AI & Marketing Research with Dr. Eva Wolf
Evidence-led briefings that translate peer-reviewed studies and important preprints on artificial intelligence, generative AI, marketing, advertising, consumer behavior, and business strategy into practical insight. Dr. Eva Wolf explains what the evidence actually says, what deserves a deeper read, and what marketers, consultants, educators, and business leaders can do next.
AI & Marketing Research with Dr. Eva Wolf
AI & Marketing Research Radar — 2026-05-07
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.
Welcome to AI and Marketing Research Radar. I'm your host. Today we screened 140 papers and selected five worth your attention. Today's theme: AI and Marketing. Specifically, how AI is reshaping creative work, market structure, live commerce, research practice, and agile execution.
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 deeper read. Three of today's papers are unreviewed preprints. We'll flag that clearly for each one.
SPEAKER_01Let's get into it.
SPEAKER_00Paper one is Personalized AI Scaffolds: Synergistic Multi-Turn Collaboration in Creative Work by Kelly DeKremer and Riedel. It's a preprint posted to Arco, not yet peer-reviewed. Full text is available.
SPEAKER_01Why should marketers care about this one?
SPEAKER_00The research question is whether giving an AI assistant detailed information about a specific user, their psychological profile, creative ability, domain expertise, and work style produces better creative marketing output than using a generic AI. The method is a randomized controlled experiment. 331 participants were split across three conditions: generic AI, partially personalized AI, and fully personalized AI. Personalization meant feeding the AI a psychometric profile built from pre-experiment surveys and a structured interview about how the person works. The task was writing a marketing campaign for a fictional startup. Campaign quality was assessed using an LLM judge validated against human expert ratings. And what did they find? Participants in the fully personalized condition produced significantly higher quality and more creative campaigns than the generic AI group. And the output exceeded what the AI would have produced alone, which the authors describe as genuine human AI synergy. They also ran causal mediation analysis to identify the mechanisms. Personalization improved outcomes by enhancing what they call collective memory, what the human AI pair retains and reuses attention, meaning what the pair focuses on, and reasoning, how the pair prioritizes choices. Importantly, the underlying AI model was the same across all conditions. The personalization was purely external, structured context fed in as scaffolding, not changes to the model itself. What about limitations? The task was a fictional startup campaign, which may not map cleanly to real professional contexts. The sample demographics are not fully described, so we cannot assess how broadly this generalizes. Campaign quality was rated by an LLM judge, which creates a potential circularity problem when you are studying LLM-based tools. And this is an unreviewed preprint, so treat the findings as preliminary.
SPEAKER_01Here is what this means in practice. If your team is using AI for creative work, campaign ideation, copywriting briefs, concept development, the standard onboarding you give a new team member is probably more valuable than you think when it comes to AI too. This paper suggests that feeding the AI structured information about the individual user, their experience level, their creative tendencies, how they like to work, produces meaningfully better output, not marginally better. The practical move is to build that context into your AI onboarding workflow before anyone sits down to use the tool. That means going beyond the usual write copy for ex audience prompt, briefing the AI on the person doing the work, not just the task. If you manage a team using AI for campaign concepting, draft a standard user profile template that team members fill in once and attach to new project sessions. That is not a niche power user move. This study suggests it is a structural performance driver. The other implication is that these tools are built for conversation, not one-shot queries. Synergy emerged through multi-turn interaction informed by that user context, not from a single well-crafted prompt. My triage verdict. Deep dive. Because this is a well-powered randomized experiment with causal mediation, it is directly about marketing campaign creation, and the finding is specific enough to act on. The preprint status is a real caution, but the experimental design holds up.
SPEAKER_00Paper two is Vertical Tacit Collusion in AI-mediated markets by Felipe Afonso. This is also an unreviewed preprint posted to AR Cocteau in January 2026. Full text is available. What is the paper asking? The research question is whether platforms and sellers, each independently trying to maximize their own returns against AI shopping agents, produce consumer harm that is greater than the sum of their individual effects. The method is a multi-agent simulation modeled as a repeated game, calibrated to published empirical measurements of known LLM biases across GPT 3.5, GPT-4, Claude 3, Gemini Pro, Lama 2, and T5. No human participants. This is simulation-based. The model has three players: the platform, which controls ranking and information architecture, sellers, who optimize their product descriptions and bids, and an AI shopping agent acting on behalf of the consumer. And the central finding? In the simulation, both platforms and sellers discover through independent profit optimization that AI agents can be manipulated through position bias, anchoring, framing, and decoy effects. When both exploit these biases simultaneously, the consumer harm is more than double what either party would produce acting alone. A super additive effect, the paper names vertical tacit collusion. The key wrinkle is that this requires no coordination. Platforms and sellers never communicate. They independently converge on the same exploitation strategy through ordinary optimization, which means it falls outside the scope of traditional antitrust frameworks that require evidence of agreement.
