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 — 2026-05-12
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. Today we screened 120 papers and selected five worth your attention. Today's theme: AI and Marketing, how AI is reshaping creative work, competitive dynamics, consumer trust, and market stability all at once.
SPEAKER_01Quick reminder: 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 deeper read. Paper one is Personalized AI Scaffolds Synergistic Multi-Turn Collaboration in Creative Work by Sean Kelly, David DeKremer, and Christoph Riedel. It's a preprint posted to Arsiv, not yet peer-reviewed, full text available. Why should marketers care? The research question is straightforward. Does giving an AI assistant detailed information about who you are, your personality, your work style, your skill level produce better creative output than opening a blank chat window? Now, to test this, the researchers ran a randomized controlled experiment. 331 participants were split into three groups. Group one used a generic AI. Group two used a partially personalized AI. Group three used a fully personalized AI, where personalization came from a psychometric survey and an AI-guided interview conducted before the task started. All groups then completed a marketing campaign task for a fictional startup. Campaign quality was evaluated by an LLM judge validated against human expert ratings, and causal mediation analysis was used to figure out why personalization worked, not just whether it did. The findings, fully personalized AI produced higher quality and more creative campaigns than generic AI, and the human plus personalized AI team outperformed what the AI could produce on its own. The mechanism is not mysterious. Personalization helped the human and the AI stay on the same page across multiple exchanges, better attention to what was said, better context tracking, better joint decision making. The paper also raises a concern that generic AI may homogenize creative output, making everyone's work look similar, while personalized AI counteracts that by drawing out each person's unique strengths. Limitations. It's a preprint. The task was for a fictional startup, not a real campaign with real stakes. The LLM judge is validated but not identical to full human expert review. Participants were not working professional marketers. And the personalization process required a pre-experiment survey and interview, which takes time.
SPEAKER_00Okay, so here's what this means for anyone managing creative work with AI. The gap between open a chat window and ask versus tell the AI who you are before you start is not small. This study found measurably better and more creative output. Measurably better. So your Monday morning change is this. Before you start a campaign brief or a creative session with AI, write two or three sentences about yourself, your experience level, the kind of creative you lean toward, what you care about in this category. That briefing paragraph is not just color. This research suggests it materially changes what comes back. If you manage a team, take it one step further. Build a lightweight onboarding process where each person has a short profile, skills, voice, domain knowledge that gets injected into their AI sessions automatically. The cost is low, a 15-minute intake form. The upside per this study is real. And if you're evaluating AI platforms for your team, this is now a selection criterion. Does the tool let you store user context and reuse it across sessions? Or does everyone start from zero every time? Generic, one size fits all AI consistently underperformed here. My triage verdict, deep dive. Because this is one of the more rigorous experimental designs we've seen on the topic. Randomized, causally analyzed, directly testing a marketing task. Preprint status and fictional task context are real caveats, but the findings are immediately actionable.
SPEAKER_01Paper two is Vertical Tacit Collusion in AI-mediated markets by Felipe M. Afonso, also a preprint posted to ArcOV, not yet peer-reviewed. Full text available.
SPEAKER_00Why should marketers care?
SPEAKER_01The research question. So this is not purely theoretical. The bias inputs come from measured behavior of models live in production today. The key finding, when a platform controlling product, rankings, and sellers controlling product descriptions each independently optimize for what earns them more money, the harm to consumers is more than double what either party could cause on their own. They amplify each other without any coordination, specifically on position bias. One benchmark found Claude Sonnet for picked products from the top half of listings 77% of the time versus 23% from the bottom half. GPT 4.1 and Gemini showed similar patterns. Neither the platform nor the sellers need to understand why their tactics work. They keep doing what gets rewarded, and the AI's predictability does the rest. The antitrust problem: current law looks for signs of coordination or agreement. None happens here, so regulators have no clear legal mechanism to intervene. Limitations, it's a preprint. It's a simulation, not a live marketplace study. Bias parameters are measured from current model versions, which will change as models are updated. The model simplifies real-world complexities, including consumer heterogeneity and platform competition, and actual consumer welfare loss in live markets is not empirically measured. The harm estimates come from the simulation itself.
SPEAKER_00I want to be direct here. If you sell products on a platform where AI recommends what to buy, and that includes Amazon Rufus, ChatGPT shopping, perplexity, your listing position and your description language now matter for an audience that is not human. Not human. AI shopping agents are not browsing like a person does. They are heavily biased toward whatever appears first, and they respond to anchoring language in descriptions in ways that are predictable across millions of interactions. That is not an opinion in this paper. Those bias magnitudes come from published benchmarks of real, deployed models. So what do you do Monday morning? First, audit how AI shopping assistants actually present your products. Test it yourself. Ask Rufus, ask ChatGPT shopping, ask Perplexity to recommend products in your category, and see where you show up and how you're described. Second, your product descriptions need to be written for AI readability, not just human readability. That means clear, structured language, anchor price positioning, and leading with the attributes that match common purchase queries. Third, and this is the longer term play. Flag this to your legal and compliance team now. The regulatory picture on AI shopping is moving fast, and brands aggressively optimizing for AI biases may face scrutiny before they expect it. My triage verdict. Deep dive. Because the finding is novel, the bias parameters come from real models, and the competitive implications for any brand in e-commerce are immediate. Preprint and simulation only are real caveats.
