The power of generative marketing: Can generative AI create superhuman visual marketing content?

International Journal of Research in Marketing The power of generative marketing: Can generative AI create superhuman visual marketing content?

ABSTRACT

Generative AI's capacity to create photorealistic images has the potential to augment human creativity and disrupt the economics of visual marketing content production. This research systematically compares the performance of AI-generated to human-made marketing images across important marketing dimensions. First, we prompt seven state-of-the-art generative text-to-image models (DALL-E three, Midjourney v six, Firefly two, Imagen two, Imagine, Stable Diffusion XL Turbo, and Realistic Vision) to create ten thousand three hundred twenty synthetic marketing images, using two thousand four hundred real-world, human-made images as input. Two hundred fifty-four thousand four hundred human evaluations of these images show that AI-generated marketing imagery can surpass human-made images in quality, realism, and aesthetics. Second, we give identical creative briefings to commissioned human freelancers and the AI models, showing that the best synthetic images also excel in ad creativity, ad attitudes, and prompt following. Third, a field study with more than one hundred seventy-three thousand impressions demonstrates that AI-generated banner ads can compete with professional human-made stock photography, achieving an up to fifty percent higher click-through rate than a human-made image. Collectively, our findings suggest that the paradigm shift brought about by generative AI can help advertisers produce marketing content not only faster and orders of magnitude cheaper but also at superhuman effectiveness levels with important implications for firms, consumers, and policymakers. To facilitate future research on AI-generated marketing imagery, we release GenImageNet that contains all of our synthetic images and their human ratings.

One. Introduction

Generative AI fundamentally disrupts the marketing industry, representing a new paradigm of automated marketing content generation. Industry reports suggest a tremendous economic potential of generative AI, quantifying its impact at United States dollars four hundred sixty-three billion in the marketing sector alone. Both marketing practice and research report astonishing anecdotal examples of generative AI's disruptive possibilities. Encouraged by such promising prospects, some firms have already successfully piloted synthetic content created by generative AI in their marketing campaigns, e.g., the award-winning "A.I. Ketchup" campaign by Heinz, which garnered more than eight hundred fifty million earned impressions around the globe.

Given the considerable excitement around generative AI, it is not surprising that firms have started exploring and experimenting with this novel technology. Industry forecasts project that large organizations will synthetically generate up to a third of their outbound marketing messages by twenty twenty-five. However, the sustainable adoption of generative AI by firms critically hinges on generative AI's effectiveness in reaching their marketing objectives and its efficiency, namely, in realizing substantial cost savings. Pioneering studies demonstrate the productivity gains and increase in output quality enabled by generative AI for automated marketing text generation. Preeminent studies outside of marketing corroborate these generative AI-enabled improvements with tangible economic benefits. However, due to the recency of the "age of generative AI" and idiosyncratic challenges pertaining to image creation, little is known about its disruptive potential for visual marketing content across diverse marketing contexts.

A better understanding of AI-generated marketing imagery's effectiveness and efficiency is important as images are a cornerstone of today's marketing communications in an increasingly media-rich environment. Firms and their ad agencies carefully design online and offline ads, influencers get paid to endorse brands across visual social media channels, online shops present products and services in the best possible conditions, consumers share their everyday consumption experiences online, and their digital traces offer a wealth of information for brand managers to visually "listen in." How do consumers perceive and respond to synthetic images compared to human-made content? How does AI-generated marketing imagery perform in a real-world context? If generative AI could create human-level visual marketing content, it could fundamentally challenge traditional human-made marketing content generation and accelerate AI adoption.

The importance of generative AI's role for the future of marketing is underscored by the substantial cost associated with creating professional marketing imagery, especially when considering large-scale, global marketing campaigns, which can require hundreds of visual assets tailored to different communication channels and target audiences. Take the following examples: Purchasing a professional stock photo typically costs around United States dollars five to ten, excluding additional expenses to acquire more permissive usage licenses. Opting for an experienced freelancer from an online marketplace to create a custom marketing image can increase the cost by an order of magnitude to around United States dollars one hundred. Employing top-tier ad agencies or organizing professional photo shoots, which involves specialized photographers and cast photo models, can even result in expenses ranging from thousands to tens of thousands of United States dollars. In contrast, generating a single image with OpenAI's DALL-E three, a state-of-the-art generative text-to-image model, costs merely United States dollars zero point zero four.

