Enabling and Understanding Personalization in AI-Generated Advertising Imagery

• Prior ad content has mostly been tailored to broad segments because more granular conventional production is costly, while existing work leaves real-data image personalization and its marketing effects unexamined.Source: §1 · pp. 1–2
• Light personalization (L1) produces the most favorable responses across ATA, ATP, and PI; stronger personalization (L2) adds no benefit and can reduce all three.Source: §1 · pp. 2–3
• Perceived personalization positively predicts ATA, ATP, and PI, while perceived creepiness negatively influences them; creepiness helps explain diminishing returns.The proposed explanation connects perceived creepiness with limitations of current GenAI model capabilities.Source: §1 · p. 3
• The study introduces and implements a framework coupling an LLM with a text-to-image generator to create advertising imagery from real customer data.Source: §1 · p. 3
• A controlled within-subjects evaluation uses real participant profiles and image-level measures of ATA, ATP, and PI, with perceived personalization and creepiness as manipulation checks.Source: §1 · p. 3
• The authors present light personalization as evidence aligned with the personalization-paradox and as guidance for low-cost, scalable GenAI image-ad deployment.Source: §1 · p. 3

• The personalization-paradox describes how additional customer data can raise relevance while also triggering privacy concerns and reactance.Source: §2 · p. 3
• Recommendation systems and DCO typically match users to existing ads or recombine preset assets, while GenAI enables personalized content generation.Source: §2 · p. 3
• LLMs have generated persuasive psychologically tailored messages at scale, motivating tests of whether similar benefits extend to visual personalization.Source: §2 · p. 3
• Studies of synthetic advertising, AI-ad disclosure, and creative-goods markets establish relevance but do not address real-profile image personalization or its outcome effects.Source: §2 · p. 4
• GenAI image-personalization work spans product matching, background edits, brand-level generation, and coarse segments, often without individual scenes or human outcome measures.Source: §2 · p. 4
• Prior work has not systematically tested fully AI-generated personalized imagery embedding products in customer-tailored scenes, based on real individual data, with varied personalization depth and established marketing outcomes.Source: §2 · p. 4

3.1.1 Conceptual framework for AI-generated personalized advertising imagery

• The framework embeds products in customer-tailored visual narratives and automates customer-varying steps so it can scale to individual customers.Its sequence adapts persona creation, creative briefing, and image production into a personalized imagery workflow.Source: §3.1.1 · pp. 4–5

3.1.2 Technical implementation of the framework for AI-generated personalized advertising imagery

• A Python script converts structured customer data into personas, an LLM creates scenario guidance, and a T2I model generates images using that guidance and a product reference.The implementation uses gpt-5 for scenario guidance and gpt-image-1 for image generation.Source: §3.1.2 · p. 5, Figure 3 · p. 6
• Prompt templates remain fixed across L0–L2; only persona data vary, so personalization emerges from the amount and specificity of customer data, apart from GenAI output variation.L0 uses no customer data; the higher levels use increasingly specific persona descriptions.Source: §3.1.2 · p. 6

3.2.1 Study variables

• Personalization levels are the independent variable: L0 uses no customer data, L1 uses general demographic and geographic information, and L2 adds more specific personal attributes.The levels are conceptually ordered by increasing individualization, not treated as equally spaced.Source: §3.2.1 · p. 6
• Confirmatory outcomes are attitude toward the advertisement, attitude toward the product, and purchase intention; perceived personalization and creepiness are exploratory measures.The study also treats perceived personalization as a manipulation check and creepiness as a possible response to personalization.Source: §3.2.1 · p. 7

3.2.2 Confirmatory hypotheses

• H1 tests whether personalization affects marketing outcomes, while H2 tests whether its effect is positive across the three levels.The hypotheses concern whether at least one level differs, and whether at least one difference is positive, respectively.Source: §3.2.2 · p. 7
• H3 predicts an inverted-U pattern, with L1 most effective and L2 potentially reducing outcomes through resistance or perceived intrusiveness.The rationale includes the possibility that greater data use heightens privacy concerns and that highly personalized imagery may provoke discomfort.Source: §3.2.2 · pp. 7–8

3.2.3 Pre-registration and ethical approval

• The study was preregistered before data collection and received ethics approval; participants provided informed consent.The passage dates preregistration to August 20, 2025, and approval to May 28, 2025.Source: §3.2.3 · p. 8

4.1 Product selection

• Four products were selected to span the FCB Grid’s involvement and decision-driver quadrants, balance stereotypical demographic interests, and vary likely indoor and outdoor depictions.The products are Polestar 4, RIMOWA Original Cabin, Tide Ultra Hygienic Clean, and Lindt Lindor Truffles.Source: §4.1 · p. 8, Table 1 · p. 8

4.2 Measurement of dependent variables

• Attitude toward the advertisement, attitude toward the product, and purchase intention use five-point Likert scales with single-item adaptations from validated scales.The product wording for purchase intention varies with the product being evaluated.Source: §4.2 · pp. 8–9

4.3 Measurement of exploratory variables

• Exploratory measures assess perceived personalization and perceived creepiness using concise single-item formulations on five-point Likert scales.The items ask whether the scene targets the respondent as a unique individual and whether the image feels creepy.Source: §4.3 · p. 9

4.4 Survey implementation

• Survey 1 collects participant information used to generate images, and Survey 2 returns those images to the same participants for evaluation.Both surveys were administered using LimeSurvey on a department-managed server.Source: §4.4 · p. 9, §4.4.1 · p. 9

4.4.1 Survey 1: Data collection

• Survey 1 collected data across five groups—basic information, life stage, lifestyle, environment, and style—chosen for image-generation relevance, segmentation relevance, and real-world collectability.Lifestyle activities and hobbies were entered as free text.Source: §4.4.1 · p. 9

