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 3, Midjourney v6, Firefly 2, Imagen 2, Imagine, Realistic Vision, and Stable Diffusion XL Turbo) to create 10,320 synthetic marketing images, using 2,400 real-world, human-made images as input. 254,400 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 173,000 impressions demonstrates that AI-generated banner ads can compete with professional human-made stock photography, achieving an up to 50% 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.

Full text

The power of generative marketing:
Can generative AI create superhuman visual marketing content?
Jochen Hartmann1,∗
Yannick Exner2,∗
Samuel Domdey3
1Jochen Hartmann is Professor of Digital Marketing at the Technical University of Munich, TUM
School of Management, Arcisstr. 21, 80333 M¨unchen, Germany. jochen.hartmann@tum.de
2Yannick Exner is Doctoral Researcher at the Technical University of Munich, TUM School of
Management, Arcisstr. 21, 80333 M¨unchen, Germany. yannick.exner@tum.de
3Samuel Domdey is Graduate Student at the Technical University of Berlin, Straße des 17. Juni
135, 10623 Berlin, Germany. s.domdey@campus.tu-berlin.de
∗These authors contributed equally to this work and share first authorship.

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 market-
ing images across important marketing dimensions. First, we prompt seven state-of-the-art
generative text-to-image models (DALL-E 3, Midjourney v6, Firefly 2, Imagen 2, Imagine,
Realistic Vision, and Stable Diffusion XL Turbo) to create 10,320 synthetic marketing im-
ages, using 2,400 real-world, human-made images as input. 254,400 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 commis-
sioned 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 173,000 impressions demonstrates that AI-generated banner ads can compete with pro-
fessional human-made stock photography, achieving an up to 50% 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.
Keywords: generative AI, marketing effectiveness, productivity, content creation, artificial
intelligence, digital marketing
September 5, 2024

1. Introduction
Generative AI fundamentally disrupts the marketing industry, representing a new paradigm
of automated marketing content generation (Peres et al., 2023). Industry reports suggest a
tremendous economic potential of generative AI, quantifying its impact at USD 463 billion
in the marketing sector alone (Chui et al., 2023). Both marketing practice and research re-
port astonishing anecdotal examples of generative AI’s disruptive possibilities (Kelly, 2023;
Noy and Zhang, 2023). Encouraged by such promising prospects, some firms have already
successfully piloted synthetic content created by generative AI in their marketing campaigns
(Acar and Gvirtz, 2024), e.g., the award-winning “A.I. Ketchup” campaign by Heinz, which
garnered more than 850 million earned impressions around the globe (The One Club, 2023).
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 2025 (Gartner, 2024). However, the sustainable adoption of genera-
tive AI by firms critically hinges on generative AI’s effectiveness in reaching their marketing
objectives (Jansen et al., 2024) and its efficiency, namely, in realizing substantial cost savings
(Ammanath et al., 2024; Gartner, 2024). Pioneering studies demonstrate the productivity
gains and increase in output quality enabled by generative AI for automated marketing text
generation (e.g., Reisenbichler et al., 2022, 2023). Preeminent studies outside of marketing
corroborate these generative AI-enabled improvements with tangible economic benefits (e.g.,
Noy and Zhang, 2023; Brynjolfsson et al., 2023). However, due to the recency of the “age
of generative AI” (Krugmann and Hartmann, 2024) and idiosyncratic challenges pertain-
ing to image creation (Borji, 2023), 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 increas-
ingly media-rich environment (Grewal et al., 2021). Firms and their ad agencies carefully
2

design online and offline ads (Pieters and Wedel, 2004; Hartmann et al., 2021), influencers
get paid to endorse brands across visual social media channels (Beichert et al., 2024), online
shops present products and services in the best possible conditions (Dzyabura et al., 2023;
Zhang et al., 2022b), consumers share their everyday consumption experiences online (Li
and Xie, 2020; Zhang and Luo, 2023), and their digital traces offer a wealth of information
for brand managers to visually “listen in” (Liu et al., 2020; Dzyabura and Peres, 2021). 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 tra-
ditional 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 as-
sets tailored to different communication channels and target audiences (King, 2024). Take the
following examples: Purchasing a professional stock photo typically costs around USD 5−10,
excluding additional expenses to acquire more permissive usage licenses. Opting for an ex-
perienced freelancer from an online marketplace to create a custom marketing image can
increase the cost by an order of magnitude to around USD 100. Employing top-tier ad agen-
cies 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 USD (Rodgers, 2021). In contrast, generating a single image with OpenAI’s DALL-E 3, a
state-of-the-art generative text-to-image model, costs merely USD .04 (Betker et al., 2023).1
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
1See https://openai.com/pricing (accessed September 2, 2024)
3

of words and the right AI model2all an advertiser needs? Considering that most methods
are developed in computer science as general-purpose AI tools without specific optimization
for marketing applications (Bommasani et al., 2021; Dzyabura et al., 2022), 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 investi-
gate the perceptual evaluation of AI-generated vs. human-made marketing images. Study
1 draws on eight different real-world marketing datasets, covering a comprehensive set of
marketing applications, structured by the source of the data (firms vs. users) and the mar-
keting objective (call to action vs. convey brand identity). We prompt seven state-of-the-art
generative text-to-image models, released between October 10, 2023 and February 1, 2024,
namely, DALL-E 3, Midjourney v6, Firefly 2, Imagen 2, Imagine, Realistic Vision, and Sta-
ble Diffusion (SD) XL Turbo to generate 10,320 synthetic images, using 2,400 real-world,
human-made images as input. 254,400 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 (MacKenzie et al., 1986). 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
(N = 1,575 Prolific panelists). Overall, DALL-E 3 produces the best synthetic images,
outperforming the human freelancers in terms of five marketing metrics, and obtaining di-
rectionally higher evaluations across the other five. Strikingly, participants attribute higher
2We use the terms “AI model” and “generative text-to-image model” interchangeably to refer to sys-
tems that make use of AI to generate images based on textual descriptions. Similarly, “AI-generated” and
“synthetic” images are used synonymously (see also https://deepmind.google/discover/blog/identifying-ai-
generated-images-with-synthid/; accessed April 4, 2024).
4

ad creativity to the AI-generated images by DALL-E 3 compared to the human-made free-
lancer images. In addition, AI-generated images are substantially more cost-efficient. The
same budget of a single freelancer image allows for creating 2,500 images with DALL-E 3.
Third, we run a real-world marketing campaign on an online marketing platform to an-
alyze the actual effectiveness of AI-generated banner ads in terms of their click-through
rates (CTRs). We collect and evaluate over 173,000 impressions to compare the synthetic
images with a high-quality, human-made stock photo selected by an online marketing pro-
fessional. DALL-E 3, the best-performing AI model, achieves an over 50% higher CTR than
the human-made banner ad, while being 225 times cheaper to create. DALL-E 3’s CTR
significantly outperforms the least effective AI model, SDXL Turbo, by 100%.
This research makes three important contributions. First, we shed light on the real-world
marketing effectiveness of AI-generated vs. human-made images. While nascent marketing
research demonstrates generative AI’s effectiveness for textual content generation (e.g., Carl-
son et al., 2023; Reisenbichler et al., 2022), this research is among the first to demonstrate
superhuman perceptual evaluations and marketing effectiveness of synthetic marketing im-
agery 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 gener-
ative 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 (e.g., MacKenzie et al., 1986; Pieters and Wedel, 2004). 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
5

create synthetic marketing imagery that humans perceive as more realistic than real images,
which is in line with recent findings on “AI hyperrealism” (Miller et al., 2023). 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 market-
ing (e.g., Hartmann et al., 2023; Krugmann and Hartmann, 2024). Despite the remarkable
performance of all AI models, we find that model choice matters. While DALL-E 3 and
Midjourney v6 constantly rank among the winning models, SDXL Turbo provides inferior
performance compared to other AI models and the human-made benchmark images across
almost all applications.
2. Related literature
2.1. The importance of marketing imagery
Images are a core component of contemporary marketing communications (Dzyabura
et al., 2023) and “worth a thousand words” (Li and Xie, 2020). Their persuasive power
is well documented in the field of marketing, explaining their widespread adoption across
diverse marketing contexts, such as online and offline advertising (Pieters and Wedel, 2004;
Hartmann et al., 2021), social media (Beichert et al., 2024; Li and Xie, 2020), online shopping
(Dzyabura et al., 2023; Zhang et al., 2022b), product design (Burnap et al., 2023), and visual
consumer reviews (Zhang and Luo, 2023). Furthermore, the abundance of visual data allows
brand managers to visually “listen in” and derive actionable insights to position their brands
(Liu et al., 2020; Dzyabura and Peres, 2021).
What explains the popularity and power of visual content in marketing communications?
Ample evidence supports the picture superiority effect, whereby images are remembered
better than words (Childers and Houston, 1984; Paivio and Csapo, 1973). Pieters and Wedel
(2004) find that the pictorial element of print ads is superior in capturing attention. Li and
6

