The power of language in advertising is a foundational concept in marketing, dating back to early work on the psychology of selling [10]. While modern digital advertising research has often prioritized visual elements and targeting technology, the textual component remains a crucial interface for communicating value, evoking emotion, and prompting action [11]. Recent work highlights how influencer characteristics and live content impact impulsive buying in e-commerce [12]. Prior research has shown that specific word choices can significantly impact consumer attitudes and behaviors. For example, the use of emotional language can enhance ad effectiveness by creating a positive affective response, which then transfers to the brand itself [13]. Similarly, cognitive language that prompts analytical thought can be effective for products requiring more deliberate consideration [14]. However, much of this research has been conducted in controlled experimental settings or has focused on English-language advertising, leaving a gap in our understanding of these dynamics in large-scale, real-world, non-English environments. Table 1 summarizes seminal linguistic studies on online advertising.

The effectiveness of linguistic strategies in advertising is deeply intertwined with cultural context. In Japan, communication often emphasizes subtlety, indirectness, and a high-context approach, contrasting with more direct communication styles prevalent in Western cultures [15]. This manifests in advertising through the nuanced use of language and non-verbal cues. For instance, while Western ads might employ direct emotional appeals, Japanese advertising frequently leverages rhetorical expressions like metaphor and metonymy to evoke feelings and convey product value indirectly through imagery and association, particularly in beauty and health sectors [15]. These approaches align with Japanese cultural values of harmony and implicit understanding.

Further, Japanese language boasts a rich vocabulary of onomatopoeic and mimetic expressions (known as EEEs - Embodied Emotional Expressions) which are intuitively used to convey nuanced emotions and sensations [16]. The pervasive “kawaii” (cute) culture also serves as a distinct persuasive strategy, where adorable characters and gentle language can bypass direct argumentation to create affective appeal and build trust [17]. Such culturally specific linguistic preferences underscore the importance of employing tools like J-LIWC2015, which is optimized for Japanese text, to accurately capture these psycholinguistic dimensions accurately. Moreover, in highly regulated sectors like health and cosmetics, Japanese advertising must navigate strict legal frameworks (e.g., Pharmaceutical and Medical Device Act, Act against Unjustifiable Premiums and Misleading Representations) that impact permissible linguistic claims, especially regarding health benefits and exaggerated expressions. Understanding these cultural and regulatory nuances is crucial for developing effective advertising strategies in the Japanese market.

To systematically analyze the psychological dimensions of language at scale, researchers have increasingly turned to computational text analysis tools. One of the most established and validated tools is the the Linguistic Inquiry and Word Count (LIWC) [18,19]. LIWC is a dictionary-based software that analyzes text by calculating the percentage of words that impact into psychologically meaningful categories. These categories include affective processes (e.g., positive and negative emotion), cognitive processes (e.g., insight, causation), social processes, and various content-related themes (e.g., health, money, risk). Unlike complex machine learning models that can be difficult to interpret, LIWC’s dictionary approach provides transparent and theoretically grounded metrics. It has been successfully applied across numerous domains to link language patterns to psychological states and real-world outcomes, such as predicting political elections from tweets [20] or assessing public opinion [21]. This study utilizes the officially validated Japanese version, J-LIWC2015 [9], to apply this robust methodology to Japanese advertising copy.

Our analysis is guided by two complementary theories of persuasion: the Elaboration Likelihood Model (ELM) [6] and Framing Theory [7]. ELM posits that persuasion occurs via two distinct routes: a central route, characterized by careful and thoughtful consideration of the argument’s merits, and a peripheral route, which relies on heuristic cues suchs as source credibility or emotional appeals. Framing Theory complements ELM by focusing on how information is presented [7]. A gain-frame impacts the positive outcomes of taking an action (e.g., “achieve radiant skin”), while a loss-frame impacts the negative consequences of not taking an action (e.g., “don’t let tired skin hold you back”). Research, including Protection Motivation Theory [22], suggests that for health-related behaviors, loss-framing can be particularly effective because it activates a desire to mitigate risk. Therefore, we expect that “supplement” ads may benefit from loss-framed language (e.g., LIWC categories like “risk”). Conversely, aspirational “cosmetic” products may benefit more from gain-framed language that highlights positive outcomes (e.g., LIWC categories like “posemo” and “achieve”). By integrating these theoretical frameworks, this study moves beyond simply identifying which words work, aiming to understand why they work in different product contexts.

