Does Generative AI Actually Improve Content Marketing?
Marketers are being told, almost daily, that generative AI will transform content marketing and writing better copy, designing better visuals, and doing it all faster and cheaper. But "faster and cheaper" is not the same as "more effective." This week I want to walk through five recent pieces of research that, together, give a more precise (and more useful) picture of what generative AI actually does and doesn't do for content marketing. One is a large-scale, three-study comparison of AI-generated versus human-made visual content; two are controlled experiments in tourism marketing; one is a strategic research roadmap; and one is a multi-study test of whether fine-tuning generative AI directly on marketing performance data beats simply prompting it.
When AI-generated images beat professional photography?
The paper of Hartmann et al. (2025) ran three studies comparing AI-generated marketing images against human-made content, using seven state-of-the-art text-to-image models (DALL-E 3, Midjourney v6, Adobe Firefly 2, Google Imagen 2, Meta's Imagine, Stable Diffusion XL Turbo, and Realistic Vision).
In Study 1, they generated over 10,000 synthetic images as "siblings" of 2,400 real, human-made marketing images spanning firm- and user-generated content across both call-to-action and brand-identity use cases, and collected 254,400 human ratings on quality, realism, and aesthetics. Four of the seven models significantly surpassed human-made images on quality, and all seven surpassed them on aesthetic appeal. One model in particular, Realistic Vision, produced images people rated as more realistic than the real, human-made photos, which is a striking "AI hyperrealism" effect that echoes similar findings in other domains.
In Study 2, the authors gave identical creative briefs to commissioned human freelancers (paid around $100 per image) and to the same seven AI models, then had over 1,500 people rate the results across ten marketing metrics. DALL-E 3 came out on top, significantly outperforming the freelancers on five metrics, including perceived ad creativity and scoring directionally higher on the rest, all while costing a small fraction of a cent per image.
Study 3 took this into the real world. A live Meta ad campaign (173,022 impressions) comparing DALL-E 3-generated banner ads against a professional, human-made stock photo chosen by an experienced marketing practitioner. The AI-generated ad achieved a meaningfully higher click-through rate than the stock photo, at a lower cost per click, while costing a small fraction of what the stock photo itself cost to license.
A few caveats worth holding onto: model choice mattered enormously (the open-source models tested performed noticeably worse than DALL-E 3 and Midjourney v6), the field study covered one campaign for one company over a few days, and the specific text-to-image models tested are already dated by AI standards. Newer versions are released constantly. So I'd treat the direction of these findings (AI-generated images can match or exceed professional human-made content on several important dimensions, at a small fraction of the cost) as well-supported, while treating the exact percentages as illustrative of this particular study rather than universal constants.
AI can polish your copy, but polish isn't the same as persuasion
Guttentag et al. (2025) ran one of the first direct comparisons of human-written versus ChatGPT-optimised marketing copy. They took real TripAdvisor descriptions for three Panamanian indigenous cultural tours, asked ChatGPT-3.5 to "optimise" them with a single simple prompt, and then had 579 research subjects evaluate either the original or the AI-optimised version.
The AI-optimised text was rated as significantly easier to read, friendlier, and more professional across the board, and it also generated more interest and was seen as somewhat more useful and credible. Interestingly, the same product images were rated as more effective at generating interest when they were paired with the AI-optimised text than with the original text even though the images themselves never changed, echoing the halo-effect logic in the Hartmann et al. study above.
But here's the finding I'd flag most for practitioners: none of this translated into a statistically significant increase in stated desire to actually take the tour, or in how much people estimated they'd pay for it. Better-sounding copy did not reliably move the needle on purchase intent. The authors' own conclusion highlights that: AI-optimised descriptions are a useful, nearly free upgrade to existing copy, but they are not a substitute for the many other factors such as price, reviews, and general interest in the category that drive an actual booking decision.
Should you tell people AI wrote it? It depends on the emotional register
A second tourism study, by Song et al. (2024) tackled a question a lot of us are quietly wondering about: does disclosing that content was AI-generated help or hurt?
Across three scenario-based experiments, they found a clear matching pattern. When an advertisement used a rational appeal (facts, statistics, functional details), visit intention was higher when the ad was declared to be AI-generated. When the ad used an emotional appeal (atmosphere, nostalgia, comfort), visit intention was higher when it was declared to be human-generated. However, for mismatches underperformed for AI-declared emotional copy, or human-declared rational copy. The authors trace this to "information processing fluency": when the declared creator matches the tone of the message, people process it more easily and evaluate it more favourably.