SPEAKER_01That's the part regulators will struggle with.
SPEAKER_00Right. The study also notes that AI agents' biases are far more uniformly exploitable than human cognitive biases. The same input produces the same biased output reliably at scale, which concentrates the attack surface. On limitations, this is a simulation, not real market data. Actual harm magnitudes in live markets are not directly measured here. The model uses simplified action spaces. Calibration draws on third-party published bias measurements, which may not represent all current implementations. And the paper is an unreviewed preprint.
SPEAKER_01Let me take this from two angles, because there are two very different audiences for this finding. If you are a brand or seller competing on platforms like Amazon, ChatGPT shopping, or Perplexity, the practical reality this paper describes is that position in AI returned results is not just more valuable than traditional search. It may be categorically different. The cited data point is that top half results are selected by AI agents 77% of the time per the benchmarks in this paper. That changes how you think about bid strategy, product title optimization, and description language. At the same time, if you are advising clients on AI shopping optimization, anchoring language, keyword placement, or pricing structure designed to exploit AI agent tendencies, you should be aware that this paper frames those practices as a likely future regulatory flashpoint. The pattern the author describes is one regulators will eventually have to act on, even without proving intent or coordination. Do not wait for enforcement to start thinking about where your optimization practices sit on that line. My triage verdict. Because the concept of vertical tacit collusion in AI-mediated markets is novel, the mechanism is clearly explained, and the implications for platform strategy and seller optimization are immediate. The simulation only design is a genuine limitation, so treat the harm magnitudes as illustrative rather than measured.
SPEAKER_00Paper three is From Avatars to Algorithms, Virtual Streamers and AI-enabled consumer behavior in live streaming commerce, a systematic review by Wang, Yep, Liu, and Lee. It was published in the Journal of Theoretical and Applied Electronic Commerce Research in February 2026. It is a peer-reviewed systematic review. Our summary is based on the abstract only. We do not have full text access.
SPEAKER_01What does the review find?
SPEAKER_00The research question is what factors drive consumer responses to virtual or AI-powered streamers in live commerce, and how those factors explain AI-enabled purchasing behavior across platforms and product types. The method is a systematic review following Prisma 2020 guidelines. That is the standard methodological protocol for systematic reviews. It synthesizes 41 peer-reviewed studies. The main finding is that three mechanisms explain how virtual streamers influence consumers. Trait-based trust, meaning whether consumers trust the personality or identity projected by the avatar, perceived social presence, meaning how real or human the interaction feels, and message framing, meaning how information is structured and delivered. The authors combine these three into what they call a triadic integration model. The review also identifies that aligning avatar traits and communication style with the specific product category and consumer expectations is critical to effectiveness. Current research in this area is geographically concentrated, relies heavily on self-reported data, and lacks longitudinal or behavioral measurement, which the authors flag as gaps.
SPEAKER_01How should we treat the triadic model itself?
SPEAKER_00Because this summary is abstract only, I want to be clear, we cannot assess the full details of the search strategy, inclusion criteria, or individual study findings. We should not present the triadic model as empirically validated. It is a synthesized theoretical framework proposed by the authors, not a tested mechanism. And the causal inference limitations across the reviewed studies are real. Most are correlational and self-reported.
SPEAKER_01The practical frame here is straightforward. If your brand is considering or already using virtual streamers, AI avatars, or automated live commerce hosts, this review gives you a three-part diagnostic to audit your setup against trust, social presence, and message framing. On trust is the avatar's personality consistent, credible, and appropriate for your category. A high-energy avatar built for fast fashion may not translate to luxury skincare. On social presence, are the interaction dynamics genuinely responsive, or does the avatar feel mechanical? Even small responsiveness cues matter for whether consumers feel they are in a conversation or watching a recording. On message framing, is the scripting optimized for the live commerce context with clear offer structure and calls to action? The transparency point from this review is also worth flagging directly. Disclosing that a streamer is AI operated is not just a regulatory consideration. The review identifies it as relevant to sustaining user trust. If you are running undisclosed AI streamers and consumers work it out, that is a brand safety risk. Build disclosure into your virtual streamer strategy now before it becomes a forced hand. My triage verdict, deep dive. Because this is a Prisma compliant systematic review of 41 studies on a commercially significant and fast-moving channel. The Triadic model is a useful diagnostic even at the framework level. Main caveat is abstract-only access, which limits how deeply we can evaluate the synthesis quality.