This study analyzes linguistic features impacting click-through rate (CTR) in Japanese Instagram ads (21,692 ads; July 2021-June 2023, Meta’s Marketing API). All data were collected from Meta’s Marketing API under standard developer agreements, ensuring compliance with their Terms of Service. Advertisements included were publicly accessible posts from Meta Business accounts. The anonymized dataset used in this study is available in the S1 File.

To ensure the relevance and consistency of our dataset, we applied specific inclusion and exclusion criteria. Data were collected from Meta’s Marketing API, a repository of advertisements run on Meta’s platforms. The collection period spanned two years, from July 1, 2021, to June 30, 2023. Our initial query targeted ads delivered to users in Japan. From this initial collection, we implemented a multi-step filtering process to construct our final analytical dataset. We first identified and isolated ads belonging to two of the most frequently advertised product categories: “supplement” and “cosmetic”. The selection of “supplement” and “cosmetic” categories was based on their high frequency within our dataset and their theoretical relevance to the Elaboration Likelihood Model and Framing Theory. These two categories also represent psychologically and commercially distinct segments—health-related versus beauty-oriented products—making them a theoretically meaningful and practically relevant comparison for testing differential persuasion mechanisms.

As shown in Table 2, these two categories accounted for the largest number of total ads and products, making them ideal for a focused comparative analysis. We focused our analysis on ads predominantly written in Japanese, ensuring their relevance to J-LIWC2015. We defined ‘predominantly written in Japanese’ as ad captions where at least 30% of the tokens (after preprocessing) were matched to the J-LIWC2015 dictionary, which is optimized for Japanese text. The 30% threshold was chosen to balance linguistic coverage and sample retention, ensuring sufficient representation of Japanese text without excessively excluding ads containing mixed or code-switched language. We did not apply a fixed token-based language filtering threshold to retain the real-world linguistic diversity of Instagram ad copy. To ensure that the ad copy was substantial enough for meaningful linguistic analysis, we applied an inclusion criterion requiring a minimum caption length of 5 words. To ensure CTR reliability, ads with missing click data and those with fewer than 50 clicks were excluded from the analysis. While the exclusion of ads with fewer than 50 clicks improves the stability of CTR estimates, it may also introduce bias by disproportionately filtering out smaller advertisers, which is noted as a limitation of this dataset. We did not exclude duplicate ad creatives. This rigorous filtering process resulted in a final dataset of 21,692 unique advertisements. Each ad was assigned a unique product ID. The dataset was composed of 12,206 “supplement” ads and 9,486 “cosmetic” ads. For each ad, we retained the full text of the primary caption, along with the total number of impressions and clicks. The distribution of advertisements across all collected product categories is summarized in Table 2. This table clearly indicates that “supplement” and “cosmetic” categories represent the largest proportion of advertisements within our collected data, justifying their selection for focused analysis in this study. It should be noted that this study focuses solely on the linguistic content of advertisements and does not include audience-level or targeting-related variables such as demographic segmentation, device type, or delivery optimization, which may also influence CTR.

This study analyzed secondary, non-identifiable advertising data obtained under a standard developer agreement with Meta. Under the authors’ institutional guidelines, research that involves no interaction with human subjects and uses only de-identified data is exempt from Institutional Review Board oversight; therefore, no ethical approval was required. All data were collected from Meta Marketing API. The dataset used in this study is anonymized and contains no personally identifiable information. A critical requirement for this study is compliance with data source policies. All data collection and analysis strictly adhered to Meta’s Marketing API Terms of Service and Platform Policies, ensuring the proper use and handling of publicly accessible advertising data.

