet.

Both the narrative synthesis and moderator analyses show that the effectiveness of digital interventions varies by modality. SMS text messaging interventions had a moderate effect on healthy and sustainable food intake, with over 75% demonstrating significant results in the narrative review. Despite the small number of social media interventions, they had the strongest effects on eating behavior, being significantly associated with larger effect sizes than other modes of delivery. While prior research associates social media use with unhealthy eating habits [118,119], these findings suggest that social media can also be used to positively influence dietary behavior. Since many consumers engage with food content on social media, platforms such as Instagram or Facebook hold great potential to promote healthier and more sustainable eating [120,121]. Social media platforms present unique environments in which users can form large social networks, allowing them to seek information about others’ behavior and receive positive reinforcements for their own behavior. Although the evidence base remains limited, social media interventions may promote stronger behavior change for several reasons: they expose users to peers’ behaviors (social modeling), leverage participants’ existing engagement with the platform to sustain intervention exposure and retention, and offer interactive features that facilitate social support [122,123]. Notably, none of the social media interventions targeted adolescents, despite this being a key demographic due to their high engagement with social media and high exposure to unhealthy food marketing [124,125].

These findings suggest that interventions delivered through accessible and familiar platforms (ie, SMS text messaging or social media) tend to be more effective than interventions requiring more intentional user engagement (ie, stand-alone websites or apps). This difference also links to the amount of agentic demand (ie, the degree to which participants are required to engage with the content in order to achieve the intended outcome) [126,127], though this requires further empirical testing. The common academic practice of developing digital interventions via research-created platforms, often used only briefly and actively throughout the study period, implies a misalignment with real-world digital behavior [35]. To enhance ecological validity and long-term behavior change, future studies could focus on accessible interventions using existing, familiar platforms [126,128]. While social media interventions were significantly more effective than interventions via other platforms, they also present specific challenges for researchers. As opposed to more controlled intervention environments such as apps or websites, social media constitute open and dynamic spaces that are saturated with unhealthy food marketing and health misinformation [129]. This may dilute or even counteract intervention effects and can lead to nutrition confusion [130].

A wide range of BCTs was incorporated across the interventions, and the choice of BCT varied depending on the type of digital medium used. BCTs targeting goals and planning were most frequently applied across all interventions, particularly in web-based interventions. Knowledge-related techniques, such as providing instructions, were frequently applied, especially in web-based interventions or games. Communicating the consequences of unhealthy eating behavior was also a commonly used cluster of techniques, mainly in SMS text messaging interventions. Yet, these information-based BCTs were not the most effective ones. Although definitive conclusions on the most effective BCT cluster cannot be made due to the limited number of studies incorporating certain clusters, moderator analyses [1] indicated that interventions incorporating prompts or cues (BCT cluster 7) were significantly more effective than those that did not, yielding a large effect size. Additionally, interventions targeting social mechanisms through social support (cluster 3) or comparison of behavior (cluster 6) demonstrated larger effects, though not significantly larger than those not including these BCTs. The narrative synthesis further suggests that social norm communication is an effective strategy, as several studies with significant effects incorporated descriptive norms to influence behavior. These findings suggest that future interventions may benefit from shifting emphasis away from purely informational strategies (eg, raising awareness or increasing knowledge) and instead testing techniques that leverage social influence, peer dynamics, and cues. Previous research on health interventions also suggests that providing social support strengthens the impact of digital interventions [131-133], and that norm communication can gradually reshape individuals’ perceptions of others’ behavior, prompting them to align with these evolving norms [134-136].

Finally, while age-related differences in the effect size of the interventions were not significant, moderator analyses yielded the strongest effect size for young adults. Due to the limited number of studies targeting adolescents, conclusions for this age group cannot be drawn. Regarding SES, most studies included SES indicators only as control variables rather than focusing on targeted recruitment, resulting in the underrepresentation of lower SES individuals. This limited our ability to conduct subgroup analyses for this population. This lack of research is concerning, given that food intake is socially structured, with unhealthy and unsustainable diets being more prevalent among lower SES individuals [137-139]. While financial constraints and educational disparities contribute to these patterns, modifiable psychosocial factors (eg, attitudes, literacy, desired identity, or social norms) also play a critical role [140-142]. Despite potential challenges related to digital access and literacy, it is crucial to reach lower SES populations in the current digital era to help mitigate health inequalities related to dietary behavior, especially adolescents, given that they are highly active on digital platforms. Leveraging platforms such as social media may help embed interventions within their daily routines and reduce barriers to participation. Future research should conduct both targeted testing of interventions within lower SES groups and exploration of SES moderation effects to understand intervention effectiveness across socioeconomic groups.