What if generative AI could substantially lower the expenses associated with the time-consuming and cost-intensive process of creating marketing imagery without compromising the content's visual appeal and marketing effectiveness? Is a prompt consisting of a couple of words and the right AI model all an advertiser needs? Considering that most methods are developed in computer science as general-purpose AI tools without specific optimization for marketing applications, it is unclear if state-of-the-art generative text-to-image models can generate effective marketing content that resonates with consumers when used off the shelf. Similarly, there is a lack of scientific evidence on which AI models provide consistent performance across marketing applications.

To systematically address this research gap, we conduct three studies. First, we investigate the perceptual evaluation of AI-generated versus human-made marketing images. Study one draws on eight different real-world marketing datasets, covering a comprehensive set of marketing applications, structured by the source of the data (firms versus users) and the marketing objective (call to action versus convey brand identity). We prompt seven state-of-the-art generative text-to-image models, released between October tenth, twenty twenty-three and February first, twenty twenty-four, namely, DALL-E three, Midjourney version six, Firefly two, Imagen two, Imagine, Stable Diffusion XL Turbo, and Realistic Vision to generate ten thousand three hundred twenty synthetic images, using two thousand four hundred real-world, human-made images as input. Two hundred fifty-four thousand four hundred human evaluations of these images, combined with algorithmic aesthetics assessments, show that AI-generated marketing imagery can surpass human-made images in quality, realism, and aesthetics.

Second, we give identical creative briefings to commissioned human freelancers and the same AI models, mimicking a real-world advertising pretest. We evaluate the perception, attitude, behavioral intention, and prompt following of the AI models and human freelancers in a between-subjects design across ten dependent variables. Overall, DALL-E three produces the best synthetic images, outperforming the human freelancers in terms of five marketing metrics, and obtaining directionally higher evaluations across the other five. Strikingly, participants attribute higher ad creativity to the AI-generated images by DALL-E three compared to the human-made freelancer images. In addition, AI-generated images are substantially more cost-efficient. The same budget of a single freelancer image allows for creating two thousand five hundred images with DALL-E three.

Third, we run a real-world marketing campaign on an online marketing platform to analyze the actual effectiveness of AI-generated banner ads in terms of their click-through rates. We collect and evaluate over one hundred seventy-three thousand impressions to compare the synthetic images with a high-quality, human-made stock photo selected by an online marketing professional. DALL-E three, the best-performing AI model, achieves an over fifty percent higher click-through rate than the human-made banner ad, while being two hundred twenty-five times cheaper to create. DALL-E three's click-through rate significantly outperforms the least effective AI model, Stable Diffusion XL Turbo, by one hundred percent.

This research makes three important contributions. First, we shed light on the real-world marketing effectiveness of AI-generated versus human-made images. While nascent marketing research demonstrates generative AI's effectiveness for textual content generation, this research is among the first to demonstrate superhuman perceptual evaluations and marketing effectiveness of synthetic marketing imagery across a comprehensive set of marketing applications and generative text-to-image models. Thereby, we shed light on the new paradigm of generative marketing-using generative AI to automate or assist marketing activities-which will likely fundamentally change the creation of marketing content in the future.

Second, our findings deepen the understanding of the human perception of AI-generated content. Studies on human perception of advertising content have a long tradition in the marketing literature. However, due to the recent advent of generative text-to-image models, little is known with respect to consumer perception of synthetic marketing imagery. Are AI-generated images only more cost-efficient to produce, or can they attain human-made images' perceptual evaluations? Our study demonstrates that AI-generated images can exceed human quality and aesthetic levels. An AI model specialized in photorealistic images, namely, Realistic Vision, can even create synthetic marketing imagery that humans perceive as more realistic than real images, which is in line with recent findings on "AI hyperrealism." In addition, we explore which visual features can explain differential perception of AI-generated imagery. For example, we observe a negative association between the color saturation and all three perceptual dimensions (quality, realism, and aesthetics).

Third, the present paper adds to the rich body of comparative method studies in marketing. Despite the remarkable performance of all AI models, we find that model choice matters. While DALL-E three and Midjourney version six constantly rank among the winning models, Stable Diffusion XL Turbo provides inferior performance compared to other AI models and the human-made benchmark images across almost all applications.