SPEAKER_00Paper four is New Tools New Roles: A Manager's Guide to Harnessing Generative AI for Marketing Insight by Netzer, Blanchard, Duane, Garvey, and O, published in NIM Marketing Intelligence Review in 2026. It is a peer-reviewed practitioner guidance article. Full text is available, open access.
SPEAKER_01Strong author names on this one. What is the article doing?
SPEAKER_00This is not an empirical study with a primary data set. It is a practitioner guidance piece authored by senior marketing academics, and the evidence quality reflects that. The research question is how marketing managers can responsibly use generative AI across the full research workflow from desk research through survey design, conversational data collection, qualitative coding and analysis while managing specific risks. On desk research, gene AI can compress days of literature scanning into hours, but because it reconstructs rather than retrieves information, it can conflate findings and hallucinate citations. Treat all Gen AI desk research outputs as first drafts requiring source verification. On survey and stimuli design, prompt precision is critical. The paper uses a concrete example, prompting Gen AI for visual attractiveness of produce-produced items about freshness and quality instead. Tightly defined prompts with explicit inclusions, exclusions, and context produce dramatically more usable outputs. On conversational data collection, AI interviewers can deliver qualitative depth at quantitative scale, but they introduce risks of prompt variability across respondents and off-script behavior. Pre-testing, reinforcement prompts, and full transcript logging are essential. On coding and analysis, AI-generated code and AI-coded qualitative data must be validated against human benchmarks before use in decisions. And the limitations on the article itself.
SPEAKER_01If your Insights team is using any generative AI tool in the research process, and most are at this point, the single most actionable takeaway is to build a verification step into every workflow stage before outputs move to decisions. That sounds obvious. But the practical gap is that teams often treat AI desk research summaries or AI coded qualitative results as ready to use rather than as drafts. So the prompt precision point is worth taking seriously in briefing culture. If you brief an AI the way you brief a junior researcher, vaguely, you will get the same kind of output you would get from a vague brief. Work that is plausible looking but conceptually off. The fix is the same as with any brief. Define what the construct is, what it is not, and what context the output needs to work within. On data security, if you are running Gen AI on internal data, past survey results or proprietary customer research, confidential consulting reports, using a public chat interface is a real risk. The guidance here is to use enterprise APIs or localized deployments. If your organization does not have that infrastructure yet, that is a conversation worth having with your IT and legal teams before it becomes a problem. My triage verdict, deep dive, because the authorship is senior, the full text is open access, the practical coverage is comprehensive, and insights teams will get immediate value from the risk mitigation guidance. The evidence quality ceiling is appropriately lower than an empirical study, but for a guidance piece, this is exactly what it needs to be.
SPEAKER_00Paper five is the prospect of AI enhanced agile marketing, boosting marketing ROI through customer engagement and sales performance by Pugh, Radix, Umar, Jeremiah, and Kwan, published in Marketing Intelligence and Planning May 2026. It is a peer-reviewed journal article. Our summary is abstract only. Full text requires journal access. The method is a quantitative survey study using partially square structural equation modeling combined with artificial neural networks as a robustness check. The theoretical framework draws on dynamic capabilities theory and the resource-based view. The sample is 317 marketing managers and executives at Chinese e-commerce SMEs. The main finding is what the authors call a parallel dual mediation pathway. AI enhanced agility simultaneously improves both customer engagement and sales performance rather than forcing a trade-off between the two. The paper argues that traditional agile marketing requires trade-offs because resources are finite, and AI removes that constraint by enabling both pathways to run concurrently. The study also finds that AI exhibits both autonomous effects on ROI and synergistic effects. It amplifies existing marketing capabilities rather than simply automating tasks. How tight are the limitations? Significant and worth stating plainly. This is association, not proof that AI adoption caused ROI improvement. Self-reported measures from managers introduce common method bias. And because our summary is abstract only, we cannot assess how the key variables were operationalized.
SPEAKER_01The finding that challenges conventional thinking here is the trade-off argument. The classic tension in agile marketing is that you can optimize for relationship building, engagement, loyalty, long-term brand equity, or you can optimize for conversion and short-term sales, but doing both at once competes for the same resources. What this paper claims is that AI removes that constraint by enabling both pathways to operate in parallel. If that holds up in other contexts, it has real implications for how marketing teams justify AI investment. The case for AI spend has often been made on efficiency grounds, same output, lower cost. This paper frames it differently. AI as a capability multiplier that unlocks performance gains that were structurally unavailable before. The practical test is to look at