Xie (2020) demonstrate a positive mere presence effect of image content on social media
engagement compared to text-only content. Their potential to shape consumers’ cognitive,
emotional, and behavioral responses makes images an appealing medium to convey a brand’s
visual identity in a memorable way and call consumers to action (Phillips et al., 2014).
However, producing visual content is a resource-intensive process. Consider the devel-
opment of a multi-modal digital ad consisting of a visual component and a corresponding
tagline. Each component demands a specialized skill set to create effective content. While
the textual tagline can be efficiently revised, the visual element adds a layer of complexity,
which arises not only from the aesthetic decisions involved in the creation process (Zhang
et al., 2022b) but also from the technical demands of dedicated image-editing software. What
if generative AI could support the efficient production of both synthetic textual marketing
content (e.g., Reisenbichler et al., 2022, 2023) and of visual marketing content?
2.2. Generative AI-enabled productivity gains and cost savings
Both within and outside of the marketing context, there is a growing interest in studying
generative AI-enabled productivity gains and cost savings. Brynjolfsson et al. (2023) show
that access to a generative AI-based conversational assistant can increase customer support
agents’ productivity by 14%, on average, with even larger benefits for novice and low-skilled
workers. For search engine optimization (SEO) content generation, Reisenbichler et al. (2022)
demonstrate that large language models can produce significantly more effective text than
human SEO experts, quasi-experts, and novices while simultaneously incurring a cost benefit
of 91%. Similarly, Reisenbichler et al. (2023) show that machine-written ad content can
increase the production efficiency of search engine advertising (SEA) by more than 60%.
Beyond marketing, Dell’Acqua et al. (2023) demonstrate for 18 realistic consulting tasks
that consultants with GPT-4 access, on average, completed 12.2% more tasks and were
25.1% faster compared to a control group. Similar to the findings of Brynjolfsson et al.
(2023), consultants with below-average skills benefited more from generative AI support.
Peng et al. (2023) present evidence that software developers with access to GitHub Copilot,
7

an AI pair programmer, completed a programming task 55.8% faster than a control group
without access. Similarly, Zhou and Lee (2024) show that generative AI can increase human
creative productivity by 25% in the context of digital artworks.
Collectively, recent research provides converging evidence for the substantial productivity
gains and cost savings enabled by generative AI across various application contexts and data
modalities. These findings are in line with the priorities of the business sector, where “tactical
benefits such as improving efficiency/productivity (56%) and/or reducing costs (35%)” are
reported as primary objectives, according to a survey among 2,835 business and technology
leaders (Ammanath et al., 2024). In addition, there seems to be a growing consensus that
generative AI exerts an equalizing effect, narrowing the skill gap between higher- and lower-
ability (marketing) content creators (e.g., Zhou and Lee, 2024; Noy and Zhang, 2023; Zhang
et al., 2024; Brynjolfsson et al., 2023; Dell’Acqua et al., 2023). However, the adoption of
generative AI as a new technology is a function not only of efficiency but also of effectiveness
gains. Hence, an important question remains: How do consumers react to AI-generated
marketing content?
2.3. Consumer response to AI-generated content
Pioneering work demonstrates favorable consumer response to generative AI in the con-
text of textual online marketing content generation (Reisenbichler et al., 2022). Similarly,
Carlson et al. (2023) show that AI-generated content can be indistinguishable from human-
made content in the automated creation of online reviews. These technological capabilities
even enable the generation of personalized persuasion at scale (Matz et al., 2024) with im-
portant implications for society (Jakesch et al., 2023).
Research on consumer reactions to AI-generated visual content shows mixed results.
Challenging the prevailing assumption that artistic, creative tasks are reserved for human
intelligence (Feuerriegel et al., 2024), Zhang et al. (2024) demonstrate that AI-assisted artists
are more likely to receive positive reactions to their creative content. In contrast, Horton
et al. (2023) demonstrate that people devalue AI-generated art even if they cannot distinguish
8

it from human-made art. Jansen et al. (2024) show that consumer responses can inform an
automated alignment process, whereby a generative AI model is steered to convey certain
brand dimensions in its outputs. Also in the context of product design, AI-generated visuals
can be appealing to consumers (Burnap et al., 2023; Zhang et al., 2022a). However, AI-
assisted product designs can also backfire when used in the wrong context (Xu and Mehta,
2022).
3. Overview of studies
To explore consumer response to AI-generated visual marketing content across various
industries and application contexts, we conduct three studies. Table 1 summarizes the
objective and setup for each of them, exploring generative AI’s potential to rival human-
made content both in the lab and in the field (van Heerde et al., 2021).
Study 1 Study 2 Study 3
Research
question
How is AI-generated marketing
imagery perceived compared to
human-made images across a
broad range of real-world mar-
keting applications?
Given the same prompt, i.e.,
creative briefing, can generative
text-to-image models reach sim-
ilar performance across key mar-
keting outcomes vs. freelancers’
human-made imagery?
In a real-world A/B test, can AI-
generated images achieve sim-
ilar click-through rates to a
high-quality, human-made stock
photo selected by an online mar-
keting professional?
Dependent
variables
•Quality (Zhang et al., 2022b)
•Realism (Cho et al., 2014)
•Aesthetics (Talebi and Milan-
far, 2018)
•Perception (Zhang et al.,
2022b; Cho et al., 2014; Talebi
and Milanfar, 2018; Smith et al.,
2007)
•Attitude (Smith et al., 2007)
•Behavioral intention (Smith
et al., 2007; Rizzo et al., 2023)
•Prompt following (Saharia
et al., 2022)
Click-through rate (CTR)
Study
setup
Lab setting (within subjects):
•2,400 human-made images
•10,320 synthetic images
•2×10 ratings / image
(254,400 ratings in total)
Lab setting (between subjects):
•4 human-made images
•28 synthetic images
•∼50 participants / condition
(1,575 participants in total)
Field study:
•1 human-made image
•7 synthetic images
•∼100 clicks / condition
(173,022 impressions in total)
Table 1: Overview of studies
First, study 1 compares 10,320 AI-generated marketing images to 2,400 human-made im-
ages on three perceptual dimensions: quality, realism, and aesthetics. Study 2 gives identical
9

creative briefings to both the AI models and commissioned human freelancers, allowing us
to quantify image production costs and obtain a broader set of standard marketing metrics
in an advertising pretest setting. Specifically, we assess image performance along image per-
ception, attitude, behavioral intention, and prompt following, i.e., compliance of the created
images with the initial creative briefing. Building on studies 1 and 2, study 3 is designed
to investigate the real-world marketing effectiveness of AI-generated banner ads via an ac-
tual online marketing campaign featuring seven synthetic images and a human-made stock
image selected by an online marketing professional. As the behavioral response, we assess
each image’s CTR based on 173,022 impressions and 907 clicks. Collectively, the three stud-
ies comprehensively assess AI-generated visual marketing content based on over a quarter
million human evaluations for 10,355 synthetic marketing images.
All studies draw on the same set of seven state-of-the-art generative text-to-image models
that reflect a heterogeneous collection of AI models with differing characteristics such as
training data, source accessibility, and release date. Table 2 presents an overview of the AI
models included in our studies. Web Appendix A.1 describes the models from a technical
perspective. The earliest model we include in our benchmark dates back to October 10, 2023
(Adobe’s Firefly 2). The newest model is from February 1, 2024 (Google’s Imagen 2).
10

Model Developer Source accessibility Release date Batch size @ Resolution Scalable access Watermark
DALL-E 3 OpenAI Proprietary Oct 19, 2023 1@1,024 ×1,024 ✓(API) ∼(invisible C2PA)
Midjourney v6 Midjourney Proprietary Dec 21, 2023 4@1,024 ×1,024 ✓(via wrapper API) ✗
Firefly 2 Adobe Proprietary Oct 10, 2023 4@2,048 ×2,048 ✗ ✗ (only in free version)
Imagen 2 Google Proprietary Feb 01, 2024 4@1,536 ×1,536 ✗∼(invisible SynthID)
Imagine Meta Proprietary Dec 06, 2023 1@1,024 ×1,024 ✗ ✓ (see Web Appendix Figure A.2)
SDXL Turbo Stability AI Open source (Hugging Face) Nov 28, 2023 1@512 ×512 ✓(local hosting) ✗
Realistic Vision Community Open source (Civitai) Dec 23, 2023 1@512 ×512 ✓(local hosting) ✗
Note: DALL-E 3 offers an automated prompt refinement to improve the output. To treat all AI models equally, we followed OpenAI’s guidelines to reduce the impact
of automated prompt refinement. For details see: https://platform.openai.com/docs/guides/images/prompting (accessed March 13, 2024).
Table 2: Overview of state-of-the-art generative text-to-image models
11

4. Study 1: Perception of AI-generated vs. human-made marketing imagery
The objective of study 1 is to investigate if AI-generated images can achieve similar per-
ceptual ratings as human-made images in terms of quality, realism, and aesthetics across
common visual marketing applications. All three perceptual dimensions are frequently stud-
ied image characteristics in marketing research with important downstream consequences
(Zhang et al., 2022b; Li and Xie, 2020; Karpinska-Krakowiak and Eisend, 2024; Kim et al.,
2019). Moreover, they represent established evaluation measures for AI-generated images in
computer science (Saharia et al., 2022; Betker et al., 2023).
4.1. Method
To ensure comparability between the AI-generated and the human-made images, each
synthetic image is created based on a textual description of its underlying human-made
source image. Figure 1 illustrates the two-step image creation process: First, converting
human-made source images to text (green trapezoid). Second, generating synthetic sibling
images from text (purple trapezoid).3
As human-made source images, we utilize 2,400 images that differ by marketing use case
and visual composition. To comprehensively reflect prevalent use cases of marketing imagery
in our benchmark, we build on established work in marketing, which differentiates between
(a) the data source, i.e., firms vs. users (Dzyabura et al., 2022; Liu et al., 2020) and (b) the
marketing objective, i.e., conveying a firm’s brand identity as a long-term objective vs. calling
prospective customers to action as a short-term objective such as purchasing a product or
clicking on an ad (Keller, 1993; Keller and Lehmann, 2006).4This conceptual framework
3Some AI models generate a batch of images for each prompt, e.g., four images for Firefly 2 and Midjour-
ney v6. For these, we consistently sample the first / top left image. To verify the quality of the transformation
process, we assess the image-text alignment between the source image and the resulting textual representa-
tion as well as between the textual representation and the generated images for a sample of images, following
the protocol outlined by Google in Saharia et al. (2022). The resulting image-text alignment is in the range
of Google’s evaluations, validating our two-step image creation approach.
4The differentiation of the marketing objective in terms of performance marketing (short-term) vs. brand
building (long-term) is also a known “marketing dilemma” in the industry (Kyriakidi, 2022).
12

Textual representation
as prompt
Source image
(human-made)
Synthetic sibling
(AI-generated)
Image encoding
(image-to-text)
Image creation
(text-to-image)
CLIP
Interrogator
Generative
text-to-image
models
a bottle of Heinz's tomato
ketchup on a white surface
[…]
x7
Note: In the first step, we employ CLIP-Interrogator in a zero-shot manner via an API endpoint provided
by replicate (Ding et al., 2023), allowing us to transform the visual information stored in the human-made
source images into textual descriptions (Radford et al., 2021). In the second step, we generate seven synthetic
sibling images for each human-made source image using the CLIP-based textual description as a prompt for
the seven generative text-to-image models.
Figure 1: Image generation procedure in studies 1 and 3
results in a 2 ×2 matrix (see Figure 2), which structures marketing imagery’s most relevant
use cases into four quadrants.
For each of the four quadrants, we obtain two representative real-world datasets from
which we randomly sample 300 images each.5As acknowledged by (Dzyabura et al., 2022),
platforms and specific data sources constantly change, but the general characterization
of marketing datasets in terms of the two primary dimensions (firm-generated vs. user-
generated and conveying brand identity vs. calling to action) are likely to prevail. Hence,
the datasets we sample shall only be considered exemplary for each quadrant. Beyond en-
suring diversity in real-world applications and industry contexts, our conceptual framework
that guides our systematic data sampling also ensures a large heterogeneity in the human-
made images’ composition and characteristics. For a detailed overview of the datasets and
their provenance, see Web Appendix Table A.1.
To obtain large-scale perceptual ratings for the 2,400 human-made images and their
10,320 synthetic siblings, ten human raters on Amazon Mechanical Turk (MTurk) assessed
each image on 7-point Likert scales (1 = lowest, 7 = highest). Before human evaluation, each
image was resized to a standardized resolution with a minimum dimension of 512 pixels in
5For those models that we could not access programmatically, e.g., via an API endpoint, we sampled 30
human-made images per dataset, resulting in a total of 240 images per model for evaluation. This corresponds
to 10% of the data compared to the other models we evaluate on 2,400 images each.
13

