Enabling and Understanding Personalization in AI-Generated Advertising Imagery
Victor Kolominsky-Rabas, Leopold Müller, Claudius Budcke, Claas Christian Germelmann, Niklas Kühl
cs.AI · cs.HC
Quick verdict
Preregistered N=100 study finds GenAI ad imagery works best at moderate personalization; credible, but L2 effects confounded with image quality.
Summary
The paper develops and evaluates an AI framework that generates personalized advertising imagery directly from customer data. A rule-based script converts survey-collected customer attributes into a persona description; an LLM (GPT-5) turns this into a scene prompt; a text-to-image model (gpt-image-1) renders the final ad. In a preregistered, ethics-approved two-stage within-subjects study (N=100 US Prolific participants; four products spanning the FCB grid), participants first supplied demographic, lifestyle, and appearance data, then rated baseline (L0), light (L1), and strong (L2) personalization on single-item measures of attitude toward the ad (ATA), attitude toward the product (ATP), and purchase intention (PI). Repeated-measures ANOVAs show level differences for all three outcomes; planned one-tailed t-tests show L1 beats L0 for ATA and ATP but not PI (p=.085), and L1 beats L2 on all three; L2 never beats L0 and is numerically lowest. Mediation analysis indicates perceived personalization raises all outcomes while perceived creepiness lowers them, with creepiness dominating at L2. Key limitations: all constructs use single items; personalization depth is confounded with generation complexity and visible artifacts at L2, which the authors acknowledge; mediators were measured after outcomes; no code or data are released.
Review complete
Assessment at a glance
5
risk flags identified
● high 1
● medium 3
● low 1
4 strengths · 5 weaknesses · 5 author questions
AI text signal = Binoculars zero-shot detector score over 20 sampled passages (Fast-DetectGPT co-reported) — absence does not prove human authorship, and a flag does not imply misconduct.
Levels reflect verification complexity — not a quality score. Evolve never rates your paper.
Key claims under review
- A fully automated LLM+T2I pipeline can generate advertising imagery whose personalization level participants reliably perceive (Table 4: L2>L1>L0 in perceived personalization).
- Light personalization (L1) maximizes ATA and ATP, significantly beating both the L0 baseline and the L2 condition (Table 3).
- Strong personalization (L2) adds no benefit over baseline on any outcome and is numerically lowest throughout (Table 6).
- Perceived personalization and perceived creepiness are competing mediators whose balance turns negative at high personalization (Tables 5, 7).
- The inverted-U pattern extends the personalization-paradox to fully AI-generated advertising imagery.
Strengths
Clean manipulation of personalization depth
Prompt templates stay constant across levels; only the rule-generated persona description varies (§3.1.2), and the manipulation check confirms monotone perception (Table 4: perceived personalization L2>L1>L0, all Holm-Bonferroni corrected p≤.004).
Preregistered, ethically approved, honest reporting
OSF preregistration (2025-08-20) and ethics approval (2025-05-28) predate data collection; partially supported H2 and mixed H3 are reported as such; Figure 7 openly displays flawed L2 outputs and §4.4.2 states no generated image was manually filtered.
Quantified competing-mechanism explanation
Concrete, well-specified framework architecture
Weaknesses
Personalization confounded with image quality at L2
All constructs measured with single items
Mediation is exploratory with post-outcome mediators
Sample and design constraints
Reproducibility gaps
Full Review Report
$1.99 · one-timeUnlock the complete preliminary analysis Evolve produced for this paper.
Report preview
Scope & Claims
The paper pursues two questions: (RQ1) how to design a technical framework that generates personalized advertising imagery from customer data, and (RQ2) how the level of personalization affects attitude toward the ad (ATA), attitude toward the product (ATP) and purchase intention (PI). Claimed contributions: a fully automated three-step pipeline (rule-based persona creation → GPT-5 scenario guidance → gpt-image-1 synthesis with a product reference image); a preregistered within-subjects evaluation manipulating personalization depth (L0 generic baseline, L1 demographic/geographic, L2 adding sociodemographic, lifestyle, household, pet and appearance attributes); and a mechanistic account in which perceived personalization (positive) and perceived creepiness (negative) compete. The authors position this as the first study combining fully generated imagery, real individual customer data, and manipulated personalization depth.
The remaining sections cover Methodology Assessment, Statistical Evidence, Results & Reproducibility, Limitations & Risks, and Verdict & Recommendations.
- All 4 strengths & 5 weaknesses, analyzed in detail
- 5 risk flags with severity, section & reviewer questions
- Methods & statistical evidence assessment
- 5 questions a peer reviewer would ask
- Full six-dimension evaluation report & recommendations
One-time purchase · this paper only
One-time purchase, this paper only. AI-generated preliminary analysis — not peer review.
Frequently asked questions
Does more personalization in AI-generated ads always improve results?
No. In this study, 100 participants rated personalized ad images generated from their own survey data. Light personalization (demographic and geographic data) produced the most positive ratings of the ad and the product, while strong personalization using detailed lifestyle, household and appearance data scored lowest — even below the generic, non-personalized baseline — because perceived creepiness outweighed the relevance gains.
What AI models were used to generate the personalized ads?
The pipeline used OpenAI's GPT-5 to turn a customer persona description into a detailed scene prompt, and OpenAI's gpt-image-1 (fed a product reference image, with 1024×1536 output) to generate the final advertising image. Persona descriptions were produced by a rule-based script from survey answers.
Was the study preregistered?
Yes. The authors preregistered hypotheses and analyses on the Open Science Framework on August 20, 2025, and obtained ethics approval from the University of Bayreuth on May 28, 2025, both before data collection.
What is the personalization-paradox finding in this paper?
The personalization-paradox holds that more customer data can raise relevance but also trigger privacy discomfort. This paper extends it to fully AI-generated ad images: perceived personalization positively predicted attitudes and purchase intention, while perceived creepiness negatively predicted them, and at the strongest personalization level the negative creepiness effect dominated the total effect.
About this review
This is an AI-generated preliminary review by Evolve.Pub, produced from the openly available text of the paper. It is not peer review and does not represent the authors' views. The canonical record of the paper is at the publisher link above.
- All figures are raster images in the PDF; Figure 3's exact prompt templates and the visual quality of the generated images (Figures 4, 6-8) could not be assessed from text extraction.
Go deeper on this paper
Wrote a paper? Stress-test it before reviewers do.
Upload your manuscript for the same structured scrutiny — reasoning, statistics, evidence, figures, and reproducibility.