AI & Marketing Research with Dr. Eva Wolf
Not another AI news podcast. This is a research radar — a twice-weekly briefing that surfaces peer-reviewed studies on AI and marketing, tells you what the evidence actually says, and helps you decide what's worth a deeper read.
AI & Marketing Research with Dr. Eva Wolf
AI & Marketing Research Radar — 2026-05-14
Use Left/Right to seek, Home/End to jump to start or end. Hold shift to jump forward or backward.
Welcome to AI and Marketing Research Radar. I'm Eva Wolf, and this is your first pass research briefing for the week. We screened 105 papers and selected five worth your time. Today's theme, AI and Marketing, how it's changing creative work, marketplace competition, brand visibility, and consumer trust all at once.
SPEAKER_00Quick reminder before we start: this is a first pass research briefing, not a final academic review. We'll tell you what these papers suggest, what they don't prove, and which ones deserve a closer read. All five papers today are flagged as deep dive candidates. Links to every source are in the show notes.
SPEAKER_01Let's get into it.
SPEAKER_00Paper one is Personalized AI Scaffold Synergistic Multi-Turn Collaboration in Creative Work by Sean Kelly, David DeKremer, and Christoph Riedel, published in 2025. It's a preprint on ARCHEFFR, not yet peer-reviewed. Access status, full text available. Why should marketers care? Most of us use AI the same way every time. We open a blank chat, describe the task, and go. This paper tests whether telling the AI who you are first actually changes what you get back. The research question: Does giving an AI assistant detailed information about your personality, creative style, and domain expertise lead to better work than using a generic AI? Now, the method. The researchers ran a randomized controlled experiment with 331 participants split into three groups. One group used generic AI, one used partially personalized AI, one used fully personalized AI. Personalization was built from pre-experiment psychometric surveys and an AI-guided interview about each person's work style. Every participant completed a marketing campaign task for a fictional startup. Campaign quality and creativity were judged using a blinded, rubric-based LLM evaluator, validated against human expert standards. The researchers also ran causal mediation analysis to identify the mechanisms. Now, the key findings. People working with the fully personalized AI produced campaigns significantly higher in quality and creativity than those using generic AI. And the personalized AI didn't just outperform generic AI. The human AI team together outperformed what the AI could produce on its own. That's real collaborative synergy, not the AI doing the work. Personalized AI users also reported feeling more helped, more trusted, more confident. The mechanism the researchers identified and the reason it worked was alignment across the whole conversation. Shared memory, focused attention, and joint reason. So with a limitation. The task was a fictional start-up assignment, which may not reflect the pressure of real professional work. Participants were recruited for a controlled experiment, so effort levels may differ from the deadline condition. The campaign judge was an LLM, though validated against human standards. The study doesn't tell us whether benefits hold up over weeks or months. And it's a preprint.
SPEAKER_01Right. So here's what I'd actually do with this. Before your next big creative project, a campaign brief, a positioning document, a content series, spend 10 to 15 minutes writing a profile of yourself and paste it into the system prompt before you start. Tell the AI your marketing background, what you're strong at, what you find difficult, how you like to work through ideas, not the task, you. This study suggests context is the difference between an AI that gives you generic feedback and one that actually pushes your thinking. If you manage a team, this gets more interesting. Build a short onboarding questionnaire, maybe five questions, that captures each person's expertise, thinking style, and creative preferences. Use those answers to pre-configure their AI assistant. The setup is 20 minutes per person. The payoff, if this paper holds up, is real. One thing to flag, okay? The task here was specific and contained. A fictional startup campaign is not a full integrated brand strategy. So treat this as a strong signal, not a guarantee. The preprint status means it hasn't been challenged by peer reviewers yet. But the design is clean and the finding is directionally clear. My triage verdict, deep dive, because it's a randomized controlled experiment on a marketing-specific task with causal analysis, a sample of 331, and findings you can act on before Friday.
