LLM-Generated Ads: From Personalization Parity to Persuasion Superiority

Abstract: As LLMs become increasingly capable of generating persuasive content, understanding their effectiveness across different advertising strategies becomes critical. This paper presents a two-part investigation examining LLM-generated advertising through complementary lenses: (1) personality-based and (2) psychological persuasion principles. In our first study (n=400), we tested whether LLMs could generate personalized advertisements tailored to specific personality traits (openness and neuroticism) and how their performance compared to human experts. Results showed that LLM-generated ads achieved statistical parity with human-written ads (51.1% vs. 48.9%, p > 0.05), with no significant performance differences for matched personalities. Building on these insights, our second study (n=800) shifted focus from individual personalization to universal persuasion, testing LLM performance across four foundational psychological principles: authority, consensus, cognition, and scarcity. AI-generated ads significantly outperformed human-created content, achieving a 59.1% preference rate (vs. 40.9%, p < 0.001), with the strongest performance in authority (63.0%) and consensus (62.5%) appeals. Qualitative analysis revealed AI's advantage stems from crafting more sophisticated, aspirational messages and achieving superior visual-narrative coherence. Critically, this quality advantage proved robust: even after applying a 21.2 percentage point detection penalty when participants correctly identified AI-origin, AI ads still outperformed human ads, and 29.4% of participants chose AI content despite knowing its origin. These findings demonstrate LLMs' evolution from parity in personalization to superiority in persuasive storytelling, with significant implications for advertising practice given LLMs' near-zero marginal cost and time requirements compared to human experts.

• The paper introduces a three-stage generation pipeline integrating knowledge graphs with LLMs to align user personality traits with product features.
• It finds that LLM-generated ads achieve parity with human ads for promotion traits but surpass them in multimodal persuasive appeal, especially for authority and consensus.
• The study highlights generational differences in detecting AI origin, revealing that high-quality AI content can overcome origin bias in ad preference.

LLM-Generated Advertising: From Personalization Parity to Persuasion Superiority

Generation Architecture and Experimental Design

The paper presents a structured generation pipeline integrating knowledge graphs (KGs) with LLMs to ensure semantic alignment between user characteristics (e.g., Big Five traits) and product features in advertisement creation. The pipeline proceeds in three stages: KG-driven prompt contextualization, LLM-powered narrative generation, and audience-specific delivery.

Figure 1: The proposed generation pipeline employs a KG for user-product mapping, followed by LLM-generated persuasive storytelling, culminating in targeted audience deployment.

Two major studies underpin the experimental protocol. Study 1 employs a survey-based, trait-matched approach focusing on Openness and Neuroticism, leveraging both human and LLM-generated (GPT-3.5) ad texts. Study 2 adopts a multimodal paradigm—LLM (GPT-5, Sonnet 4, LLama 4) texts with Midjourney diffusion model-generated images—targeting four classic persuasion principles: Authority, Consensus, Cognition, and Scarcity.

Study 1 participants complete Likert-scale ratings on isolated ad messages and then indicate direct preference in a forced-choice setting for four ad variants.

Figure 2: Survey framework: (i) trait-targeted Likert questions, (ii) competitive ad preference selection, (iii) Big Five inventory for trait scoring; four survey variants based on message source and trait.

Study 2 participants view matched multimodal ads side-by-side, select their preferred version, supply qualitative reasoning, and finally attempt to identify which ad is AI-generated.

Personalization: Trait-Targeted Ad Generation and Performance

Statistical analysis reveals that LLMs achieve parity with human creatives in personalized advertising for the Openness trait: matched participants rate LLM and human ads equivalently for product attitude, purchase intent, and engagement ($p > 0.05$ across metrics). Personalized appeals to Openness exhibit robust positive deltas when matched, but no significant efficacy for Neuroticism, for both machine and human sources. Thus, the Openness trait is receptive to direct creative signaling, while Neuroticism-driven appeals fail to induce engagement—attributable to psychological trait-driven avoidance patterns.

In forced-choice direct competition among four ad messages (human/AI × openness/neuroticism), aggregated preferences show marginal statistical parity (LLM: 51.1%, Human: 48.9%).

These results confirm that zero-shot LLMs, when structured via KG-based prompt engineering, match human-level ad efficacy for promotion-oriented traits but not for prevention-oriented ones, consistent with regulatory focus theory.

