Algorithmic Intimacy: How AI-Personalised Marketing Shapes Brand Trust and Brand Loyalty

3. METHODOLOGY

In this study, the researchers have used the between-subjects randomised controlled experiment to investigate the effect of AI-personalised marketing on brand trust and brand loyalty in the fashion industry. The research aimed to determine the influence of varying degrees of AI personalisation on consumer’s attitudes and behaviour (i.e., brand trust and brand loyalty). The experiment design was a between-subjects design, with participants randomly assigned to one of three groups. Random assignment was done by the online survey tool to eliminate potential biases by giving each participant an equal chance of being in one of the conditions. Different groups were also exposed to different levels of AI personalisation, allowing for comparisons of the impact of brand personalisation on brand trust and brand loyalty. Experimental conditions were the following:

Control: Exposed to traditional non-personalised marketing (i.e., no personalisation of ads).

Experiment Group 1: Exposed to ads with basic personalisation, which included personalised information such as the consumer’s name and location.

Experiment Group 2: Exposed to highly personalised ads, which included personalised information such as the consumer’s purchase history, browsing history and preferences.

For each group, 300 people were recruited through online crowdsourcing websites (MTurk & Prolific). But it is noted that the sample was somewhat unbalanced in terms of age and gender, and this may lead to sampling bias and a lack of external validity. This type of recruitment was aimed at reducing potential sampling bias and to ensure a diverse sample across key demographics such as age, gender and technological success. The sample is eclectic which enhances external validity.

3.1. Pre-Test and Post-Test Surveys

The study had two main phases: the pre-test and post-test survey.

Pre-test survey: This survey gathered demographic information (such as age, gender, and technology literacy) and participants’ baseline brand trust in the digital marketing environment. It also assessed participants’ awareness of AI technology to avoid biases due to experience. These were used as control variables to minimise the impact of these differences on the experiment.

Marketing stimulus experiment: The participants were exposed to 3-5 advertisements, depending on the group they were assigned to. The advertisements were shown for only a certain amount of time to standardise the amount of time spent viewing advertisements. To overcome the realism problem, the mock-up ads were designed to be representative of e-commerce ads or social media ads (confirmed by expert opinion). The design, colour and content of the advertisements were kept standardised to prevent confounding factors and to ensure that any effects were due to the personalisation of the message. However, manipulation checks were also undertaken to ensure that the participants had the same perception of personalisation in the two conditions.

Post-test survey: Participants also completed a post-test survey after exposure to the marketing to assess perceived algorithmic intimacy (PAI), brand trust, brand loyalty, and privacy concerns. Questions about the creepiness and authenticity of the personalised ads were also included in the questionnaire. Respondents were also asked whether they would repurchase the brand or refer their friends to purchase it, both of which are essential indices of brand loyalty.

3.2. Manipulation Checks and Construct Validity

The manipulation checks were introduced to guarantee the construct validity of the study. The questionnaire required the respondents to assess the extent of personalisation they felt in the advertisements they had watched. A one-way ANOVA was run on the manipulation check items to ensure that there were significant differences between the three experimental groups, indicating that the different levels of AI personalisation were perceived by participants.

3.3. Instruments Used

The Trust Scale (McKnight et al., 2002) was used to measure the brand trust. The scale uses the benevolence, integrity and competence dimensions of a brand, which are important in brand trust in the online context. This scale is commonly employed in the literature to measure brand trust in online e-commerce platforms and AI marketing (Febrian, 2025). The Brand Loyalty Scale was used to measure brand loyalty (Bobâlcă et al., 2012). The scale is used to assess the preference for the brand, brand attachment and purchase intention. These are critical in measuring the effect of personalised marketing on brand loyalty in the long term.

The Perceived Algorithmic Intimacy (PAI) measure was a Likert-type scale that has been developed to measure emotional and social responses of consumers to AI-driven interactions. The tool was conceptually based on previous research on parasocial interaction, computational intimacy, and human-computer relationship theory, highlighting perceived proximity, emotional attachment, and personalised responsiveness in mediated relationships (Horton & Wohl, 1956; Daft & Lengel, 1984; Ferreira & Pereira, 2025). The scale items measured the perceptions of emotional connection, personal relevance, and perceived closeness to the brand caused by AI personalisation. The internal consistency analysis ensured satisfactory reliability by linking items with recognised conceptual aspects of intimacy in online relationships and took into account content validity. This approach validates PAI as a mediating factor between AI exposure and awareness, and brand loyalty.