SPEAKER_00Paper 2 is Vertical Tacit Collusion in AI-mediated markets by Felipe M. Afonso, published in 2026. It's a preprint on Air Quassiv, not yet peer-reviewed. Access Status, full text available. Why should marketers care? If you sell products on any marketplace where AI recommends what consumers buy, Amazon with Rufus, ChatGPT shopping, perplexity commerce, this paper is about the system you're competing in right now. The research question: when a platform independently optimizes its rankings to exploit AI shopping agents' weaknesses, and sellers independently optimize their listings to do the same. Does the combined harm to consumers exceed what either party causes alone? Now, the method. The researcher built a multi-agent simulation using a formal repeated game model. The platform and sellers are modeled as independent reinforcement learning agents, each maximizing returns over time without communicating. Critically, the model was calibrated to published empirical measurements of real LLM cognitive biases. These aren't assumed weaknesses. They come from documented benchmarks across GPT 4.1, Claude Sonnet 4, Gemini 2.5 Flash, and Lama. Consumer harm is measured as the gap between what the AI agent recommends and what a perfectly informed consumer would actually want. So, the findings. When both parties optimize independently, harm to consumers more than doubles what either party causes alone. The strategies amplify each other. Now here's the biased data that makes this alarming. Claude Sona 4 picked products from the top half of a list 77% of the time, regardless of actual quality. GPT-4.1 and Gemini 2.5 Flash showed similar patterns. And unlike individual humans, who vary in how easily they're influenced, every user of the same AI model is affected the same way, making these biases scalable and predictable. Platforms and sellers never need to communicate. Each just optimizes on their own. The paper calls this vertical, tacit collusion, and the regulatory gap is real. Current antitrust law is built to catch coordination and conspiracy. When there's no coordination, existing law has no clear mechanism to intervene. The limitations. It's a simulation, not a field study. Real marketplace dynamics are more complex. The bid weight values are discretized: 00, 0.33, 0.671, which simplifies the strategy space. The LLM bias measurements come from published benchmarks that may not perfectly represent proprietary, continuously updated shopping systems. The 2x Harm figure comes from the model's parameters, not real market data, and it's a preprint.
SPEAKER_01Okay, I'm telling you. This one changes how I think about e-commerce strategy. The rules for winning in AI-mediated shopping are already different from the rules for winning in search or browse-based discovery. Different. This is a new game, and the paper gives you the vocabulary to explain why. My triage verdict, deep dive. Because the finding is directly actionable for any brand selling on AI-powered marketplaces. The calibration to real LLM bias data is rigorous, and vertical tacit collusion is a concept every e-commerce strategist needs to understand now.
SPEAKER_00Paper 3 is From Stereotypes to Strategy, Addressing Gender Bias in AI-powered marketing by Katerina Fox and Gabriele Schuster, published in 2026 in the International Conference on Gender Research. It's a conference paper, likely peer-reviewed. Access Status, full text, available. Why should marketers care? This paper goes inside the day-to-day work of marketing professionals already using AI tools and asks, are you introducing gender bias without knowing it? And what are you actually doing about it? The research question, how do marketing professionals perceive gender bias in AI-driven marketing? And what strategies do they use to identify and reduce it? Now the method is qualitative. Eleven guided, semi-structured expert interviews with marketing and media professionals in Germany. Content creators, creative directors, art directors, agency managing directors, event agency executives all had practical AI experience and operational responsibility. Interviews ran 45 to 60 minutes each by video conference between October and December 2024. Data were analyzed using Maring's qualitative content analysis methodology. So, the key findings. Marketing professionals have very different levels of awareness. Some treat bias as a PR risk, others as a genuine social harm. AI bias in marketing isn't only a technical data problem, it also gets baked into the creative process itself, through the prompts humans write and the outputs they fail to scrutinize. AI systems can lock stereotypes in place through self-reinforcing loops. For example, if women historically clicked more on household product ads, the algorithm keeps showing those ads to women, and the pattern strengthens over time. And the most commonly cited fixes, diverse teams, peer review of AI outputs, more careful prompting, structured reflection time, were consistently blocked by time pressure, budget constraints, and lack of organizational support. The limitations 11 experts is a very small sample, all based in Germany, so generalizability is limited. Qualitative design captures what people say they do, not what they actually do or what actually works. Purposive sampling may overrepresent professionals already paying attention to bias. And there's no quantitative measurement of actual bias levels or how well any mitigation strategy performs.