Universal Persuasion: Multimodal Superiority of AI Generators

Expanding scope, Study 2 demonstrates that AI-generated advertisements—comprising both LLM-crafted persuasive copy and diffusion-generated imagery—significantly outperform expert human ads in forced-choice head-to-head testing (AI: 59.1%, Human: 40.9%; $p<0.001$; Cohen's $h=0.37$). The largest effect sizes are observed for Authority (+26%, $d=0.52$) and Consensus (+25%, $d=0.50$) principles; Cognition sees a moderate advantage, while Scarcity yields no significant differential ($p = 0.671$).

Figure 3: Representative scarce-condition advertisements—left: human creative, right: AI-generated; both stressed urgency and time-limited offers.

Qualitative thematic analysis attributes AI superiority to (i) aspirational and sophisticated messaging, (ii) tight visual-textual coherence, and (iii) more precise operationalization of psychological constructs (especially credibility and social proof in Authority/Consensus).

Figure 4: Qualitative preference drivers—those selecting AI cite color/tone and model quality; human-preferring participants value authenticity and clarity.

Preference rates are robust across age and gender demographics, indicating wide generalizability of the observed AI advantage.

Human vs. Machine Attribution: Detection, Bias, and the Perception Paradox

Detection analysis uncovers a significant generational gradient in the ability to identify AI ads: digital natives (18-24) achieve 73.0% accuracy, while the oldest cohort (65+) operates at chance (48.3%).

Figure 5: Age-stratified detection accuracy; sharp generational divergence, with older participants failing to reliably identify AI-generated content.

Critically, correct identification of AI origin induces a substantial "detection penalty" in preference (Δ$_{\text{bias}}$ = 21.2pp), but this penalty does not fully negate AI performance—preference for AI remains at 50.3%, still above human ads. Approximately 29.4% of participants select an AI ad with full awareness of its origin, signaling resilience of execution quality over origin bias.

Figure 6: The Perception Paradox—preference for AI persists after detection, indicating that execution quality often dominates origin bias.

Younger cohorts demonstrate less attachment to authenticity provenance, prioritizing aesthetic and narrative "vibe," while older cohorts' search for authentic cues ironically leads them to attribute high-quality AI content as human-made.

Theoretical and Practical Implications

Empirical evidence challenges the premise of an enduring human creativity advantage—LLMs, augmented with powerful generative models, have internalized persuasive architectures (especially for promotion-focused principles) and now routinely outperform human experts. The frictionless integration of visual and narrative coherence delivered by GenAI systems establishes high processing fluency, directly impacting judgment heuristics and narrative transportation.

However, tactical resistance remains: classic scarcity appeals trigger entrenched consumer persuasion knowledge, blunting AI's superior execution. Personality-driven personalization is trait-dependent; avoidance-focused traits are impervious to both human and algorithmic creative targeting.

Implications include:

• Substantial cost and time savings for agencies deploying AI models (near-zero marginal cost).
• Uniform scalability across demographic groups.
• Shifting consumer emphasis toward content quality and structural coherence rather than authorial authenticity—a trend poised to intensify with generational turnover.
• Persistent bias from revealed AI origins but with decreasing practical relevance as quality standards rise.

Limitations and Future Directions

Key limitations arise from the reliance on US-based online panels, brief single-exposure paradigms, and constrained trait/principle domains. Absence of model fine-tuning or adaptive feedback further suggests that current performance benchmarks represent conservative lower bounds.

Future work should (i) generalize across cultures and product categories, (ii) introduce iterative personalization, (iii) track longitudinal impact on trust, manipulation, and brand outcomes, and (iv) refine detection and transparency mechanisms to optimally balance creative utility and ethical safeguards.

Conclusion

This research establishes that generative AI systems, when rigorously architected for contextual alignment and multimodal integration, reach and often surpass human advertising performance, both for personalized and universal persuasion strategies. While origin bias remains a measurable factor, its practical impact is diminishing as content sophistication escalates. The resulting paradigm shift towards a "post-authenticity" era in advertising foregrounds execution quality, narrative appeal, and coherence, with profound implications for both creative industries and consumer engagement frameworks. Societal and ethical governance structures must now confront the challenge of harnessing persuasive GenAI capabilities to benefit, not exploit, end-users.