Firm-generated content (FGC) User-generated content (UGC)
Convey brand identity Call to action
Instagram YelpAmazon Booking.com
Twitter UnsplashLions awards Image ads
Note: Each quadrant contains two representative datasets. All displayed images are AI-generated.
Figure 2: Overview of datasets
either width or height to avoid image distortions or cropping (Sauer et al., 2023). We adopt
the image quality scale from Zhang et al. (2022b): “Give a score to an image on a scale of 1-7
on its aesthetic quality where 1 is ‘very bad’ and 7 is ‘excellent’.”. In addition, we provide
the same detailed instructions as Zhang et al. (2022b) to ensure reliable ratings (see Web
Appendix Figure A.3). To assess perceived realism, we adopt a scale item from Cho et al.
(2014), which Karpinska-Krakowiak and Eisend (2024) also use in the context of deepfake
content: “The visual elements of the ad are realistic.”, anchored by “strongly disagree” and
“strongly agree”. Below the question, we define realistic as “accurately representing what is
natural or real” (see Web Appendix Figure A.3).6
This results in a total of 254,400 human ratings ((2,400 + 10,320) images ×2 ques-
tions/image ×10 raters/question). In addition, to enrich these human evaluations with an
algorithmic aesthetics assessment, we apply Neural Image Assessment (NIMA) to all our im-
6In study 2 we use the full six-item scale for perceived realism by Cho et al. (2014) (Cronbach’s α= .88).
14

ages. NIMA is a convolutional neural network-based classifier for human perception of image
aesthetics, defined as xN IM A ∈[1,10], where 10 is the highest score (Talebi and Milanfar,
2018). Web Appendix Figure A.4 plots the highest vs. lowest rated AI-generated images
across quality, realism, and aesthetics. The lowest-realism image, for example, features an
incoherent object representation, showing a baseball player with an ill-positioned arm.
4.2. Results
Perception
Dependent variables: Quality Realism Aesthetics
Model: (1) (2) (3)
AI models
DALL-E 3 -.0366∗∗∗ -.5273∗∗∗ .1602∗∗∗
(.0090) (.0103) (.0129)
Midjourney v6 .0925∗∗∗ -.1142∗∗∗ .0704∗∗∗
(.0090) (.0103) (.0129)
Firefly 2 .1564∗∗∗ -.0909∗∗∗ .3994∗∗∗
(.0228) (.0261) (.0326)
Imagen 2 .0156 -.2365∗∗∗ .2798∗∗∗
(.0227) (.0260) (.0326)
Imagine .0589∗∗ -.2229∗∗∗ .2880∗∗∗
(.0227) (.0260) (.0326)
SDXL Turbo -.1268∗∗∗ -.2945∗∗∗ .0983∗∗∗
(.0090) (.0103) (.00129)
Realistic Vision .0960∗∗∗ .0708∗∗∗ .2808∗∗∗
(.0090) (.0103) (.0129)
Fixed effects
Prompt Yes Yes Yes
Respondent Yes Yes No
Fit statistics
Observations 127,200 127,200 12,720
R2.3742 .4003 .5567
Within R2.0074 .0348 .0584
Note: Standard errors in parentheses. Human-made image as reference.
For aesthetics, NIMA provides a single deterministic prediction whereas
ten respondents rated each image’s quality and realism. This explains
the ten-fold difference in observations for Models (1-2) vs. Model (3).
***: p<.001, **: p<.01, *: p<.05
Table 3: Results for perception of AI-generated vs. human-made marketing imagery
Table 3 displays the OLS regression results for the three perceptual dimensions. To
account for unobserved differences across the prompts and human raters, we include fixed
15

effects for both. The low to medium correlations among the three dependent variables suggest
that they capture three distinct visual characteristics of the images (see Web Appendix
Figure A.5).
Regarding image quality, four out of the seven AI models significantly surpass the human-
made images, with Firefly 2 (βQuality =.1564, p < .001) generating the highest-quality
images. In contrast, SDXL Turbo obtains the lowest quality perception scores, exhibiting
inferior performance compared to the human benchmark (βQuality =−.1268, p < .001).
However, given that ratings are captured on 7-point Likert scales, with an overall mean
quality rating of 5.26, the effect sizes suggest only marginally lower quality evaluations
than the real-world marketing imagery. DALL-E 3’s images, for example, are rated less
than four hundredths of a Likert-scale point lower than the human-made benchmark images
(βQuality =−.0366, p<.001).
Realistic Vision lives up to its name, standing out as the only AI model that outperforms
the human-made images in terms of perceived realism (βRealism =.0708, p < .01). This
remarkable performance can be attributed to its specialized fine-tuning to generate highly
realistic images (see Web Appendix A.1 for details). In contrast, all other AI models fail to
achieve the realism level of human-made images.
Regarding aesthetic appeal, all seven AI models significantly surpass the human-made
marketing images (p<.001). Overall, Firefly 2 and Realistic Vision produce the images
with the highest aesthetic appeal.
4.3. Discussion
Study 1 demonstrates that the best AI models can outperform human-made marketing
visuals across three important perceptual dimensions: quality, realism, and aesthetics. Fur-
thermore, the results suggest that model choice matters and depends on the advertiser’s
objective. Regarding image quality and aesthetics, Firefly 2 emerges as the winning method.
In terms of perceived realism, Realistic Vision’s synthetic images are perceived as more real-
istic than real, human-made images. This observation is consistent with findings by Jakesch
16

et al. (2023), demonstrating that AI-generated text can be perceived as “more human than
human”. Similarly, Miller et al. (2023) document so-called “AI hyperrealism” for the percep-
tion of AI-generated human faces. For advertisers, these findings are important as realistic
product portrayals can facilitate consumers’ mental simulation of product consumption or
usage (Kim et al., 2019), which in turn can translate into purchase intention (Ceylan et al.,
2024).
In terms of the mechanism — that is, the images’ visual features correlating with differ-
ential perceptual assessments — our analysis identifies several visual ingredients that shape
consumers’ perception of AI-generated marketing imagery (see Web Appendix Table A.3 for
details). For example, excessive color saturation appears to diminish favorable consumer
responses. Similarly, content creators need to exercise caution when prompting AI models
to generate text or human faces, as these elements are prone to visual imperfections, which
in turn can hamper the synthetic images’ appeal.
Next, study 2 addresses three limitations of study 1. First, study 1 assesses only percep-
tual dimensions, which we expand in study 2 to a broad set of marketing metrics commonly
used advertising pretests, including ad attitudes and purchase intentions (MacKenzie et al.,
1986). Second, study 2 simplifies image creation through an alternative prompting proce-
dure. Instead of the two-step pipeline (first, converting images to text; second, generating
images from text; Figure 1), both AI models and commissioned human freelancers receive
the same creative briefing, streamlining the process to a text-to-image task. Third, we could
not observe the production costs of the human-made images in study 1. Study 2 allows us
to obtain these costs and compute a back-of-the-envelope calculation on the cost savings for
generative AI vs. human labor for visual content creation.
5. Study 2: Generative AI vs. human freelancers
Study 2 is designed to investigate how AI models perform compared to experienced
human freelancers. To set the generative AI models and the designers on equal footing, we
17

instruct both with the identical creative briefing and do not share any additional information.
Figure 3 displays a schematic overview of the image creation process. Compared to study 1,
this approach is significantly more costly and less scalable. However, the smaller number of
images allows us to obtain consumer responses to a broader range of important marketing
metrics, thereby mimicking a “diagnostic pretesting” setting (MacKenzie et al., 1986).
Large outdoor advertising
banner, including "Gives
You Wings" slogan […]
Briefing/Prompt
•Firm-generated content
•Convey brand identity
Marketing application Output
Human
freelancer
Generative
text-to-image
models
Image creation
x4
x7
Note: For complete briefing/prompt, see Figure 4. The blue circles indicate that we use seven generative
text-to-image models to obtain AI-generated sibling images for four different marketing applications.
Figure 3: Image generation procedure in study 2
5.1. Method
Guided by our framework introduced in study 1 (see Figure 2), we define one creative
briefing for each quadrant of our 2 ×2 matrix covering prevalent marketing imagery ap-
plications (marketing objectives: call to action vs. convey brand identity; data source:
firm-generated content (FGC) vs. user-generated content (UGC)). Specifically, we include a
marketing decal showcasing a new limited edition product (i.e., FGC; call to action), a large
outdoor banner ad (i.e., FGC; convey brand identity), a consumer brand selfie (i.e., UGC;
call to action), and a consumer action shot with merchandise (i.e., UGC; convey brand
identity). We choose “Red Bull” as an exemplary brand due to its frequent examination
in marketing research (e.g., Seiler et al., 2021), being renowned for its “buzz marketing”
18