SPEAKER_01So, the finding I keep coming back to is the self-reinforcing loop. This isn't a one-time mistake you can fix and move on from. If your targeting data reflects historical stereotypes, women seeing household ads, men seeing financial ads, and your AI is optimizing on that data. The bias compounds with every campaign. It gets stronger over time unless you actively interrupt it. Here's something concrete you can do this week. Next time you prompt an AI to generate images, copy, or audience targeting, add an explicit instruction. Ask for diverse representations. Specify that you want the output reviewed for default stereotyping, and check the result before it goes anywhere near a client or consumer. That prompt adjustment takes 30 seconds. The experts in this study said the biggest barrier to fixing bias was time. So build the check into the step that already exists, not into a separate meeting. If you manage a team, push for one standing quarterly review of AI-generated content for fairness, not a committee, not a workshop. One meeting per quarter where you look at what the AI produced and ask whether it defaulted to stereotypes. The paper's experts said structured reflection time is the single biggest thing that's missing. My triage verdict, deep dive. Because this is one of the first studies to gather on-the-ground practitioner perspectives on how AI bias actually plays out in real agency work, the methodology is clearly described, and the mitigation strategies are immediately practical. The small Germany-only sample caps evidence quality, but the practitioner angle is valuable and underrepresented.
SPEAKER_00Paper 4 is The Impact of AI Generated Marketing Imagery on Consumer Trust and Purchase Intentions. Examining Effect of Human AI-assisted images on marketing by Rushikesh Lahan, Mahek Ahuja, Mehak Sharma, Amrita, and Atif Javed Kazi, published in 2026 in the International Journal for Research in Applied Science and Engineering Technology. Peer-reviewed journal article. Access status, full text available. Why should marketers care? This is the paper that puts numbers on the trust gap between AI-generated and human-made advertising, and it tests whether a hybrid approach actually fixes the problem. The research question: How do consumers respond to marketing images created by humans AI, or a human-AI combination in terms of trust and willingness to buy? Now, the method, a factorial design experiment with two independent variables, creator type, human AI, or human AI hybrid, and product category, fast food versus cosmetics, and skincare. Each participant saw one ad with a disclosure tag identifying who or what created it. Sample size, 214 participants. So the findings. Human-made ads scored 5.63 out of 7 for consumer trust. AI-made ads scored 4.24, a statistically significant gap. Human ads also led to higher purchase intent, 5.62 versus 4.65 for AI. Now here's the important nuance. Hybrid ads, human and AI together, scored 5.08 for trust, meaningfully better than pure AI at 424. But the hybrid condition did not significantly improve purchase intent compared to fully AI-made ads. So correcting the trust problem didn't automatically fix the purchase problem. Consumers also perceived AI-made ads as less realistic and as requiring less effort to produce. Human-made ads felt more carefully crafted. Hybrid ads fell in the middle on both counts. The limitations. The sample of 214 is moderate, and demographic details aren't specified in the available text. Participants saw only one ad each, so we don't know how repeated exposure changes responses. Only two product categories were tested. The experiment used explicit disclosure tags. In real-world settings, disclosures aren't always present, which could significantly alter reactions. And the study was conducted at a single institution in India, which limits cross-cultural generalizability.
SPEAKER_01This finding is counterintuitive in a specific way, and I want to be direct about it. Fixing the trust problem with hybrid labeling did not fix the willingness to buy problem. That gap matters. Marketers moving to hybrid framing, thinking it gives them the best of both worlds, may be improving how their brand is perceived without moving the metric that matters most. Before you invest in a hybrid labeling strategy, test actual purchase behavior, not just survey responses. Here's the practical application. If you're in a category where regulations or platform rules require AI disclosure, or where you're considering voluntary disclosure as a trust signal, label your creative as AI-assisted, directed by human name or team, rather than simply AI generated. The data suggests that framing raises trust, but track conversion, not just sentiment. For high-stakes trust-sensitive categories, skincare, cosmetics, health products, be especially cautious about replacing human creatives entirely. The trust gap was 1.39 points on a seven-point scale. That's large, and it was consistent across both product categories. One more thing, this study used explicit disclosure. If your consumers aren't told whether an image is AI made, their responses may be completely different. Don't extrapolate to undisclosed contexts. My triage verdict, deep dive, because it's the paper in this batch that most directly answers a question brand managers are facing right now. Does it matter whether we tell people our ads are AI generated? The answer, at least in disclosed settings, is yes. And the hybrid caveat is critical information.