SPEAKER_01Paper 3 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 and four co-authors. It's a peer-reviewed journal article published in the International Journal for Research in Applied Science and Engineering Technology, IDRA set, in 2026. Full text available, though the text was truncated in what we reviewed.
SPEAKER_00Why should marketers care?
SPEAKER_01The research question: When you tell consumers an ad image was made by a human, made by AI, or made by both working together, a hybrid, how does that disclosure change their trust in the brand and their intention to buy the product? The method is a factorial experiment, two by three between subjects design. Two hundred and fourteen participants were each shown a single advertisement labeled with one of three creator types, human, AI, or hybrid, for one of two product categories, fast food or cosmetics, and skincare. Trust and purchase intention were measured on self-report scales. The sample was likely a convenience or student sample from Lovely Professional University in India. The findings, human labeled ads scored an average trust rating of 5.63 out of 7 versus 4.24 for AI labeled ads, a gap of about 1.4 points. Statistically very reliable at P less than 0.001. Purchase intention followed a similar pattern, with human ads scoring 5.62 versus 4.65 for AI ads. Now here's the finding worth paying close attention to.08, but did not significantly improve purchase intention compared to fully AI-made ads. In other words, the human involvement label made people trust the ad more, but that extra trust did not automatically translate into more willingness to buy. Consumers also perceived AI-made ads as lower effort and less authentic. Hybrid ads landed in the middle on both dimensions. Limitations. The sample is 214 people, likely students in India. Generalizability is limited. Each participant saw only one ad, so there's no data on repeated exposure effects. Only two product categories were tested. Purchase intention is self-reported, not actual behavior. And the full text was truncated, so some methodological details could not be fully verified.
SPEAKER_00There are two things to take from this paper, and they point in slightly different directions. First, if you're using AI to create ad images and you're disclosing it, or you may be required to disclose it soon, labeling the work as made with AI and human creative direction earns meaningfully more trust than a pure AI label. That's not a big production change. You're not changing the image. You're changing the disclosure framing. Do that. Second, and this is the counterintuitive part, do not assume more trust equals more sales. This study found the trust bump from the hybrid label did not move purchase intent significantly over a pure AI label. So if your goal is conversion, the label framing helps your brand perception but may not directly help your bottom line. What that tells me is this run your own tests on conversion, not just sentiment surveys. The trust gap is real, and it likely matters more in high consideration categories, skincare, health, anything where authenticity is a core purchase driver. In fast food or lower stakes categories, the gap may be less meaningful. Know your category before you change your workflow. My triage verdict, deep dive. Because the three-way comparison between human, AI, and hybrid ads is genuinely novel. Marketing teams are navigating right now.
SPEAKER_01Paper four is AI-driven digital marketing and responsible consumption, the mediating role of marketing intelligence in advancing SDG 12 by Ephraim Habtimikhail Reda, published in Sustainability, a peer-reviewed MDPI journal in 2026. Full text available, though the text was truncated.
SPEAKER_00Why should marketers care?
SPEAKER_01The research question Does AI-driven digital marketing directly improve responsible consumption outcomes? Outcomes aligned with the UN's Sustainable Development Goal 12, which covers responsible consumption and production? Or does it only work through specific types of AI marketing intelligence? The method is a quantitative survey of 120 marketing and business professionals at multinational corporations operating in South Africa. The study tests a mediation model, examining whether AI-enabled marketing intelligence, specifically predictive consumer analytics, and sentiment-based consumer understanding, explains the relationship between AI marketing tools and responsible marketing outcomes. The findings AI-driven digital marketing does not directly improve responsible consumption outcomes on its own. Deploying AI for personalization and ad targeting alone does not make marketing more ethical or sustainable. The only pathway that showed a significant effect was through sentiment-based consumer understanding, using AI to genuinely understand what consumers feel, value, and care about, including their ethical concerns. Predictive analytics, using AI to forecast purchase behavior, showed no significant effect on responsible outcomes. So the distinction the paper draws is between AI as a listening and understanding tool versus AI as a targeting and conversion tool. Limitations: the sample is 120 self-reporting professionals, which is small for a mediation analysis. All respondents are from multinationals in South Africa, which limits geographic and sector generalizability. The design is cross-sectional, so causality cannot be established. Social desirability bias is likely. Professionals may overstate their firm's sustainability commitment. And the full text was truncated.