(Steenkamp et al., 2010). Figure 4 presents the creative briefings alongside the AI-generated
images by DALL-E 3 and human-made images.7
AI-generated
Convey brand identity Call to action
Decal showcasing new limited-edition RedBull flavor,
energetic green colors, highlighting iconic can design with a
twist representing the new mint flavor, engaging call to
action, digital illustration, photorealistic detailing,
attention-grabbing, high resolution.
First-person perspective picture, one hand holding RedBull
can, blurry beach backdrop with a palm tree, laidback chill
vibe, bright sunlight, inspiring engagement, photorealistic,
high resolution, clear focus on can.
Large outdoor advertising banner, including "Gives You
Wings" slogan, Formula 1 setting, RedBull race car in front
of the banner on the street, car captured from the side,
high resolution design with a balanced color contrast,
avoiding not hyper-realistic appearance
Instagram-style image, consumer skydiving with RedBull
logo parachute, thrilling adventure, freedom sense, clear
sky background, high resolution, photorealistic, dynamic
action pose, vibrant colors, detailed equipment.
Firm-generated content (FGC) User-generated content (UGC)
Human-made AI-generated Human-made
Note: DALL-E 3 created the AI-generated images (left). Freelancers created the human-made images (right).
Figure 4: Overview of four creative briefings and resulting images
To obtain human-made images from experienced human freelancers, we created an indi-
vidual request for proposal (RFP) for each creative briefing on Freelancer.com and waited
until we received more than 50 bids. Based on expert input from an independent freelancer
unaware of our hypothesis, we defined a budget range of USD 30 to USD 250 per image. Be-
sides the creative briefing, the RFP included additional information, such as two statements
7The creative briefings are formulated in a similar style to standard prompts, ensuring they are compatible
with all seven AI models without modifications and human freelancers can easily understand them.
19

that prohibit the use of any generative AI tools8as well as the desired quadratic aspect ratio
of the output (see Web Appendix Figure A.6 for one of the four RFPs used to brief the
freelancers). Among the bids, we then filtered for freelancers that (a) commit in writing not
to use any generative AI tools, (b) have a rating of 4.5 out of 5 stars or higher, (c) have a
bid above the average bid, and (d) are verified by Freelancer.com. This systematic selection
results in three unique freelancers offering their service at USD 100 per image. Their profiles
state hourly rates ranging from USD 15 to USD 30.
Based on the same creative briefings shared with the human freelancers, we prompt all
generative text-to-image models. This approach results in a total of 32 images where four are
human-made and 28 are AI-generated (i.e., 7 AI models ×4 images). We use these images
as stimuli for a between-subjects experiment on Prolific, where each participant is exposed
to only one of the 32 images. We collect ten dependent variables, which can be subsumed
into the following four groups:
•Perception via quality,realism,aesthetics, and ad creativity
•Attitude via ad attitude and brand attitude
•Behavioral intention via purchase intention and social media engagement
•Prompt following via image-text alignment and brand recognition
All questions are rated on 7-point Likert scales (1 = lowest, 7 = highest), using established
multi-item scales. Web Appendix Table A.5 gives a detailed overview of these scales and
respective references from the marketing literature.
5.2. Results
1,575 of the 1,604 Prolific panelists pass all three attention checks, resulting in an average
of 49 participants per condition (Mage = 43.35 years, 50.16% women). Table 4 presents the
OLS regression results.
8This restriction on AI use mirrors current trends in the formulation of agency contracts (Sloane, 2024).
20

Perception Attitude Behavioral intention Prompt following
Dependent variables: Quality Realism Aesthetics Ad creativity Ad attitude Brand attitude Purchase intention Engagement Alignment Brand recognition
Model: (1) (2) (3) (4) (5) (6) (7) (8) (9) (10)
AI model
DALL-E 3 .6462∗∗∗ .0854 .6273∗∗∗ .4198∗∗ .3763∗.0939 .0078 .1313 .5607∗∗∗ .0050
(.1422) (.1071) (.0330) (.1446) (.1641) (.1566) (.1879) (.1870) (.1547) (.0280)
Midjourney v6 .6335∗∗∗ .3843∗∗∗ .4891∗∗∗ .1667 .3901∗.2070 .2153 .1186 .1540 -.0519
(.1421) (.1070) (.0330) (.1445) (.1640) (.1565) (.1878) (.1869) (.1546) (.0280)
Firefly 2 .1829 -.3178∗∗ .5515∗∗∗ .0560 .0447 -.0279 .1931 .2912 -1.371∗∗∗ -.9548∗∗∗
(.1449) (.1091) (.0336) (.1474) (.1672) (.1596) (.1915) (.1906) (.1577) (.0285)
Imagen 2 .2347 .3965∗∗∗ .6277∗∗∗ -.2490 .0963 -.0954 -.2010 -.1854 .3862∗-.0479
(.1431) (.1077) (.0332) (.1455) (.1651) (.1576) (.1891) (.1882) (.1557) (.0282)
Imagine .0897 -.0002 .3433∗∗∗ -.0855 -.0758 -.0503 -.2597 -.1829 .2047 -.0254
(.1419) (.1068) (.0329) (.1443) (.1637) (.1562) (.1875) (.1866) (.1544) (.0279)
SDXL Turbo -.8379∗∗∗ -.6207∗∗∗ .4900∗∗∗ -.5582∗∗∗ -.8801∗∗∗ -.7553∗∗∗ -.5985∗∗ -.2199 -1.396∗∗∗ -.6751∗∗∗
(.1428) (.1076) (.0331) (.1453) (.1649) (.1573) (.1888) (.1879) (.1555) (.0281)
Realistic Vision -.7618∗∗∗ -.3223∗∗ .6665∗∗∗ -.7357∗∗∗ -.8831∗∗∗ -.5189∗∗ -.2696 -.1630 -.8023∗∗∗ -.3196∗∗∗
(.1431) (.1077) (.0332) (.1455) (.1651) (.1576) (.1891) (.1882) (.1557) (.0282)
Fixed effects
Prompt Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes
Fit statistics
Observations 1,575 1,575 1,575 1,575 1,575 1,575 1,575 1,575 1,575 1,575
R2.1436 .1278 .5559 .1136 .0992 .1002 .1622 .0730 .2396 .6227
Within R2.1314 .0939 .2797 .0720 .0938 .0987 .1565 .0707 .1933 .6066
Note: Standard errors in parentheses. Human-made image as reference. All regression models control for the participants’ brand familiarity with Red Bull.
***: p<.001, **: p<.01, *: p<.05
Table 4: Results for perception, attitude, behavioral intention, and prompt following of AI-generated vs. human-made marketing imagery
21

Overall, DALL-E 3 produces the best synthetic images, significantly outperforming the
human freelancers in terms of five marketing metrics, and obtaining directionally higher
evaluations across the other five. Midjourney v6 ranks second, surpassing the freelancers on
four evaluation criteria, and being directionally better on the other ones, except for brand
recognition. The open-source models, SDXL Turbo and Realistic Vision, exhibit the worst
performance and are almost consistently inferior to the freelancers’ marketing imagery.
Regarding perceptual dimensions, participants rate the aesthetic appeal of all AI-generated
images significantly higher than the freelancers’ visuals, which aligns with our observation
in study 1. Also for quality and realism, the best AI models outperform the human-made
images (DALL-E 3 and Midjourney v6 for quality, and Midjourney v6 and Imagen 2 for
realism). DALL-E 3 is the only model that obtains significantly higher assessments of ad
creativity (βAdCreativity =.4198, p<.01), while SDXL Turbo and Realistic Vision exhibit
significantly lower ad creativity compared to the freelancers (βAdC reativity =−.5582, p<.001
and βAdCreativity =−.7357, p<.001, respectively)
In terms of attitudinal response and behavioral intentions, DALL-E 3 and Midjour-
ney v6 are the only two AI models that obtain directionally higher assessments compared to
the freelancers, outperforming the human-made images significantly regarding ad attitudes
(βAdAttitude =.3763, p<.05, and βAdAttitude =.3901, p<.05, respectively).
Lastly, DALL-E 3 obtains the highest text-to-image alignment (βAlignment =.5607,
p < .001), capturing participants’ response to the question: “How accurately does the cap-
tion describe the above image?”. In other words, the best AI model manages to generate
images that adhere better to the creative briefing than the commissioned human freelancers.
Imagen 2 ranks second, also obtaining a significantly higher image-text alignment than the
human-made images (βAlignment =.3862, p < .001). These findings are plausible as both
DALL-E 3 and Imagen 2 are trained on enhanced, synthetic image-text pairings to improve
their prompt following capability (see Web Appendix A.1).
The open-source models, SDXL Turbo and Realistic Vision, as well as the proprietary
22

Firefly 2 are inferior to the human-made images, both in terms of image-text alignment
and brand recognition. This is plausible as Firefly 2 does not generate brand logos at all
and the open-source models generate Red Bull logos, which tend to be distorted or highly
corrupted, making participants are less likely to correctly recognize them as Red Bull (see
Web Appendix Figure A.7 for example images).
5.3. Quantifying generative AI-enabled cost savings compared to human freelancers
Observing the fees paid to the commissioned human freelancers allows us to compare
these with the production costs incurred by the AI models. Specifically, we compute a back-
of-the-envelope calculation to determine the cost per image for each of the seven AI models
and the freelancers. For details on the underlying input parameters and assumptions, refer
to Web Appendix Table A.6.
For those AI models offered via an API endpoint with a transparent pricing scheme the
cost per image is straightforward to obtain, e.g., for DALL-E 3 at USD .04 per image or
Imagen 2 at USD .02 per image. Other AI models are priced based on the usage of GPU
hours or packaged image credits, which translate into costs of USD .07 for Midjourney v6
and USD .05 for Firefly 2. For the open-source models, SDXL Turbo and Realistic Vision,
which can be self-hosted on local or cloud Graphics Processing Units (GPUs), we obtain
the lowest costs at USD .00005 and USD .00026 per image, respectively. In other words,
creating a single image with SDXL Turbo costs only one two-hundredth of a cent ( 5
1000 =
1
200 ). This is reasonable as SDXL Turbo is optimized for rapid inference through adversarial
diffusion distillation, enabling image creation in just a single diffusion step (Sauer et al.,
2023). The ability to create nearly 1,000 synthetic images with SDXL Turbo for the same
budget compared to generative text-to-image models with an API endpoint such as DALL-
E 3, highlights the substantial cost variations among AI models. Note in addition to their
cost advantage and inference speed (Sauer et al., 2023), open-source software and models
are appealing in terms of control and customizability (Bonaccorsi and Rossi, 2003).
The cost-efficiency advantage of AI models over human freelancers is even stronger. For
23