For AI Exposure and Awareness, we used the study of (Gao & Liang, 2025) and modified the statements to suit the environment in our study. It focuses on the impact of exposure and awareness of AI on purchase decision-making and brand loyalty.

In this study, (Kudapa, 2024) scale was adopted to measure Privacy Concerns. Some words were replaced to reflect AI-controlled personalised marketing in the fashion sector, to ensure relevance. Businesses that are not exposed to AI were excluded, and terms were altered to capture interest in data collection, openness and invasiveness. Reliability Analysis of Scales. Cronbach’s Alpha was used to test the reliability of the scales used in the study, as per Table 1. The results showed:

Table 1. Reliability.

These measures suggest that the scales used to measure brand trust, brand loyalty, privacy concerns, AI exposure and awareness and algorithmic intimacy (PAI) were consistent in measuring the constructs.

3.4. Data Collection and Instrument Questionnaire

The aim of the study was to ensure there were consistent procedures for all participants. Firstly, all participants answered the pre-test questionnaire (Appendix A), and then viewed the personalised advertising. The participants were given the post-test survey. This pre-post quantitative two-stage experiment ensured that the variation in brand trust and brand loyalty were valid measures that could be measured before and after the marketing campaign. The questionnaire used in the data collection process is attached to the study. It includes all questions on brand trust, brand loyalty, algorithmic intimacy (PAI), and privacy issues.

The paper has used a detailed data testing procedure that includes One-Way ANOVA and the Process Macro Model 4 to test the aim and objectives/hypotheses of this study. One-Way ANOVA was selected because it is appropriate for comparing mean differences across more than two independent groups under experimental conditions. The means of the three experimental groups (non-personalised, basic AI personalisation, and deep AI personalisation) were compared using a one-way ANOVA to test H0, H1, and H5. This has enabled the determination of high and low levels of AI personalisation across trust, loyalty, and privacy. Process Macro Model 4 has subsequently been used to test the mediation effect of Algorithmic Intimacy (PAI) in the correlation between AI personalisation and brand trust and brand loyalty. This technique was chosen due to its robustness in testing indirect effects within behavioural research. This hybrid methodology has presented a powerful construct through which AI personalisation can be seen to affect consumer outcomes and the mediating effect of algorithmic intimacy in such relationships.

4. RESULTS

4.1. Frequency Analysis

The frequency analysis for the pre-test survey is given in Table 2:

Age: The majority of respondents are young adults (25-44 years old at 54.7%) which suggests tech-savvy consumers with awareness of AI marketing are well represented. However, this may mean that the results are more representative of tech-savvy consumers than the general population.

Gender Distribution: There is a strong male bias (77.3%) in the sample, which can impact the interpretation of the results on brand trust and brand loyalty. This skew is a potential limitation for generalising the findings to both genders

Understanding of Digital Marketing: Many respondents (67.4%) are familiar with digital marketing technologies, in line with increasing familiarity with AI and awareness of AI-based marketing practices. This justifies the investigation into the impact of AI personalisation but also emphasises privacy concerns. Privacy Concerns: Although the majority of respondents are at least slightly concerned about privacy 29.7% are slightly concerned and 27.3% are moderately worried this highlights the need to address privacy issues in AI-driven marketing techniques. This is consistent with the theoretical debate about the privacy-personalisation paradox.

Table 2. Pre-test survey analysis.

4.3. One-Way Analysis of Variance (ANOVA)

All the variables had statistically significant Kolmogorov-Smirnov and Shapiro-Wilk tests (p < .001), which is an indication of non-perfectly normal data (Table 3). ANOVA is however, said to be robust to normality violations due to the large sample (n = 300). Also, the skewness and kurtosis values were visually inspected, and it showed that the distribution patterns were acceptable. Hence, it was considered that parametric testing was relevant. The statistical test shows that the groups for each dependent variable differ significantly, confirming the hypotheses (H1-H4) that AI personalisation affects perceptions of algorithmic intimacy, Brand trust, loyalty, and privacy concerns. This analysis directly tests the hypotheses derived from the theoretical framework discussed earlier.

