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 Ads, Persona Research & Consumer Trust: 3 Marketing Papers
Use Left/Right to seek, Home/End to jump to start or end. Hold shift to jump forward or backward.
Thanks for listening to AI & Marketing Research Radar by Big Plans Media.
I’m Dr. Eva Wolf, and I help marketers, educators, consultants, and business owners turn AI marketing research into practical strategy, smarter workflows, and real business opportunities.
More episodes: https://bigplans.media/ai-marketing-research-radar/
Consulting: https://bigplans.media/ai-marketing-consulting/
Big Plans Media — Where Big Ideas Meet Smart Marketing.
You're listening to Evita, an AI-generated research briefing avatar trained on the research framework and methodology of Dr. Eva Wolfe, marketing professor, AI researcher, and founder of Big Plans Media. Every day, Evita scans emerging research in AI, marketing, consumer behavior, psychographics, and business strategy to identify the most relevant developments, opportunities, and risks worth watching. These daily radar reports are designed to help busy professionals stay informed without having to read hundreds of research papers themselves. And every Friday, join Dr. Eva Wolfe live for her personally recorded weekly AI marketing radar roundup, where she breaks down the biggest stories, explains what actually matters, and shares practical insights and strategic implications for marketers, educators, entrepreneurs, and business leaders. Now here's today's radar report. Here's the signal I can't ignore today. What if the two biggest bets your team is making right now, building detailed customer personas for AI research and trusting AI to write your ads are both quietly working against you. Because that's what today's papers suggest. And there's a third one that goes even deeper into the actual plumbing of how AI decides which ads to show in the first place. We screened 368 papers. Three cleared the full text bar and made the radar. Quick caveat. This is a first pass research briefing, not a final academic review. Two of today's papers are preprint. I'll tell you what they suggest, what they don't prove, and which ones deserve a deeper read. Okay, let's get into it. Paper one. Here's the business question. When your team uses AI to simulate how a customer would react to a message, are your detailed ICPs actually making those simulations better or worse? This is a preprint out of Archive. Bodhacharya and colleagues tested how well LLMs represent persona descriptions internally and whether richer personas led to more accurate simulations of real human behavior. Two approaches. First, they measured how the AI's internal representations shifted as you added attributes. Then they checked whether the AI's predictions actually matched what real humans said in surveys and behavioral data. So what happened? The more detail you added to a persona, the worse the simulation got. I'm telling you. That one consistently lost to a two-field persona, just age and gender. The researchers call it persona manifold collapse. The AI starts treating different people as more similar to each other the more you describe them. You add detail trying to differentiate, and the AI internally flattens them into the same blob. That is not a prompt writing problem. That is a structural failure in how LLMs process persona information. And here's what makes it worse. Which attributes you pick matters enormously. Swap one of three attributes and you get a completely different simulation quality. So the teams building ICPs with whatever feels important, they're basically guessing. That's the part I keep coming back to. You can't just trim your ICP down to three fields and call it fixed. You have to figure out which three fields actually produce reliable simulations for your specific use case. But here's the catch. This is a preprint. The specific models tested and the exact human benchmark data sets aren't fully detailed. So how broadly this generalizes, we genuinely don't know yet. Plain English payoff. Stop building long, detailed customer profiles for AI persona simulations. Simpler personas predict real human responses more accurately. Okay, here's where this becomes commercially interesting. Money move. Build a persona audit service for market research teams. Input your existing ICP, get back the minimal at priority set that actually produces reliable AI simulations. Higher accuracy, lower complexity. Action step. Take one AI research workflow your team runs this week. A synthetic focus group, a message test, a simulated audience reaction, and run it again with just age and gender as the persona. Compare the outputs. That's your baseline check. Evidence check. Preprint, not peer-reviewed yet. Model names and data set sizes aren't fully disclosed. Treat this as a strong directional signal, not a final verdict. Radar verdict, read now. Counterintuitive, directly actionable, and it challenges a core assumption behind a lot of AI-powered market research running right now. This next one also challenges an assumption, but this time about the ads themselves, not the research behind them. Paper two. Here's the business question. When you use AI to generate ads, does the content actually work? And what makes consumers trust it or walk away? This is from a peer-reviewed journal. Wang used two datasets, a curated AI-generated advertisement dataset, and engagement data from Xiaohongs, a Chinese social platform. Four quality dimensions were measured: visual quality, style consistency, semantic accuracy, and creativity. Then regression models looked at what predicted engagement. Across all four dimensions, current AI tools have real room to improve. But the bigger finding is about trust. How people feel about whether an AI ad is real and trustworthy predicts engagement more than how pretty it looks. Sentiment and perceived authenticity drove results more than visual polish. And it's especially bad in high trust industries. Healthcare, finance, luxury. Those are the sectors where an AI-generated ad that feels generic or robotic is most likely to backfire. That is not a creative quality problem. That is a trust problem that becomes a brand problem. But here's the catch. I want to be direct. The venue is a lower-tier journal. Sample sizes for both data sets aren't reported anywhere in the available text, and the scoring methods aren't fully described. This actually bothers me. The finding about high trust industries matches what we see in practice. But I want stronger data behind it before I'd build a strategy around it. Plain English payoff. In healthcare, finance, and luxury, AI-generated ads that feel fake