the human freelancers’ images, we paid an average USD 100 per image. In contrast, the
same budget allows for the creation of 2,500 images with DALL-E 3 or two million images
with SDXL Turbo, showcasing the disruptive efficiency gains offered by generative AI for
the automated creation of visual marketing content.
5.4. Discussion
Study 2 investigated the effectiveness and efficiency of AI-generated marketing imagery
compared to commissioned human freelancers’ work in a between-subjects setting, mimicking
an advertising pretest. DALL-E 3 emerges as the winning generative text-to-image model,
significantly outperforming the human-made visuals in 50% of the marketing metrics, with
directionally higher ratings on the other 50%. Notably, DALL-E 3 is the only AI model
to achieve significantly higher ad creativity ratings compared to the freelancers’ visuals.
Midjourney v6 ranks as the second most effective AI model. The findings align with the
disruptive impact of generative AI on creative freelancer work, as suggested by Hui et al.
(2023). While not without flaws, generative AI’s output is often perceived as comparable to,
if not better than, human-made content, but at a substantially lower cost. Strikingly, the AI
models generated images that closely matched Red Bull’s iconic design elements, such as the
distinctive shape of the can and the brand’s characteristic color palette (see Figure 4). This
suggests that even without brand-specific fine-tuning the AI models had sufficient examples
within their extensive training datasets to learn and replicate Red Bull’s visual language.
A common issue for both developers and users of generative AI is ensuring that the
AI model follows instructions precisely, producing outputs that match the given prompts
(Betker et al., 2023). Study 2 suggests that this concern may extend to commissioned
human freelancers who sometimes deviate from the creative briefing provided to them. For
example, despite the request for a mint-flavored limited edition Red Bull can, one freelancer
created a lime-flavored design, see Figure 4. Similarly, another freelancer was instructed
to create a “photorealistic” and “high resolution” brand selfie but delivered an illustration
instead. These instances highlight that even careful freelancer selection does not guarantee
24

flawless results and that some images might be more complex to create for humans than
others. While iterations with the freelancers could have corrected some of these flaws, we
did not do so in order to maintain a consistent one-shot creation process for both the AI-
generated and the human-made images. Note that had we opted for iterations, we could
have iterated the marketing visuals much faster with the AI models than with the human
freelancers.
Next, study 3 focuses on the real-world performance of AI-generated banner ads in a
field study designed to collect additional evidence on the effectiveness and efficiency gains
enabled by generative AI.
6. Study 3: The real-world effectiveness of AI-generated banner ads
To increase the ecological validity of our findings (van Heerde et al., 2021; Hulland and
Houston, 2021), study 3 systematically assesses the real-world effectiveness of synthetic im-
ages compared to a professional human-made stock photo. Specifically, we measure perfor-
mance in terms of the AI-generated banner ads’ CTR, i.e., clicks divided by impressions.
6.1. Method
For the field study, we collaborated with an education provider specializing in online
marketing courses. To mimic a real-world marketing campaign as closely as possible, we
co-designed the experiment with the education provider’s CEO, an online marketing expert
holding a Ph.D. in marketing with 25 years of industry experience.
First, the online marketing expert selected and purchased the human-made stock photo
in line with the education platform’s marketing objective. The image displays two hands
holding two puzzle pieces against a horizon at sunset (see Figure 5). Note that hands are
considered notoriously difficult to create for generative AI models (Chayka, 2023). Hence,
the online marketing expert did not choose an easy baseline for the generative text-to-
image models to compete with. Moreover, the selected stock photo originates from an
25

iStock account that offers more than 4,500 images in its portfolio (oatawa, 2024), suggesting
professional experience in the creation of stock photography.
To generate seven synthetic sibling images, one per AI model, we followed our validated
image creation approach from study 1 (see Figure 1). All banner ads had the same title
(“Understand online marketing with [company name].”) and a detailed description (“Learn
to piece together the components for your online marketing success! With [company name]
- your expert for online marketing training. We train on over 70 online marketing topics in
small groups!”).
Firefly 2
Realistic Vision
DALL-E 3
Imagine
Midjourney v6
SDXL Turbo
Stock Photo
Imagen 2
Note: CLIP-Interrogator transformed the professional stock photo into the following prompt that served as
input to all seven generative text-to-image models: “a person holding two pieces of a puzzle, a stock photo
by [artist], shutterstock contest winner, objective abstraction, stockphoto, stock photo, congruent”
Figure 5: AI-generated banner ads and professional stock photo for field study
To ensure fair competition between the banner ads, we chose Meta’s online marketing
platform which offers a sophisticated A/B testing functionality.9As Meta restricts the num-
ber of ads that can be included in a single campaign when the A/B testing functionality is
9See https://www.facebook.com/business/ads/ab-testing for details (accessed March 14, 2024).
26

activated, we created two concurrent A/B testing campaigns that both include the profes-
sional stock photo as the human-made benchmark. Both campaigns had identical campaign
settings, targeting an audience in the 25-55 age group interested in marketing. In addition,
we set “traffic” as the campaign objective to avoid the unobservable Meta algorithm opti-
mizing the ads’ distribution on our dependent variable, namely, clicks. We set a daily budget
of $10 per condition and aimed for an expected 100 clicks per condition, i.e., 900 clicks in
total. Both campaigns started on February 27, 2024, and ran until March 2, 2024.
6.2. Results
The campaigns cost a total of $449.91, generating 173,022 impressions and 907 clicks.
This translates into an average CTR of .52%. Table 5 reports the CTRs for all banner ads.
Rank Model CTR Impressions Clicks CPC Spend Campaign
1 DALL-E 3 .80% 16,579 133 $.38 $49.99 A
2 Midjourney v6 .54% 19,310 105 $.48 $49.99 A
3 Imagine .54% 19,851 107 $.47 $49.99 A
4 Stock Photo .53% 18,531 98 $.51 $49.99 A
5 Stock Photo .52% 19,606 101 $.49 $49.99 B
6 Imagen 2 .51% 19,170 97 $.52 $49.99 B
7 Firefly 2 .49% 19,612 97 $.52 $49.99 B
8 Realistic Vision .43% 20.348 88 $.57 $49.99 B
9 SDXL Turbo .40% 20,015 81 $.62 $49.99 A
Total 173,022 907 $449.91
Table 5: CTRs of professional, human-made stock photo vs. AI-generated banner ads
DALL-E 3 generated the best-performing banner ad, obtaining a CTR of .80% and sig-
nificantly outperforming the professional stock photo within the same A/B testing campaign
by more than 50% (CTRStockPhoto =.53%; χ2(1, N = 35,110) = 9.5915, p < .01). Given
the same budget for both conditions of USD 49.99 this results in an over 34% higher cost
per click (CPCDALL-E3 = USD .38 vs. CPCStockPhoto = USD .51). In addition, we find that
model choice matters. Specifically, DALL-E 3 significantly outperforms the worst-performing
model SDXL Turbo (χ2(1, N = 36,594) = 23.968, p < .001) by 100% (CTRDALL-E3 =.80%
vs. CTRSDXLTurbo =.40%).
27

6.3. Quantifying generative AI-enabled cost savings compared to professional stock photos
Again, we quantify the cost savings enabled by generative AI. While the professional
stock photo is more than ten times cheaper than a freelancer image (USD 9 vs. USD 100), it
is orders of magnitude more expensive than the most effective image in the online campaign
generated by DALL-E 3 (USD 9 vs. USD .04). Put differently, an advertiser can create 225
images with DALL-E 3 for the price of a single stock photo.
In addition to the production costs, the CTR of DALL-E 3, which is over 50% higher
than that of the stock photo, substantially lowers its CPC by over 25%. Spending the same
budget of USD 49.99, DALL-E 3 obtains 133 clicks while the stock photo within the same
campaign obtains only 98 clicks. The least effective AI model, SDXL Turbo, achieves only
81 clicks, resulting in a CPC over 63% higher than DALL-E 3.
6.4. Discussion
Study 3 showed in a real-world environment that the best AI models can generate syn-
thetic banner ads that surpass the CTR of high-quality, human-made stock photography
selected by an online marketing professional. The CTR increase by more than 50% between
DALL-E 3 and the best human-made image is noteworthy, especially as we did not conduct
any prompt engineering or fine-tuning of the AI models.10
Closely inspecting the image by DALL-E 3 and the stock photo (see Figure 5) reveals
only a subtle difference in their image composition. While the presence of another person
in the stock photo might deter observers from engaging in self-referencing (Hartmann et al.,
2021), the first-person perspective of the image generated by DALL-E 3 could encourage
observers to imagine themselves holding the puzzle pieces, fostering mental simulation and
10These results align with those from a field study we conducted on Taboola’s online marketing platform
in December 2022 involving 13 generative text-to-image models. In this field study, the top-performing AI
model exceeded the human-made stock photo’s CTR by over 20%. For illustrative purposes, Web Appendix
Figure A.8 presents the human-made benchmark image alongside the AI-generated images that performed
best and worst (SD v1-3 and Disco Diffusion, respectively). Note the substantial advancements in image
quality observed when comparing the latest generations of AI models to their earlier versions from 1.5 years
ago, highlighting the rapid pace of technological progress in automated image creation.
28

translating into positive downstream consequences (Ceylan et al., 2024). Furthermore, the
physical arrangement of objects is more symmetric in the DALL-E 3 image, which can evoke
favorable consumer response (Zhang et al., 2022b).
Advertisers commonly run campaigns with multiple assets on online marketing platforms
that redistribute the campaign budget based on each asset’s effectiveness (Schwartz et al.,
2017). Our findings suggest that generative AI fits well into this A/B testing paradigm, as
it allows for creating many visual assets at a fraction of the cost of human-made content
with the potential to provide at least the same marketing effectiveness. The adoption of
a “human-in-the-loop” system can improve the effectiveness of AI applications in the field
even more (Reisenbichler et al., 2022). This approach involves human experts evaluating the
quality of multiple candidate ads before launching a marketing campaign. Alternatively, a
specialized predictive AI system, informed by historical performance data, can assume the
function of the human experts in an “AI-in-the-loop” arrangement, resulting in an iterative
interplay between a generative AI and a predictive AI model.
7. General Discussion
7.1. Summary
Generative AI represents a new paradigm, fundamentally disrupting the marketing in-
dustry (Peres et al., 2023). The present paper demonstrated the effectiveness and efficiency
gains that state-of-the-art generative text-to-image models can enable across a broad set of
marketing use cases. The AI models’ ability to rival human-made content across key mar-
keting metrics suggests that firms may soon find it necessary to embrace generative AI for
visual marketing content generation in their day-to-day operations to stay competitive.
Study 1 systematically evaluated consumer perceptions of AI-generated vs. human-made
marketing imagery, drawing on 254,400 human evaluations and algorithmic aesthetics assess-
ments. The results showed that the best AI models can generate synthetic sibling images that
can significantly outperform their human-made benchmark images in terms of quality, real-
29