This review reveals some major gaps in the literature. First, individuals from lower socioeconomic backgrounds remain underrepresented in research testing digital interventions. None of the included studies used subjective measures of SES (ie, perceptions of own social status) [2] despite evidence that perceived SES is a strong predictor of health (behavior) even after controlling for objective indicators [143-145]. Future research should therefore not only systematically measure SES but also incorporate subjective indicators more frequently [146]. Second, while the included interventions targeted various age groups, few studies specifically focused on adolescents. The handful of studies that included adolescents as participant group primarily tested web-based or mobile app interventions, which were among the least effective digital modalities. Both the lack of research as well as the prominent focus on websites or apps may explain the smaller effect size for this demographic. Future research should prioritize low-agentic interventions for adolescents of different socioeconomic backgrounds, particularly through social media, as adolescents not only increasingly use these platforms, but research also shows that exposure to social media food content influences their eating behavior [118,124]. Finally, most studies focused on specific food groups, particularly fruits and vegetables, framing them as key components of a healthy diet. These findings are similar to previous systematic reviews on interventions aiming to improve food consumption [33,147]. While fruits and vegetables are indeed a critical part of a healthy and sustainable diet, interventions should also highlight other important components, such as legumes and whole grain products. Furthermore, rather than solely focusing on reducing meat consumption, interventions can promote replacing meat with whole-food plant-based alternatives, as substituting certain products is often more achievable than complete elimination [18,64].

By identifying the conditions in which digital interventions are most effective for encouraging plant-based eating, this study offers timely guidance for practitioners in a rapidly evolving digital landscape. The results of this review suggest that digital interventions can effectively be implemented to support people in transitioning to healthier and more sustainable eating behavior. Given the variability in effectiveness following intervention characteristics, health care providers and practitioners should consider selecting digital platforms that best fit the preferences, literacy, and lifestyle of the target group. Social media may be particularly useful for delivering low-threshold interventions in community settings, targeting mechanisms of social influence for behavior change. Practitioners are encouraged to adopt evidence-based BCTs that leverage social influence (ie, social support and demonstration of behavior) and thus shift emphasis from purely informational strategies (eg, raising awareness or increasing knowledge). Collaboration between behavioral scientists and health care professionals can help ensure that interventions are evidence-based and aligned with participants’ needs. Special attention is needed for the participation of underrepresented groups, more specifically, adolescents and individuals with lower SES. Future research should use specific strategies to recruit and engage these vulnerable populations, for example, through school-based programs or partnerships with community organizations. Moreover, it is important to ensure that intervention materials are accessible and inclusive to help reduce disparities in participation and retention. One way to ensure this is by actively involving members of the vulnerable group or trusted intermediaries in the design of the intervention [148].

Despite using extensive search strategies, relevant research may have been overlooked if published in languages other than English or if they were inaccessible through the selected databases. The majority of included studies demonstrated moderate risk of bias, with measurement of outcome bias being the most prevalent concern due to reliance on self-reported dietary measures rather than objective assessments. This may lead to overestimation of intervention effects, as self-report measures are susceptible to social desirability bias and recall error. Future research should prioritize objective outcome measures to strengthen evidence for real-world applicability. In total, 11 studies could not be included in the meta-analysis due to insufficient data for effect size calculation. Additionally, the inconsistency in measuring SES and the scarcity of research involving participants from lower SES backgrounds hindered our ability to assess the differential effects of digital interventions across SES groups. Moreover, the small number of studies contributing data to certain moderators may have limited the statistical power to detect potential moderation effects [149] and restricted our ability to assess their combined influence. Ideally, a sufficiently large number of studies would be available to allow the modeling of joint effects, rather than examining each moderator separately. Furthermore, most evidence was derived from studies with a moderate risk of bias, and substantial heterogeneity was present. Although the subgroup analyses explained some of the observed variation, these factors may still affect the overall certainty of the evidence. We did not perform a formal Grading Recommendations Assessment, Development, and Evaluation assessment to systematically rate the certainty of the evidence, which limits our ability to evaluate the confidence in our findings [150].