don't just underperform, they erode trust. And that shows up in the engagement data. Here's the business hiding inside the research. Money Move. Build an AI ad trust scoring service, specifically for high-stakes industries, that evaluates generated creative for authenticity and credibility before it goes live. Not a visual quality checker, a trust checker. Action step. If your team is running AI-generated ads in any high trust category, pull the last campaign. Check the comment sentiment, not the click rate. If people are calling it out as AI or feeling it's impersonal, that's your early warning sign. Evidence check. Low-tier venue, unreported sample sizes, and the engagement data is from Xiao Hong Shu, which may not translate directly to other markets. Radar verdict. Use cautiously. The direction is right and worth acting on, but the methodology is too thin to treat this as hard evidence. Okay, paper three is the one that surprised me most. Stay with me, because it looks like a technical paper for engineers, but the business implication is enormous. Paper three, here's the business question. As AI chatbots become a primary way people discover products, how do platforms monetize that without trashing the user experience? And what does it mean for advertisers? This is a preprint from Han and Dai. They built two auction frameworks. I won't read you the names because they're dense. But the core idea is this ad systems for AI generated responses that use quality filtering before any bid is accepted. The mechanism works like this. The system sets a baseline, what the AI would say with no ads at all. Then it only allows an ad to appear if it actually fits the topic and doesn't degrade that baseline answer. If the ad would make the response worse or more confusing, it gets rejected. And in their experiments, this quality gating approach earned more revenue per ad shown and kept the AI's answers much closer to the no ad version compared to existing methods that just inject whatever pays the most. One more thing. It's designed so advertisers have no incentive to lie about their bids. Bidding honestly is the dominant strategy. That's a big deal for platform stability. Here's why this matters right now. We are about to watch search advertising get rebuilt from scratch inside AI assistance. The platforms that figure out how to monetize without destroying answer quality will win. The ones that just inject the highest bidder will face a trust backlash. Not highest bidder wins, highest relevant bidder wins. That is a fundamental shift in what winning an ad auction means. But here's the catch. Zero real users tested. Zero real advertisers. This is entirely computational simulation. The click-through estimates the model relies on may not hold when you deploy this on a live system with messy, unpredictable real-world bids. Genuinely surprising. The quality filter actually increased revenue. I expected a trade-off. The paper argues there isn't one. That's the claim I'd want to see tested in a live environment before I'd bet on it. Plain English payoff. The next generation of AI advertising will reward relevance over budget. So advertisers who write tightly targeted copy will outperform those who just outbid everyone. Here's the monetizable angle. Money move. Build an LLM native ad relevance scoring service. Audit how likely existing ad copy is to pass quality filters on AI platforms. And help advertisers rewrite briefs to improve placement rates before those filters go live. Action Step. In your next AI advertising strategy conversation, add one question. What's our relevance score? Not just our bid. Start building the muscle now, before the platforms force the issue. Evidence check. Preprint. Simulation only, no live users, no real advertisers. Foundational theoretical work. The direction is credible, the numbers aren't validated in the real world yet. Radar verdict. The theoretical framework is solid and the implications are big. But simulation only evidence means you monitor this, you don't act on it today, except to get ahead of the strategic shift it's pointing at. Okay, here's the pattern. At first glance, these papers look separate, but together they show one overarching signal. The assumptions we've been building AI marketing workflows on are cracking. We assumed more detail makes AI smarter. Paper one says no, more persona detail makes your simulations worse. We assumed AI generated ads just need to look good. Paper two says no, they need to feel real. We assumed ad auctions reward the highest bidder. Paper three says that model is getting replaced by relevance gating, not more complexity, more precision. That's the shift. And here's the tension I keep coming back to. All three are pointing at the same failure mode. AI systems being asked to do more than they're currently calibrated to do. More persona attributes than they can represent. More creative polish than they can make feel authentic. More ad revenue than they can generate without degrading the response. In every case, the fix isn't more AI. It's better insertion points, simpler inputs, tighter relevance. The teams that figure that out first are going to have a real operating advantage. But here's what I'd push back on in myself. Two of these are preprints, and one has a low credibility venue. So this pattern I'm describing, treat it as a hypothesis worth testing, not a conclusion worth betting the roadmap on. Here's the playbook from today. One, strip your AI persona prompts down to two or three attributes and compare the simulation outputs to your full ICP. Run the experiment before your next synthetic research project. Two, pull sentiment data, not just click metrics, on your last AI-generated campaign in any high trust category. Look for early signs the trust gap is already costing you. 3. In your next AI advertising strategy conversation, introduce the question of relevance gating. Start building a point of view on how your ad copy will perform when platforms filter by fit, not just by bid. Evidence check on all of that. Two of today's papers are preprints. One has a low credibility venue and unreported sample sizes. Use these to decide what to test, not what to blindly believe. Links to all three papers are in the show notes. Read the originals before making major decisions. Want the human expert take? Join Dr. Eva Wolf every Friday for the AI Marketing Radar Roundup, where she extracts no nonsense, money-making tips, practical strategy, and real business opportunities from the week's research. Subscribe on Apple Podcasts, Spotify, YouTube, and wherever you listen to podcasts. This is Evita for Big Plans Media, and I'll be back in the next radar brief.