ism, and aesthetics. Strikingly, consumers perceived synthetic images produced by Realistic
Vision as more realistic than real images, which has important implications beyond the mar-
keting discipline (Miller et al., 2023; Nightingale and Farid, 2022). Study 2 benchmarked the
same AI models with experienced human freelancers, giving both the same creative briefings.
The results showed that the best AI models (DALL-E 3 and Midjourney v6) can outperform
human-made marketing visuals across a broad battery of marketing metrics at a fraction of
the cost. Study 3, a field study with over 170,000 impressions, provided evidence on the
real-world effectiveness of AI-generated banner ads. DALL-E 3, the best AI model, yielded
an over 50% higher CTR than a high-quality, human-made stock photo selected by an online
marketing professional. Also, AI model choice matters. Compared to DALL-E 3, the CPC of
SDXL Turbo was more than 34% higher. Hence, choosing the wrong AI model can translate
into substantial economic costs (Hartmann et al., 2019).
7.2. Contribution and Implications
This research’s large-scale evaluation of AI-generated marketing imagery provides three
important contributions for scholars, managers, and policymakers. First, we provide evidence
that generative AI can match and even surpass human-made images’ consumer perception
and marketing effectiveness. While comparative studies are well-established in marketing re-
search (e.g., Andrews et al., 2002; Hartmann et al., 2023), our work, to our best knowledge, is
the first systematic comparison between human-made marketing content, such as professional
stock photos and visual assets from commissioned human freelancers, and AI-generated im-
ages produced by multiple state-of-the-art generative text-to-image models. Our findings
aim to assist marketing researchers and practitioners select appropriate AI models for their
substantive applications.
Second, the present paper contributes to understanding the human perception of AI-
generated visual marketing content. Perceptual studies are important for identifying im-
provement levers of marketing materials (e.g., Pieters and Wedel, 2004). For example, we
find that excessive saturation in AI images is negatively associated with perceived quality,
30

realism, and aesthetics. Flaws in human representation can also lead to adverse perceptual
reactions by consumers. To facilitate future research on the mechanism between visual fea-
tures and consumer response, we contribute all our AI-generated images as “GenImageNet”
to the research community.11
Third, inspired by Reisenbichler et al. (2022), we quantify the productivity gains of
generative AI for the creation of visual marketing content. While each freelancer image
cost USD 100 (study 2) and the professional stock photo USD 9 (study 3), producing AI-
generated images is orders of magnitude more cost-efficient. For example, generating a
single image in standard resolution with DALL-E 3, the winning method of our comparative
study, costs only USD .04. For the open-source models, SDXL Turbo and Realistic Vision,
that can be hosted and run locally, the costs per image converge to zero (USD .00005 and
.00029, respectively). Figure 6 visually summarizes the results of our back-of-the-envelope
cost calculations (for details, see Web Appendix Table A.6). Note that the cost per image on
the y-axis is log-scaled, highlighting the substantial cost differences across the image sources.
Beyond these three core contributions, our findings are a harbinger of AI-generated im-
ages’ potential to disrupt visual marketing content generation in the near future. Fitting well
into existing marketing paradigms such as rapid A/B testing, where multiple visual assets
are evaluated competitively and budget is allocated to the highest-performing ad (Schwartz
et al., 2017), advertisers can expect synthetic images to play a role of growing importance
in real-world marketing campaigns. Consistent with our findings, leading online marketing
platforms such as Google and Taboola recently announced a seemless integration of gener-
ative AI into their ad managers (Dischler, 2023; Feeney, 2024), enhancing generative AI’s
accessibility, adoption, and appeal. Fueled by its cost efficiency, the widespread application
of generative AI can contribute to more targeted ads with higher individual quality, as slight
variations in an advertising message’s verbal and visual language can increase its appeal for
different target audiences (Matz et al., 2017). At the same time, this outlook of “person-
11GenImageNet is available for download at: https://osf.io/8ctjy/
31

0.04 0.07
0.05
0.02
0.00005
0.00026
100
9
0.000
0.001
0.010
0.100
1.000
10.000
100.000
SDXL Turbo
Realistic Vision
Imagen 2
DALL−E 3
Firefly 2
Midjourney v6
Stock Photo
Freelancer
Image source
Cost per image (USD, log scale base 10)
Human−made
AI−generated
Note: Cost per image as of their creation dates in Q1 2024.
Figure 6: Cost per image for AI-generated vs. human-made images
alized mass persuasion” warrants scrutiny by policymakers and scholars across disciplines
(Matz et al., 2024).
What are the implications for companies of different sizes? Acar and Gvirtz (2024) sug-
gest that generative AI exerts a leveling effect with “the potential to close the content, insight,
and technology gaps that large corporations typically have over their smaller counterparts”.
The cost and performance advantages that we document across our studies, coupled with the
high accessibility of generative text-to-image models, support this notion that generative AI
can lead to a democratization of effective visual marketing content. Even if the prediction
that “gen AI makes everyone an ad agency” (Thomas, 2024) does not fully materialize, it
will likely substantially reweigh the tasks in the content creation process (Carlson et al.,
2023; Noy and Zhang, 2023).
Lastly, from a societal perspective, our findings have implications for policymakers and
contribute to the broader societal debate on the dissemination and detection of deepfakes
and disinformation (Karpinska-Krakowiak and Eisend, 2024). Should firms be allowed to
32

use synthetic images in marketing without an AI label? How do consumers react to such
a disclosure?12 Our results indicate that already today, consumers can perceive synthetic
images generated by specialized generative text-to-image models as more realistic than real
images. Considering the societal implications of generative AI, early adopters of this disrup-
tive technology must carefully monitor and navigate the rapidly evolving regulatory and legal
landscape. Future legislation might, for example, enforce digital provenance standards, such
as watermarks and disclosures. Similarly, users of generative AI risk infringing intellectual
property rights, requiring firms to exercise caution when selecting an AI model or provider
(Inman et al., 2024; Feuerriegel et al., 2024; Wang et al., 2023).
7.3. Limitations and future research directions
We acknowledge that our studies are subject to certain limitations that can inspire future
research. While we cover a diverse set of marketing applications, future research can explore
the effectiveness of generative AI for additional upper- and lower-funnel outcome measures,
e.g., brand awareness or sales. Building on our advertising pretest with commissioned human
freelancers from Freelancer.com, future research can benchmark state-of-the-art generative
text-to-image models against more expensive professional ad agencies, with and without
generative AI access (Sloane, 2024; Thomas, 2024).
Furthermore, by design of our multi-step image creation pipeline (see Figure 1) and to
ensure a fair comparison, we did not conduct any fine-tuning of the AI models. Similarly,
we did not conduct prompt engineering to avoid injecting a human bias into the prompt
creation. As both of these levers can further enhance generative text-to-image models’
12Images generated with Meta’s AI model, Imagine, include a visible watermark disclosing that they were
generated with AI, reading “Imagined with AI”. To ensure a fair, undisclosed comparison between all AI
models and the human-made benchmark images, we obscured the watermark through a blurring function
(see Web Appendix Figure A.2). However, to explore if Meta’s AI disclosure relates to human perception
of realism, we kept 20% of the images with a watermark (control) and only blurred the watermark on the
remaining 80% (treatment) of the images. Based on 2,400 human evaluations of these Imagine-generated
images, we find only marginally significant evidence that the disclosure treatment reduces human realism
perceptions (βwatermar k =−.1241, p =.077). Apparently, the AI watermark’s presence is not sufficient to
alter consumer perceptions (see Karpinska-Krakowiak and Eisend (2024) for similar results).
33

Research topic Research questions (examples)
Advertising •How will AI-generated ads alter the overall creativity of advertising over time?
•Can generative AI help in increasing (authentic) diversity in advertising?
•How can fine-tuning of generative AI models help in achieving different marketing
objectives (e.g., clicks vs. conversions of banner ads)?
•How can different data modalities (e.g., image, text, video, audio) be integrated
in generative AI-enabled advertising pipelines?
•How can the integration of generative AI and A/B testing functionality improve
the “learn-and-earn” trade-off? (Schwartz et al., 2017)
•How do usage patterns of generative AI evolve over time for different users groups?
Product design •How can generative AI support product design processes, ranging from ideation
to hyper-customization for different customer segments?
•Which product, brand, or customer characteristics moderate the perception of
AI-generated product designs?
•Can AI models learn a firm’s “visual brand essence” to create on-brand designs?
Social media •Do social media users prefer human or virtual influencers?
•What factors moderate consumers’ response to virtual influencers?
•What is the role of trust in virtual influencers’ marketing effectiveness (over time)?
•How does generative AI affect online communities and two-sided markets, e.g.,
freelancer or user-generated artwork?
•How do virtual influencers affect consumers’ social well-being?
Online shopping •Can generative AI enable personalized visual “website morphing”? (Hauser et al., 2009)
•Can AI-enabled product presentations or virtual try-ons reduce product returns?
•How might generative AI inflate customer expectations?
•How can generative AI enhance the interactivity of online shopping, e.g., via
visual assistance or dynamic product presentations such as instant color changes?
Productivity •How will AI-enabled ad makers offered by online marketing platforms disrupt the
value chain of content creation, e.g., by commoditizing certain tasks?
•How much human supervision is required when using generative AI in fast-paced,
multi-asset marketing campaigns?
•Can multi-modal models help in emulating customer behavior, e.g., their brand
perception by simulating brand elicitation exercises?
•What are the benefits and costs of model fine-tuning vs. using commercial
“off-the-shelf” foundation models?
•Can zero-shot image analytics using multi-modal large language models replace
conventional supervised image classification models?
Moderators •How might new policies and regulation affect the adoption of generative AI in
marketing, e.g., enforced disclosure and watermarking of AI-generated content?
•Are certain consumer segments (e.g., older, less educated) more vulnerable to
AI-enabled deepfakes, misinformation, and deception?
•How do consumers react to AI-generated content over time?
•What are the implications of biases in generative text-to-image models used for the
increasingly automated generation of marketing imagery?
•How can potential biases be uncovered and remedied?
•How to train AI models while protecting sensitive and private user data?
•How should policymakers react to potential threats of personalized mass persuasion?
Table 6: Examples for research questions related to AI-generated marketing imagery
34

effectiveness (Jansen et al., 2024; Rombach et al., 2022), our results likely represent a lower
bound for the performance of AI-generated marketing imagery. The emergence of future
generative AI models will likely improve the synthetic images’ perceptual ratings and real-
world effectiveness, especially when combined with task-specific data for model calibration
(Feng et al., 2023).
Lastly, while the presented studies show that generative AI can achieve better results
than human-generated content, more research is needed that analyzes the ingredients, i.e.,
the visual characteristics, that explain consumers’ response to AI-generated marketing im-
ages (e.g., Zhang and Luo, 2023; Zhang et al., 2022b). While we identified relevant structural
and content variables associated with differential perceptual evaluations regarding quality,
realism, and aesthetics, future research with larger sample sizes can explore the relationship
between additional visual features and lower-funnel effectiveness measures such as conver-
sions and sales. Table 6 lists further research ideas, including risks and moderators that
might hamper the adoption of generative AI.
Generative AI fundamentally disrupts visual marketing content generation. This research
investigated the disruptive potential of generative AI in marketing in terms of its effec-
tiveness and efficiency. By systematically benchmarking seven state-of-the-art generative
text-to-image models against human-made content, we showed that AI-generated marketing
imagery can achieve superhuman perceptual evaluations and effectiveness levels in real-world
applications at a fraction of the cost of human-made content. We hope our paper inspires
future research in the rapidly evolving area of generative marketing.
35