Another potential limitation is the inclusion of both RCTs and NRS in the meta-analysis. The appropriateness of combining different study designs remains debated in the methodological literature, with studies having inconsistent conclusions about the risks of pooling RCTs and NRS (eg, achieving statistical significance only after inclusion of NRS) [109,151,152]. We retained NRS in our primary analyses for 2 reasons. First, sensitivity analyses excluding NRS demonstrated that our findings were robust, with effect estimates, CIs, and heterogeneity statistics remaining consistent. Second, NRS are increasingly recognized as valuable sources of evidence, also within digital intervention research, as ecological validity and real-world application are important criteria. In this context, NRS can provide complementary evidence alongside RCTs [153,154]. However, the pooled estimates should therefore be interpreted with awareness of some methodological considerations. RCTs and NRS are susceptible to different types of bias [151,152]. RCTs face risks related to randomization, allocation concealment, and blinding, while NRS are prone to confounding and selection bias, which is reflected in the different risk-of-bias assessment tools used for each design. The pooled estimates, therefore, represent weighted averages that incorporate these different types of bias, which may complicate direct translation of the results to practice [151,152]. Moreover, these study designs provide different types of evidence: RCTs answer questions of effectiveness under controlled conditions, while NRS can, for instance, address generalizability or real-world performance [151,153]. Consequently, the primary findings should be interpreted as reflecting the overall evidence base for digital interventions, rather than as a pure estimate of intervention effectiveness derived solely from controlled trials. Importantly, the consistency of the findings in the sensitivity analyses restricted to RCTs supports the robustness of the study’s main conclusions regarding intervention effectiveness.

Other important limitations are with regard to the BCTs. Categorizing intervention components into BCTs based on study descriptions and protocols proved challenging, as studies used varying terminology and levels of detail on the intervention content. Similar issues have been highlighted in previous systematic reviews [36,146]. Therefore, despite the use of a coding manual and following a training in BCT taxonomy, the presence of some techniques may have been overlooked. To enhance the replicability of interventions and accuracy of coding BCT presence for meta-analyses, future research should describe intervention content in greater detail and more systematically by following taxonomies or reporting guidelines [24,155]. It is also essential to recognize that BCTs are not exhaustive; they possess a certain level of superficiality. While BCTs can be implemented in various ways, the content was generalized to a certain type of BCT, potentially masking variations in their implementation. For instance, informing participants about the components of a healthy diet or providing personalized recipes both fall under the same BCT category (4.1 Instruction on how to perform the behavior) but represent distinct approaches that may yield different effects. Moreover, some characteristics of persuasive communication are overlooked by focusing on BCTs (eg, framing of information).

This review provides a comprehensive overview of digital interventions, suggesting that digital interventions can effectively improve eating behavior, though their success varies by intervention design and population targeted. Social media emerge as particularly promising, likely due to their unique social and interactive features. Importantly, the evidence base mainly consists of studies with a moderate risk of bias, highlighting the need for more high-quality studies to confirm current results. Moreover, the meta-analytic results have a broad PI, indicating that while the average effect is positive, individual interventions may range from highly effective to potentially ineffective, depending on context and design. To our knowledge, it is one of the first reviews to systematically code the characteristics of digital interventions, including their mode of delivery (ie, digital medium), content (ie, BCTs), behavioral goal orientation (prevention vs promotion), and targeted demographic (ie, age and SES), and to link these with the intervention effect size. Despite our aim to explore effects specifically among low SES groups, the limited available research restricted our ability to conduct subgroup analyses for this population. Our findings offer valuable insights for practitioners and researchers interested in leveraging digital media for behavior change by providing an evidence base on the contexts and types of digital interventions that most effectively promote plant-based eating.