The strategic view: where GenAI fits in the marketing process
Zooming out from individual ads, Cillo and Rubera (2025) offer a research roadmap for GenAI across the whole innovation and marketing process by developing, testing, communicating, and engaging. A few points stand out for content marketing specifically. First, GenAI's output is genuinely novel and often "appropriate," but this only holds reliably for tasks where there's no single right answer; the moment you need grounded facts or current information, these models still hallucinate. Second, on disclosure, the authors note the evidence is genuinely mixed at the brand level: it could either reinforce a brand's innovative image or clash with it, depending on existing brand associations. Third, they flag that firms who let GenAI substitute too heavily for their own market research risk losing a valuable intellectual asset: real customer knowledge.
Can you fine-tune generative AI to write winning ads, rather than just imitate existing ones?
Most of the research above test what happens when you prompt an off-the-shelf model like DALL-E 3. Heitmann et al. (2026) ask a more ambitious question: what if you fine-tune the generative model directly on marketing performance data, teaching it what a high-performing ad actually looks like, rather than relying on the model's default aesthetic sense?
Working with the Polestar 3 electric vehicle as their main case, the authors fine-tuned an open-source Stable Diffusion model on the highest-rated ads from a large pool of automotive advertising, using consumer ratings across the classic AIDA funnel (attention, interest, desire, activation) to identify which visuals to train on. The resulting AI-generated ads scored significantly higher on AIDA measures (M = 4.55 on a seven-point scale) than both market-competitor ads (M = 3.79) and the brand's own actual ads (M = 3.80), and only 3 of the 50 generated ads scored below the conventional-ad average. Importantly, this lift wasn't simply about AI images looking more polished because a mediation analysis in the paper attributed only around 15% of the performance gap to aesthetic quality.
The same workflow let them train in specific brand-personality traits (e.g., "rugged" vs. "luxury") without sacrificing AIDA performance, and a live Meta ad campaign found the fine-tuned AI ads achieved a meaningfully higher click-through rate than the brand's actual ads (I'd flag this as one campaign run over a few days, so treat the exact figures as illustrative rather than a permanent benchmark).
Where the approach didn't work is just as informative for practitioners. The performance lift held up regardless of brand familiarity and across both durable (cars) and nondurable (sunscreen) product categories, but it essentially disappeared for a niche, highly differentiated product (a two-seater city car), and the fine-tuned model couldn't reproduce humour where a human-made humorous ad in one comparison still beat its AI-generated counterpart on brand recall. The authors' interpretation is that fine-tuning on category-typical winners tends to produce visually "fluent," average-feeling ads, which helps when a brand's goal is familiarity and trust, but can work against you when the whole point of the creative is to stand out or surprise.
Practical AI and content marketing takeaways
• Generative AI images can genuinely rival and in some tested cases outperform professional human-made marketing visuals on quality, aesthetics, ad creativity, and even real-world click-through rate, at a small fraction of the production cost. Model choice matters a lot, so treat this as a reason to test, not a reason to skip testing.
• Use GenAI text to upgrade weak product descriptions expect better perceived tone and readability, not a guaranteed lift in bookings or sales.
• Match your disclosure decision to your message: AI-declared for rational/functional appeals, human-declared for emotional appeals.
• Treat GenAI output as a strong first draft requiring human review, especially where facts, statistics, or current information are involved.
• Off-the-shelf prompting and true performance-based fine-tuning are not the same thing: the biggest gains come from training generative AI directly on real consumer-response data, not just relying on a model's default aesthetic capabilities, and this approach seems to work best for mass-market, familiarity-driven products rather than niche or humour-led campaigns.
References
Cillo, P., & Rubera, G. (2025). Generative AI in innovation and marketing processes: A roadmap of research opportunities. Journal of the Academy of Marketing Science, 53, 684–701. https://doi.org/10.1007/s11747-024-01044-7
Guttentag, D. A., Litvin, S. W., & Teixeira, R. (2025). Human vs. AI: Can ChatGPT improve tourism product descriptions? Current Issues in Tourism, 28(22), 3601–3619. https://doi.org/10.1080/13683500.2024.2402563
Hartmann, J., Exner, Y., & Domdey, S. (2025). The power of generative marketing: Can generative AI create superhuman visual marketing content? International Journal of Research in Marketing, 42(1), 13–31. https://doi.org/10.1016/j.ijresmar.2024.09.002
Heitmann, M., Jansen, T. P. J., Reisenbichler, M., & Schweidel, D. A. (2026). Picture perfect: Engaging customers with visual generative AI. Journal of Marketing, 90(4), 74–96. https://doi.org/10.1177/00222429251356993
Song, M., Chen, H., Wang, Y., & Duan, Y. (2024). Can AI fully replace human designers? Matching effects between declared creator types and advertising appeals on tourists' visit intentions. Journal of Destination Marketing & Management, 32, Article 100892. https://doi.org/10.1016/j.jdmm.2024.100892