Table 3. Tests of normality.

One-way ANOVA was conducted after the experiment. To analyse the effects of different levels of AI personalisation on four primary dependent variables, the One-Way Analysis of Variance (ANOVA) as presented in Table 4 was performed to examine the effects of various degrees of AI personalisation on the four main dependent variables namely Perceived Algorithmic Intimacy (PAI), Brand Trust (BT), Brand Loyalty (BL) and Privacy Concerns (PC). Assumptions of normality and homogeneity of variance were assessed prior to hypothesis testing. Visual inspection of skewness and kurtosis values indicated acceptable normality across constructs.

Table 4. ANOVA results for hypothesis testing.

The ANOVA results indicate statistically significant differences across the three experimental conditions for all dependent variables. Perceived Algorithmic Intimacy, Brand Trust, and Brand Loyalty showed strong effects of AI personalisation (p < 0.001), while Privacy Concerns also differed significantly across groups (p = .008). The difference in PAI among the three groups in the experiment is highly significant, as indicated by an F-value of 25.423 and a p-value < .001. The difference in BT (Brand Trust) is also significantly different across the experimental groups, as the F-value of 27.546 and p-value < .001 again support the hypothesis that brand trust increases when the level of AI personalisation is high. BL (Brand Loyalty) outcomes are also significantly different between the experimental groups, with an F-value of 15.910 and a p-value < .001. The PC (Privacy Concerns) indicates that the three groups in the experiment differ significantly, with an F-value of 4.864 and a p-value of 0.008. In line with the Parasocial Interaction and Media Richness theories, greater AI personalisation appears to lead to heightened emotional closeness. The results of the one-way ANOVA confirm that AI personalisation is statistically significant in influencing brand trust, brand loyalty, and privacy concerns. Still, the extent of this influence varies across these variables. The most affected were brand trust and brand loyalty, as the greater the level of personalisation, the greater the increase in both.

4.4. Post-Hoc Test (Tukey HSD)

The Post-Hoc Tukey HSD Test was conducted in a bid to test pairwise dissimilarity between the three experimental conditions (Non-personalised, Basic AI personalised and Deep AI personalised) and four dependent variables (Perceived Algorithmic Intimacy (PAI), Brand Trust (BT), Brand Loyalty (BL) and Privacy Concerns (PC)). This analysis clarifies which levels of personalisation drive the observed ANOVA effects. Table 5 analyses the results of each of these variables. The post hoc Tukey HSD Test Results are characterised by asterisks, i.e., the statistically significant difference between the groups of experiment with a p-value that is less than the traditionally accepted value, i.e., 0.05. These considerable differences indicate that the differences that were found between the groups are not likely to have happened by accident. For example, in Non-personalised vs. Deep AI personalised PAI, the mean difference is -0.83333 with a p-value < .001, marked with an asterisk, indicating it is statistically significant. In contrast, comparisons that lacked asterisks, e.g., Non-personalised vs. Basic AI personalised in PAI, have p-values exceeding 0.05, indicating no statistically significant difference between the groups. They also use asterisks to mark points of substantial change in the data clearly. This suggests a threshold effect, where emotional and relational outcomes intensify only at higher levels of AI personalisation.

Table 5. Post-hoc turkey HSD test results.

Note: ***: significant at 1%; **: significant at 5%

The means plots (Figs. 2–5) show the results of the experiment for each group across all variables used in this study. At the same time, privacy concerns also increase, highlighting a clear benefit–risk trade-off rather than a uniformly positive effect. The average in the Deep Personalised AI groups is clearly large across all variables, indicating that Deep Personalised AI had an impact. That is, the higher the AI personalisation, the higher brand trust, brand loyalty, perceived algorithmic familiarity and perceived privacy issues. The importance of the final variable, privacy concern, is, however, an issue that requires further research to resolve.