References
Acar, O.A., Gvirtz, A., 2024. GenAI Can Help Small Compa-
nies Level the Playing Field. URL: https://hbr.org/2024/02/
genai-can-help-small-companies-level-the-playing-field.
Ammanath, B., Dutt, D., Perricos, C., Sniderman, B., 2024. Now decides next: Insights
from the leading edge of generative AI adoption: Deloitte’s State of Generative AI in
the Enterprise Quarter one report. URL: https://www2.deloitte.com/content/dam/
Deloitte/us/Documents/consulting/us-state-of-gen-ai-report.pdf.
Andrews, R.L., Ainslie, A., Currim, I.S., 2002. An Empirical Comparison of Logit Choice
Models with Discrete versus Continuous Representations of Heterogeneity. Journal of
Marketing Research 39, 479–487. doi:10.1509/jmkr.39.4.479.19124.
Beichert, M., Bayerl, A., Goldenberg, J., Lanz, A., 2024. Revenue Generation Through
Influencer Marketing. Journal of Marketing doi:10.1177/00222429231217471.
Betker, J., Goh, G., Jing, L., Brooks, T., Wang, J., Li, L., Zhuang, J., Lee, J., Guo, Y.,
Manassra, W., Dhariwal, P., Chu, C., Jiao, Y., Ramesh, A., 2023. Improving Image
Generation with Better Captions. URL: https://cdn.openai.com/papers/dall-e-3.
pdf.
Bommasani, R., Hudson, D.A., Adeli, E., Altman, R., Arora, S., von Arx, S., Bernstein,
M.S., Bohg, J., Bosselut, A., Brunskill, E., Brynjolfsson, E., Buch, S., Card, D., Castellon,
R., Chatterji, N., Chen, A., Creel, K., Davis, J.Q., Demszky, D., . . . Liang, P., 2021. On
the Opportunities and Risks of Foundation Models. doi:10.48550/arXiv.2108.07258.
Bonaccorsi, A., Rossi, C., 2003. Why Open Source software can succeed. Research Policy
32, 1243–1258. doi:10.1016/S0048-7333(03)00051-9.
Borji, A., 2023. Qualitative failures of image generation models and their application in
detecting deepfakes. Image and Vision Computing 137, 104771. doi:10.1016/j.imavis.
2023.104771.
Brynjolfsson, E., Li, D., Raymond, L., 2023. Generative AI at Work doi:10.3386/w31161.
Burnap, A., Hauser, J.R., Timoshenko, A., 2023. Product Aesthetic Design: A Machine
Learning Augmentation. Marketing Science 42, 1029–1056. doi:10.1287/mksc.2022.1429.
36

Carlson, K., Kopalle, P.K., Riddell, A., Rockmore, D., Vana, P., 2023. Complementing
human effort in online reviews: A deep learning approach to automatic content generation
and review synthesis. International Journal of Research in Marketing 40, 54–74. doi:10.
1016/j.ijresmar.2022.02.004.
Ceylan, G., Diehl, K., Wood, W., 2024. From mentally doing to actually doing: A meta-
analysis of induced positive consumption simulations. Journal of Marketing 88, 21–39.
doi:10.1177/00222429231181071.
Chayka, K., 2023. The Uncanny Failures of A.I.-Generated Hands: When it comes to
one of humanity’s most important features, machines can grasp small patterns but
not the unifying whole. URL: https://www.newyorker.com/culture/rabbit-holes/
the-uncanny-failures-of-ai-generated-hands.
Childers, T.L., Houston, M.J., 1984. Conditions for a Picture-Superiority Effect on Consumer
Memory. Journal of Consumer Research 11, 643. doi:10.1086/209001.
Cho, H., Shen, L., Wilson, K., 2014. Perceived realism: Dimensions and roles in narrative
persuasion. Communication Research 41, 828–851. doi:10.1177/0093650212450585.
Chui, M., Hazan, E., Roberts, R., Singla, A., Smaje, K., Sukharevsky, A., Yee, L., Zem-
mel, R., 2023. The economic potential of generative AI: The next productivity frontier
URL: https://www.mckinsey.com/capabilities/mckinsey-digital/our-insights/
the-economic-potential-of-generative-ai-the-next-productivity-frontier.
Dell’Acqua, F., McFowland, E., Mollick, E.R., Lifshitz-Assaf, H., Kellogg, K., Rajendran, S.,
Krayer, L., Candelon, F., Lakhani, K.R., 2023. Navigating the Jagged Technological Fron-
tier: Field Experimental Evidence of the Effects of AI on Knowledge Worker Productivity
and Quality. SSRN Electronic Journal doi:10.2139/ssrn.4573321.
Ding, Y., Tian, C., Ding, H., Liu, L., 2023. The CLIP Model is Secretly an Image-to-Prompt
Converter. doi:10.48550/arXiv.2305.12716.
Dischler, J., 2023. Introducing a new era of AI-powered ads with Google. URL: https:
//blog.google/products/ads-commerce/ai-powered-ads-google-marketing-live/.
Dzyabura, D., El Kihal, S., Hauser, J.R., Ibragimov, M., 2023. Leveraging the Power of
Images in Managing Product Return Rates. Marketing Science 42, 1125–1142. doi:10.
1287/mksc.2023.1451.
37

Dzyabura, D., El Kihal, S., Peres, R., 2022. Image Analytics in Marketing, in: Homburg, C.,
Klarmann, M., Vomberg, A. (Eds.), Handbook of Market Research. Springer International
Publishing, Cham, pp. 665–692. doi:10.1007/978-3-319-57413-4_38.
Dzyabura, D., Peres, R., 2021. Visual Elicitation of Brand Perception. Journal of Marketing
85, 44–66. doi:10.1177/0022242921996661.
Feeney, K., 2024. GenAI Ad Maker: A New Way to Create High-Quality Ads Effortlessly.
URL: https://blog.taboola.com/create-ads-effortlessly/.
Feng, X., Zhang, S., Srinivasan, K., 2023. Marketing Through the Machine’s Eyes: Image
Analytics and Interpretability, in: Sudhir, K., Toubia, O. (Eds.), Artificial Intelligence
in Marketing. Emerald Publishing Limited. Review of Marketing Research, pp. 217–237.
doi:10.1108/S1548-643520230000020013.
Feuerriegel, S., Hartmann, J., Janiesch, C., Zschech, P., 2024. Generative AI. Business &
Information Systems Engineering 66, 111–126. doi:10.1007/s12599-023-00834-7.
Gartner, 2024. Gartner Experts Answer the Top Generative AI Questions for Your Enter-
prise: Generative AI isn’t just a technology or a business case — it is a key part of a
society in which people and machines work together. URL: https://www.gartner.com/
en/topics/generative-ai.
Grewal, R., Gupta, S., Hamilton, R., 2021. Marketing Insights from Multimedia Data: Text,
Image, Audio, and Video. Journal of Marketing Research 58, 1025–1033. doi:10.1177/
00222437211054601.
Hartmann, J., Heitmann, M., Schamp, C., Netzer, O., 2021. The Power of Brand Selfies.
Journal of Marketing Research 58, 1159–1177. doi:10.1177/00222437211037258.
Hartmann, J., Heitmann, M., Siebert, C., Schamp, C., 2023. More than a Feeling: Accuracy
and Application of Sentiment Analysis. International Journal of Research in Marketing
40, 75–87. doi:10.1016/j.ijresmar.2022.05.005.
Hartmann, J., Huppertz, J., Schamp, C., Heitmann, M., 2019. Comparing automated text
classification methods. International Journal of Research in Marketing 36, 20–38. doi:10.
1016/j.ijresmar.2018.09.009.
Hauser, J.R., Urban, G.L., Liberali, G., Braun, M., 2009. Website morphing. Marketing
Science 28, 202–223. doi:10.1287/mksc.1080.0459.
38

Horton, C.B., White, M.W., Iyengar, S.S., 2023. Bias against AI art can enhance perceptions
of human creativity. Scientific Reports 13, 19001. doi:10.1038/s41598-023-45202-3.
Hui, X., Reshef, O., Zhou, L., 2023. The Short-Term Effects of Generative Artificial Intelli-
gence on Employment: Evidence from an Online Labor Market. SSRN Electronic Journal
doi:10.2139/ssrn.4527336.
Hulland, J., Houston, M., 2021. The importance of behavioral outcomes. Journal of the
Academy of Marketing Science 49, 437–440. doi:10.1007/s11747-020-00764-w.
Inman, J.J., Meyer, R.J., Schweidel, D.A., Srinivasan, R., 2024. Do great powers come
with great responsibility? Opportunities and tensions of new technologies in marketing.
International Journal of Research in Marketing 41, 18–23. doi:10.1016/j.ijresmar.
2024.01.006.
Jakesch, M., Hancock, J.T., Naaman, M., 2023. Human heuristics for AI-generated language
are flawed. PNAS 120, e2208839120. doi:10.1073/pnas.2208839120.
Jansen, T., Heitmann, M., Reisenbichler, M., Schweidel, D.A., 2024. Automated Alignment:
Engaging Customers with Visual Generative AI. SSRN Electronic Journal doi:10.2139/
ssrn.4656622.
Karpinska-Krakowiak, M., Eisend, M., 2024. Realistic Portrayals of Untrue Information:
The Effects of Deepfaked Ads and Different Types of Disclosures. Journal of Advertising
, 1–11doi:10.1080/00913367.2024.2306415.
Keller, K.L., 1993. Conceptualizing, Measuring, and Managing Customer-Based Brand Eq-
uity. Journal of Marketing 57, 1–22. doi:10.1177/002224299305700101.
Keller, K.L., Lehmann, D.R., 2006. Brands and Branding: Research Findings and Future
Priorities. Marketing Science 25, 740–759. doi:10.1287/mksc.1050.0153.
Kelly, C., 2023. Coke asks consumers to generate art with new AI platform. URL: https:
//www.marketingdive.com/news/coca-cola-coke-generative-ai-marketing-art/
645465/.
Kim, B.K., Choi, J., Wakslak, C.J., 2019. The image realism effect: The effect of unre-
alistic product images in advertising. Journal of Advertising 48, 251–270. doi:10.1080/
00913367.2019.1597787.
39