We conducted linguistic preprocessing and feature extraction using the J-LIWC2015 dictionary. All ad texts were tokenized using MeCab with the IPA dictionary, and the frequency of each LIWC category was computed. Stopwords and platform-specific tokens such as hashtags and URLs were excluded prior to analysis. An overview of the preprocessing and feature extraction workflow is shown in Figure 2. We used word count as a control variable to account for message length. Full preprocessing details, including normalization schemes and illustrative tokenization examples, are provided in S1 Text (see S2 File).

The primary outcome variable was the click-through rate (CTR), a standard indicator of user engagement in digital advertising. CTR was calculated for each advertisement as (Total Clicks / Total Impressions) × 100. CTR was treated as a continuous variable because it represents proportional engagement rather than a binary outcome, and modeling it continuously preserves more information about the variance in advertising performance. Given the typically skewed distribution of CTR, a logarithmic transformation log10(CTR+0.0001) was applied to approximate normality for linear regression. The small constant ensured that ads with zero CTR were retained in the analysis. The main predictor variables were the J-LIWC2015 category scores, representing the percentage of caption words matching each psycholinguistic category. All LIWC scores were standardized as z-scores (mean = 0, standard deviation = 1) to enable comparability of effect sizes across categories. Two covariates were included to isolate the linguistic effects. First, caption word count was controlled for to account for variation in message length. Second, product-level fixed effects were implemented using dummy variables for each product ID, controlling for unobserved, time-invariant characteristics such as brand equity, pricing, visual design, and targeting configuration.

This study employs a quantitative, cross-sectional design to examine the relationship between linguistic features and advertising performance. Specifically, we address two primary research questions for Instagram advertisements within the “supplement” and “cosmetic” product categories:

To answer these questions, we constructed a series of multivariate Ordinary Least Squares (OLS) regression models. Model 1 assessed the continuous effect of linguistic intensity by regressing log-transformed CTR on the standardized proportion of words in each J-LIWC2015 category. Separate models were estimated for each product type. The specification was:

where i indexes the advertisement, k indexes LIWC categories, αp denotes product-level fixed effects, and ϵi is the error term. Model 2 focused on threshold effects, using binary indicators of LIWC category presence as predictors, while retaining the same structure as Model 1. To investigate RQ2 directly, we also estimated models using the full dataset (N = 21,692) that included interaction terms between each LIWC feature and a product category dummy (1 = “cosmetic”, 0 = “supplement”):

A significant interaction coefficient (β3) indicates differential effects of linguistic features by category.

To assess multicollinearity, we calculated Variance Inflation Factors (VIFs) for all predictors. To avoid redundancy from hierarchical LIWC categories, we excluded parent-level constructs and retained only subcategories. Maximum VIFs were 2.78 for Models 1 and 2 and 4.12 for interaction models, all below the threshold of 5. Robustness was verified through additional logistic regressions using a dichotomized CTR: ads in the top 25th percentile (CTR > 0.985%) were coded as 1 (“high CTR”), others as 0 (“low CTR”). This analysis serves as a robustness check, confirming that the key findings are not dependent on the choice of the regression model. Results largely confirmed the direction and statistical significance of the coefficients for the primary linguistic predictors observed in the OLS models. For supplements, “risk” and “discrepancy” retained positive significance; for cosmetics, “see”, “positive emotion”, and “motion” were positive predictors, while “body” and “negative emotion” had negative effects. These results, including coefficients and p-values, are detailed in S1 Appendix (Tables 5 and 6). Effects remained stable after adjusting for word count and log-impressions. Finally, to correct for multiple testing across LIWC categories, we applied the Benjamini-Hochberg procedure with a 5% false discovery rate. This adjustment mitigates the risk of Type I errors and improves the reliability of the reported findings.

A descriptive analysis of the J-LIWC2015 categories showed notable differences in the language typically used for each product type (Table 3; see also Figure 3). Supplement advertisements, on average, contained a higher proportion of words related to “Drives” (e.g., achievement, reward), “Relativity” (e.g., time, space), and, most notably, “Risk”. In contrast, cosmetic advertisements featured a significantly higher prevalence of words related to “Affective Processes” (especially “Positive Emotion”) and “Perceptual Processes” (especially “See”), indicating a greater focus on emotional appeal and visual outcomes.