4.4.2 Survey 2: Image evaluation

• For each product, participants rated L0–L2 images on five items; product-page order was randomized, and participants were not told the images were AI-generated.No generated images were filtered or replaced, so evaluations included occasional artifacts or low-quality outputs.Source: §4.4.2 · p. 9, Figure 4 · p. 10

4.5 Recruitment strategy

• Participants were recruited through Prolific, restricted to desktop devices, and screened in Survey 1 for ages 18–99, native English, and U.S. location.Survey 2 was limited to the saved Survey 1 group without additional screening.Source: §4.5 · p. 10

5.1 Sample characteristics

• The final analyses included 100 participants, who provided 1,200 ratings across four products and three personalization levels.The sample followed exclusions for failed attention checks across two surveys.Source: §5.1 · p. 10

5.2 Analysis

• Across ATA, ATP, and PI, means rise from L0 to L1 then fall at L2, whose means are below L0; level effects are significant for all three.Figure 5 presents means and standard deviations by construct and personalization level.Source: Figure 5 · p. 11, §5.2 · p. 11
• L1 significantly exceeds L0 for ATA and ATP, but not PI; L2 does not significantly exceed baseline for any construct.The L2 means are numerically below L0 across all three constructs.Source: §5.2 · p. 11, Table 3 · p. 12
• L1 significantly exceeds L2 for all three constructs, but the inverted-U pattern is supported only for ATA and ATP, not PI.H3 requires L1 to exceed both L0 and L2.Source: §5.2 · p. 11, Table 3 · p. 12
• Perceived personalization and creepiness both increase stepwise from L0 to L1 to L2; the L1 creepiness effect is nonsignificant, and the mediators predict outcomes in opposite directions.The analysis used parallel multilevel mediation models with cluster bootstrap estimates.Source: §5.2 · p. 12, Table 4 · p. 12, Table 5 · p. 13
• At L1, perceived personalization and creepiness yield net positive indirect effects for ATA and ATP, but not PI; L2 indirect effects are nonsignificant across outcomes.Total effects peak at L1: ATA and ATP rise versus L0, while PI is nonsignificant; at L2, ATA and PI decrease, while ATP's effect is nonsignificant.Source: §5.2 · p. 12

6 Discussion

• L1 improves attitude toward the advertisement and product, while L2 adds no benefit and can reduce outcomes; ATA and ATP follow an inverted-U pattern.The framework produces visibly different personalization levels.Source: §6 · p. 13, Figure 6 · p. 13
• Perceived personalization positively predicts ATA, ATP, and PI, but at L2 rising creepiness neutralizes those benefits across all three outcomes.At L1, personalization benefits outweigh creepiness costs for ATA and ATP.Source: §6 · p. 13
• Creepiness alone cannot explain the L1-to-L2 decline; image-quality loss and algorithm aversion may also contribute, with L2 examples showing distorted features and text errors.The authors suggest richer prompts may increase image-generation complexity and artifacts.Source: §6 · pp. 13–14, Figure 7 · p. 15
• The results extend personalized-advertising and DCO research by examining personalization depth in fully AI-generated visuals, where customer data can increase relevance and creepiness.The discussion also relates personalization’s effects on earlier attitudes, but not purchase intention, to a hierarchy-of-effects perspective.Source: §6 · p. 14
• GenAI could lower the cost of tailoring ad content, but the authors identify image quality as a bottleneck to realizing its potential.They envision platforms managing both which product to show and the level of personalization.Source: §6 · pp. 14–15
• Future research should isolate customer attributes, test levels of customer likeness, and examine when deeper personalization is acceptable across contexts.The authors also call for separating personalization depth from image quality and testing quality-control approaches.Source: §6 · p. 15
• AI-generated personalized advertising raises concerns about customer data awareness and consent, product misrepresentation, visual flaws, trust, and autonomy.The authors call for transparency, accountability, and appropriate human oversight.Source: §6 · p. 15

7 Conclusion

• A three-step automated pipeline converts customer data into imagery through persona creation, LLM prompt generation, and text-to-image synthesis.The study varies personalization across L0–L2 by amount and specificity of user information.Source: §7 · p. 16
• In a preregistered within-subject study of 100 participants and four products, participants reliably perceived differences across personalization levels.The evaluation used two stages and measured ATA, ATP, and PI.Source: §7 · p. 16
• L1 had the highest descriptive ATA, ATP, and PI values and significantly outperformed L0 on ATA and ATP, whereas L2 did not improve outcomes and could reduce them.The reported effect of personalization was not monotonic.Source: §7 · p. 16
• Higher personalization increased perceived personalization, which positively predicted ATA, ATP, and PI, and also increased creepiness, which negatively predicted all three.Creepiness increased more strongly at high personalization, shifting the net effect toward less favorable reactions.Source: §7 · p. 16
• The authors conclude that effective GenAI-driven image personalization is achievable but benefits from conservative use of personal cues.They recommend further study of attributes, visual fidelity, and boundary conditions such as product type, disclosure, and platform context.Source: §7 · p. 16

A Appendix

• Table 6 presents descriptive statistics organized by construct and personalization level.The caption does not provide the table’s cell values.Source: Table 6 · p. 17
• Table 7 reports mediation effects by personalization level as standardized effects with 95% bootstrap confidence intervals.Bold coefficient values indicate intervals excluding zero.Source: Table 7 · p. 17
• Figure 8 arranges four products in columns and personalization levels in rows, using different participant personas.The caption describes the layout but does not state an outcome comparison.Source: Figure 8 · p. 18