King, A., 2024. AI to the Rescue? — BMG Says a ‘Single Project’ Can Involve
Up to 700 Digital Assets. URL: https://www.digitalmusicnews.com/2024/01/30/
bmg-digital-assets-management-ai-to-the-rescue/.
Krugmann, J.O., Hartmann, J., 2024. Sentiment Analysis in the Age of Generative AI.
Customer Needs and Solutions 11. doi:10.1007/s40547-024-00143-4.
Kyriakidi, M., 2022. Modern marketing dilemmas: Where does performance mar-
keting meet brand building? URL: https://www.kantar.com/inspiration/brands/
modern-marketing-dilemmas-where-does-performance-marketing-meet-brand-building.
Li, Y., Xie, Y., 2020. Is a Picture Worth a Thousand Words? An Empirical Study of
Image Content and Social Media Engagement. Journal of Marketing Research 57, 1–19.
doi:10.1177/0022243719881113.
Liu, L., Dzyabura, D., Mizik, N., 2020. Visual Listening In: Extracting Brand Image Por-
trayed on Social Media. Marketing Science 39, 669–686. doi:10.1287/mksc.2020.1226.
MacKenzie, S.B., Lutz, R.J., Belch, G.E., 1986. The role of attitude toward the ad as
a mediator of advertising effectiveness: A test of competing explanations. Journal of
Marketing Research 23, 130–143. doi:10.2307/3151660.
Matz, S.C., Kosinski, M., Nave, G., Stillwell, D.J., 2017. Psychological targeting as an
effective approach to digital mass persuasion. PNAS 114, 12714–12719. doi:10.1073/
pnas.1710966114.
Matz, S.C., Teeny, J.D., Vaid, S.S., Peters, H., Harari, G.M., Cerf, M., 2024. The potential
of generative AI for personalized persuasion at scale. Scientific Reports 14, 4692. doi:10.
1038/s41598-024-53755-0.
Miller, E.J., Steward, B.A., Witkower, Z., Sutherland, C.A.M., Krumhuber, E.G., Dawel,
A., 2023. AI Hyperrealism: Why AI Faces Are Perceived as More Real Than Human
Ones. Psychological Science 34, 1390–1403. doi:10.1177/09567976231207095.
Nightingale, S.J., Farid, H., 2022. AI-synthesized faces are indistinguishable from real faces
and more trustworthy. PNAS 119, e2120481119. doi:10.1073/pnas.2120481119.
Noy, S., Zhang, W., 2023. Experimental evidence on the productivity effects of generative
artificial intelligence. Science 381, 187–192. doi:10.1126/science.adh2586.
oatawa, 2024. User portfolio. URL: https://www.istockphoto.com/de/portfolio/
oatawa.
40

Paivio, A., Csapo, K., 1973. Picture superiority in free recall: Imagery or dual coding?
Cognitive Psychology 5, 176–206. doi:10.1016/0010-0285(73)90032-7.
Peng, S., Kalliamvakou, E., Cihon, P., Demirer, M., 2023. The Impact of AI on Developer
Productivity: Evidence from GitHub Copilot. doi:10.48550/arXiv.2302.06590.
Peres, R., Schreier, M., Schweidel, D., Sorescu, A., 2023. On ChatGPT and beyond: How
generative artificial intelligence may affect research, teaching, and practice. International
Journal of Research in Marketing 40, 269–275. doi:10.1016/j.ijresmar.2023.03.001.
Phillips, B.J., McQuarrie, E.F., Griffin, W.G., 2014. How Visual Brand Identity Shapes
Consumer Response. Psychology & Marketing 31, 225–236. doi:10.1002/mar.20689.
Pieters, R., Wedel, M., 2004. Attention Capture and Transfer in Advertising: Brand, Picto-
rial, and Text-Size Effects. Journal of Marketing 68, 36–50. doi:10.1509/jmkg.68.2.36.
27794.
Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell,
A., Mishkin, P., Clark, J., Krueger, G., Sutskever, I., 2021. Learning Transferable Vi-
sual Models From Natural Language Supervision, in: Proceedings of Machine Learning
Research, pp. 8748–8763. URL: https://proceedings.mlr.press/v139/radford21a.
Reisenbichler, M., Reutterer, T., Schweidel, D., 2023. Applying Large Language Mod-
els to Sponsored Search Advertising URL: https://www.msi.org/working-paper/
applying-large-language-models-to-sponsored-search-advertising/.
Reisenbichler, M., Reutterer, T., Schweidel, D.A., Dan, D., 2022. Frontiers: Supporting
Content Marketing with Natural Language Generation. Marketing Science 41, 441–452.
doi:10.1287/mksc.2022.1354.
Rizzo, G.L.C., Berger, J.A., Villarroel Ordenes, F., 2023. What Drives Virtual Influencer’s
Impact? SSRN Electronic Journal doi:10.2139/ssrn.4329150.
Rodgers, B., 2021. How Much Does Commercial Product Photogra-
phy Cost? URL: https://digitalartthatrocks.com/blog/2021/12/8/
how-much-does-commercial-product-photography-cost.
Rombach, R., Blattmann, A., Lorenz, D., Esser, P., Ommer, B., 2022. High-Resolution
Image Synthesis with Latent Diffusion Models, in: 2022 IEEE/CVF Conference on Com-
puter Vision and Pattern Recognition (CVPR), IEEE. pp. 10674–10685. doi:10.1109/
CVPR52688.2022.01042.
41

Saharia, C., Chan, W., Saxena, S., Li, L., Whang, J., Denton, E.L., Ghasemipour,
K., Gontijo Lopes, R., Karagol Ayan, B., Salimans, T., Ho, J., Fleet, D.J.,
Norouzi, M., 2022. Photorealistic Text-to-Image Diffusion Models with Deep
Language Understanding, in: Koyejo, S., Mohamed, S., Agarwal, A., Bel-
grave, D., Cho, K., Oh, A. (Eds.), Advances in Neural Information Processing
Systems. URL: https://proceedings.neurips.cc/paper_files/paper/2022/hash/
ec795aeadae0b7d230fa35cbaf04c041-Abstract-Conference.html.
Sauer, A., Lorenz, D., Blattmann, A., Rombach, R., 2023. Adversarial diffusion distillation
doi:10.48550/arXiv.2311.17042.
Schwartz, E.M., Bradlow, E.T., Fader, P.S., 2017. Customer acquisition via display
advertising using multi-armed bandit experiments. Marketing Science 36, 500–522.
doi:10.1287/mksc.2016.1023.
Seiler, S., Tuchman, A., Yao, S., 2021. The Impact of Soda Taxes: Pass-Through, Tax
Avoidance, and Nutritional Effects. Journal of Marketing Research 58, 22–49. doi:10.
1177/0022243720969401.
Sloane, G., 2024. Brands Add AI Restrictions To Agency Contracts—Behind The Grow-
ing Trend. URL: https://adage.com/article/digital-marketing-ad-tech-news/
ai-restrictions-added-ad-agency-contracts/2548696.
Smith, R.E., MacKenzie, S.B., Yang, X., Buchholz, L.M., Darley, W.K., 2007. Modeling
the determinants and effects of creativity in advertising. Marketing Science 26, 819–833.
doi:http://www.jstor.org/stable/40057228.
Steenkamp, J.B.E., van Heerde, H.J., Geyskens, I., 2010. What Makes Consumers Willing
to Pay a Price Premium for National Brands over Private Labels? Journal of Marketing
Research 47, 1011–1024. doi:10.1509/jmkr.47.6.1011.
Talebi, H., Milanfar, P., 2018. NIMA: Neural Image Assessment. IEEE Transactions on
Image Processing doi:10.1109/TIP.2018.2831899.
The One Club, 2023. A.I. Ketchup. URL: https://www.oneclub.org/awards/theoneshow/
-award/48285/ai-ketchup/ai-ketchup.
Thomas, D., 2024. AI advertising start-up valued at $4bn after fundraising. URL: https:
//www.ft.com/content/4c7bee10-51d3-489b-873a-765157af8aac.
42

van Heerde, H.J., Moorman, C., Moreau, C.P., Palmatier, R.W., 2021. Reality Check:
Infusing Ecological Value into Academic Marketing Research. Journal of Marketing 85,
1–13. doi:10.1177/0022242921992383.
Wang, W., Bell, J.J., Dotson, J.P., Schweidel, D.A., 2023. Generative AI and Artists:
Consumer Preferences for Style and Fair Compensation. SSRN Electronic Journal doi:10.
2139/ssrn.4428509.
Xu, L., Mehta, R., 2022. Technology devalues luxury? Exploring consumer responses to AI-
designed luxury products. Journal of the Academy of Marketing Science 50, 1135–1152.
doi:10.1007/s11747-022-00854-x.
Zhang, H., Bai, X., Ma, Z., 2022a. Consumer reactions to AI design: Exploring consumer
willingness to pay for AI–designed products. Psychology & Marketing 39, 2171–2183.
doi:10.1002/mar.21721.
Zhang, M., Luo, L., 2023. Can Consumer-Posted Photos Serve as a Leading Indicator of
Restaurant Survival? Evidence from Yelp. Management Science 69, 25–50. doi:10.1287/
mnsc.2022.4359.
Zhang, S., Lee, D., Singh, P.V., Srinivasan, K., 2022b. What Makes a Good Image? Airbnb
Demand Analytics Leveraging Interpretable Image Features. Management Science 68,
5644–5666. doi:10.1287/mnsc.2021.4175.
Zhang, X., Zhou, M., Lee, G.M., 2024. Generative AI and Creator Economy: Investigating
the Effects of AI-Generated Voice on Online Video Creation. SSRN Electronic Journal
doi:10.2139/ssrn.4676705.
Zhou, E., Lee, D., 2024. Generative artificial intelligence, human creativity, and art. PNAS
Nexus 3, pgae052. doi:10.1093/pnasnexus/pgae052.
43