SPEAKER_00Paper 5 is Cultural Encoding in Large Language Models: The Existence Gap in AI-mediated brand discovery by Huang Junyao, Situ Ruamin, and Ye Renkin, published in 2025 on R Quest. Preprint, not yet peer-reviewed. Access status, full text available. Before we go further, there's a significant conflict of interest to disclose. All three authors are employees of OmniEdge, also known as Z Bianj, the exact company used as the primary case study in the paper. Weigh the findings with that in mind and wait for independent replication before acting heavily on the specific numbers.
SPEAKER_01Why should marketers care?
SPEAKER_00The core question: if consumers are increasingly asking AI chatbots for product recommendations and your brand isn't mentioned, do you effectively not exist? This paper argues yes. And it tries to measure how big that visibility gap is. The research question Do AI chatbots recommend brands differently, depending on whether those brands are well represented in training data? And does this create an unfair invisibility barrier for brands that exist primarily in one language or region? Now, the method. The researchers submitted 1,909 identical English language queries across six large language models: GPT-40, Claude, Gemini, Quin3, DeepSeek, and Doughau, and recorded which brands each model mentioned. They compared mention rates for 30 brands. Statistical significance was tested using Chi Square analysis. A case study of one brand, Zizi Bianjai, the author's own company, was used to illustrate the effect. That case study is where the conflict of interest is most concentrated. So the findings. Chinese AI models mention brands 89% of the time on average. International AI models mention brands 58% of the time, a 31-point gap, even when questions were asked in identical English. The gap is driven by training data composition, not query language. The author's own product, Zizibianji, was recommended by Chinese AIs 66% of the time, and by Western AIs 0% of the time across 32 identical queries. The paper calls this the existence gap. If your brand isn't in an AI's training data, it simply won't be mentioned, and there's no equivalent of a second page of search results. The limitations. Fully describe how those brands were selected. It's observational, not causal. Other factors like model recency, knowledge cutoffs, and category handling weren't fully ruled out. Only English language queries were used. The 18-month strategic roadmap offered in the paper isn't empirically tested, and it's a preprint.
SPEAKER_01So, the concept here is real and important, even if the specific numbers need independent verification. When a consumer asks an AI, what's the best tool for X, or what should I use for Y, your brand either gets mentioned or it doesn't. There's no third position, no sponsored slot, no way to buy your way in after the response is generated. That's a fundamentally different visibility problem than anything that existed in search. Here's what I'd actually do with this right now. Go to GPT-40, Claude, and Gemini and type the questions your customers would ask. What's the best platform for X? What tool do marketers use for Y? See whether your brand appears. If it doesn't, that's a gap you can work on. Publishing English language content, technical documentation, case studies, blog posts, forum contributions creates the kind of training signal that influences what future model versions know about you. This is sometimes called generative engine optimization, or GEO, and it's a real strategy, separate from traditional SEO. For any brand that operates primarily in one country or one language, this paper is a direct warning. Your local market dominance may be completely invisible to the AI systems your customers in other markets are using. The conflict of interest is real, and I'm flagging it. But the existence gap as a concept is independently testable, and I'd encourage you to run the test yourself. My triage verdict, deep dive. Because the phenomenon is measurable. The research question is directly actionable for brand and content marketers, and the conflict of interest, while significant, doesn't invalidate the core methodology. Treat the OmniEG case study as illustrative, not as independent evidence. That's our briefing for today. All five papers earned a deep dive verdict, which is unusual. Here's how I'd prioritize. If you're doing AI assisted creative work, start with paper one on personalized AI. If you sell on AI-powered marketplaces, paper two on vertical, tacit collusion is the most urgent read. If you're running paid or creative campaigns with AI tools, papers three and four, gender bias in AI workflows and AI image disclosure effects are the most immediately actionable. And if your brand has a visibility problem in generative search, paper five on the existence gap gives you a framework for diagnosing and addressing it.
SPEAKER_00As always, these are paper briefings, not final academic reviews. Papers 1, 2, and 5 are unreviewed preprints. Paper 3 is from a conference and has likely been peer reviewed. Paper 4 is from a peer reviewed journal with noted limitations on sample size and cross cultural generalizability. Links to all sources are in the show notes. Read the originals before making major decisions.