SPEAKER_00I want to be honest about where this paper fits. Small sample, one country, self-reported, cross-sectional. So I am not going to tell you this paper proves anything about what AI tools to buy. What I will say is that the distinction it draws is worth keeping in your head. If you have a sustainability mandate or ESG commitments attached to your marketing function, and more and more brand teams do, the question your leadership is likely asking is, is our AI stack helping us be more responsible or just more efficient? This paper suggests the tools focused on predicting purchase behavior are not the answer to that question. The tools focused on understanding consumer sentiment, values, and ethical concerns are. So if you're ever in a budget or strategy conversation about AI marketing tools and someone brings up sustainability or brand trust, the framing is listening AI versus targeting AI. Listening AI versus targeting AI, they are not the same thing. And this study, with all its limitations, is an early signal the distinction matters. My triage verdict, skim later, because the evidence base is genuinely weak, the sample is narrow, and this is more useful for ESG-focused practitioners than for core AI marketing decisions. But the conceptual distinction between predictive AI and sentiment understanding AI is worth filing away.
SPEAKER_01Paper five is Reasonably Reasoning AI Agents Can Avoid Game Theoretic Failures in ZeroShot, Provably, by Enoch Hayunwuk. A preprint posted to Archiv, not yet peer-reviewed. Full text available, though the text was truncated.
SPEAKER_00Why should marketers care?
SPEAKER_01This one needs a bit of setup. The research question: Can off-the-shelf AI language model agents, without any special extra training, reliably settle into stable, strategically rational behavior when they compete repeatedly with other AI agents in markets, pricing, promotions, ad bidding. The method combines a theoretical proof and simulation. The author extends Bayesian learning theory to model AI agents as what he calls posterior samplers. Agents that update their beliefs about what opponents will do based on observed history, then respond to those beliefs. He proves mathematically that under these conditions, agents will eventually converge to Nash Equilibrium, a stable, competitive state where no agent can unilaterally do better by changing behavior. Then he runs simulations using Quinn 3.527B, a small open reasoning model across five repeated competitive game scenarios, including a marketing promotion game, comparing three agent types, a base model, a myopic agent, and the theoretically grounded reasonably reasoning agent. The reasonably reasoning agent reached stable equilibrium behavior. The simpler myopic agent, which only optimized for the immediate next move, did not sustain stable behavior over longer repeated interactions. And the stability guarantees held even when the agent did not know the full payoff structure up front and had to learn from noisy results over time. Limitations: it's a preprint. Simulations use only one model in symmetric self-play. Both agents are identical, which does not reflect real markets where different companies deploy different AI systems. The theoretical convergence guarantees are for eventual equilibrium, with no specification of how many rounds that takes in practice. And the specific conditions required, Bayesian updating, and asymptotic best response learning, are theoretical assumptions. It is not verified that real deployed agents consistently satisfy them.
SPEAKER_00So here's why this matters, even though the paper is heavy on theory. Right now, if you're running programmatic ad bidding, dynamic pricing, or any kind of automated promotional negotiation, you probably have an AI agent making decisions on your behalf in a market where other companies also have AI agents. The question this paper addresses is: will those agents settle into predictable patterns or will they behave erratically and create instability that's bad for everyone? The theoretical answer: agents that track history and adapt based on what they've seen will eventually stabilize. Agents that only react to the last move will not. So the first practical action is to ask your AI vendor or your internal team a direct question. Does our bidding or pricing agent maintain a history of competitor behavior and update its strategy based on that, or is it responding move by move? That distinction is the one this paper identifies as the difference between stable and unstable competitive behavior. The second thing, and this is the more strategic point, if you know your competitors are also using AI agents in the same market and you have enough historical data, you may actually be able to anticipate how their agents will behave over time. That's not speculation from me. It follows directly from the convergence logic in this paper. Build that monitoring capability now, before Your competitors do. My triage verdict: Deep Dive. Because the theoretical contribution is novel, the marketing promotion simulation gives it a direct anchor, and as AI agents take over more of competitive marketing operations, understanding whether they will behave predictably is no longer an academic question. Preprint status and single model simulation are real caveats. That's our briefing for today. Here's the deep dive cue. Paper one on AI personalization and creative work, paper two on vertical tacit collusion in AI-mediated markets, paper three on AI-generated imagery and consumer trust, and paper five on AI agent stability in competitive markets. Paper four, the responsible consumption study, is cued as Skim Later. More relevant when you're working on sustainability or ESG-aligned marketing discussions.
SPEAKER_01As always, these are paper briefings, not final academic reviews. Three of the five papers today are preprints and have not been peer reviewed. One is from a lower tier journal with a limited sample. Links to all sources are in the show notes. Read the originals before making major decisions. See you next Tuesday.