The regression analyses identified several significant linguistic predictors of CTR for each product category, after controlling for word count and product-level fixed effects (Table 4 and Figure 4). For supplement advertisements (Model 1), the strongest positive predictor of CTR was the “Risk” category (β=0.108), indicating that ad copy impacting health risks or concerns was highly effective at driving engagement. The “Discrepancy” category, which includes words that highlight a gap between a current and desired state (e.g., “lack”, “should”), was also a significant positive predictor (β=0.051). Conversely, words related to “Motion” (β=−0.090) and “Negative Emotion” (β=−0.076) were negatively associated with CTR.

For cosmetic advertisements (Model 2), a different set of features emerged as dominant. The strongest positive predictor was the “See” category (β=0.134), which includes words related to visual appearance (e.g., “shine”, “bright”, “flawless”), strongly supporting the idea that visual and aesthetic benefits are key drivers of engagement. The “Positive Emotion” category (β=0.088) also showed a statistically significant positive association. Interestingly, and in direct contrast to supplements, “Motion”-related words were positively associated with CTR in this context (β=0.046). Language that directly referenced the “Body” (β=−0.103) was negatively associated with CTR. The “Negative Emotion” category (β=−0.053) was also negatively associated with CTR. These results provide a direct answer to RQ1, demonstrating that specific psycholinguistic features in Japanese Instagram ad copy significantly predict CTR, and these effects vary distinctly across product categories. The results highlight a distinct contrast in persuasive strategies: while supplements benefit from language addressing risks and discrepancies, cosmetics thrive on visual and emotional appeals. This stark difference underscores the category-specific nature of effective ad copy.

The pooled regression model (referred to as the interaction model in Table 4) confirmed that the impacts of several key linguistic features were statistically different across the two product categories, as supported by a joint Wald test (χ2=111.46, df = 7, p < 0.001). The interaction term “Risk x Cosmetic” was significantly negative (β=−0.113,q<0.001), indicating that the positive impact of risk-related language on CTR was significantly weaker for cosmetics compared to supplements. Conversely, the interaction term “See x Cosmetic” was significantly positive (β=0.149,q<0.001), confirming that the positive impact of visual, perceptual language was significantly stronger for cosmetics. Most strikingly, the “Motion x Cosmetic” interaction was large and highly significant (β=0.136,q<0.001), formally demonstrating that the impact of motion-related words reverses direction between the two categories: it is detrimental for supplement ads but beneficial for cosmetic ads. The interaction term “Body x Cosmetic” was also significant and negative (β=−0.118,q<0.01), indicating that the impact of “Body” language observed in cosmetic ads was significantly weaker (or negative) when compared to supplements (where “Body” was not a strong predictor). These significant interaction effects provide strong evidence that the optimal linguistic strategy for driving user engagement is highly dependent on the product being advertised. The significant interaction effects observed in this model directly address RQ2, providing empirical evidence that the impact of linguistic features on CTR is contingent on the product category. This finding reveals a key reversal in linguistic effectiveness: language that is negatively associated with CTR for supplement ads can be highly beneficial for cosmetic ads, and vice versa. This provides strong empirical support for the theoretical proposition that optimal linguistic strategies are context-dependent.

This study is subject to several limitations that also present opportunities for future research. The most significant limitation is that our analysis relied solely on J-LIWC2015, which quantifies psycholinguistic dimensions based on tokenized words. As a dictionary-based approach, it can miss contextual nuance (e.g., irony, sarcasm), which may attenuate or obscure certain effects. This lexical scope means our models inherently exclude persuasive cues conveyed through multimodal signals, such as emojis and images, as well as instances of English code-switching within Japanese captions. Future studies should integrate computer-vision features or cross-lingual embeddings to enrich the understanding of multimodal persuasion. Another key limitation is that our current analysis does not account for specific audience characteristics, such as age, gender, or interests. Future studies would benefit from incorporating audience-level data to test for interaction effects between linguistic features and user profiles. Lastly, the cross-sectional nature of our data limits our ability to infer causality. Longitudinal studies or controlled experiments would be valuable for establishing causal relationships between linguistic features and advertising effectiveness over time.