Apple Podcasts  ·  Spotify  ·  Buzzsprout

Paper A

Personalized to Persuade: The Effects of Contextualization and Warmth on Trust and Reliance in Conversational AI

Mert Yazan, Suzan Verberne, Frederik Bungaran Ishak Situmeang — 2026 — arXiv (preprint — not yet peer reviewed)

preprint  ·   ·  read now

Key findings

• Tailoring AI responses to a user's background (e.g., framing advice in terms of their profession or context) actually made the AI *less* persuasive when used alone — not more. The AI had a harder time changing people's minds when it personalized its explanations.
• However, when that same personalized AI also used a warm, friendly tone, persuasion came back. The two features together — personalization plus warmth — roughly equaled the persuasiveness of an AI with neither feature. Neither lever is a magic button on its own.
• People relied on the AI's advice (meaning they followed it over the expert's recommendation) across all four conditions. In other words, it didn't matter much what tone or personalization style the AI used — users still deferred to it over human experts.
• People who knew more about AI (higher 'AI literacy') actually trusted the AI chatbot *less*, but still followed its advice *more* and were *more* persuaded by it. Knowing that AI can be flawed did not make these users more skeptical in practice — if anything, it made them more likely to go along with the AI.

Marketing implications

• If you're building or prompting a chatbot to be more persuasive, don't rely on personalization alone — it may backfire. Test warmth and personalization together, not separately, because the combination is what works.
• Don't assume that making your AI chatbot friendlier or more tailored will make customers trust it more or act on its recommendations more. This study suggests reliance happens anyway — people follow AI advice over experts regardless of tone.
• If your customers include tech-savvy, AI-aware users, don't assume they'll be more resistant to AI recommendations. This study found the opposite: knowing about AI doesn't make people push back on it.

Paper B

LLM-HYPER: Generative CTR Modeling for Cold-Start Ad Personalization via LLM-Based Hypernetworks

Luyi Ma, Wanjia Sherry Zhang, Zezhong Fan, Shubham Thakur et al. — 2026 — ArXiv.org

peer reviewed journal article  ·   ·  read now

Key findings

• Brand-new ads normally struggle to get shown to the right people because the system has no click history to learn from. LLM-HYPER solves this by having an AI read the ad's text and images and predict which types of shoppers are most likely to click — without needing any real click data first. In offline tests, this approach ranked ads 55.9% better than the previous best cold-start method.
• The system works by finding past ads that look and sound similar to the new ad, then using those past ads' performance data as examples to guide the AI's predictions. Think of it like showing a new employee a few examples of past successful ads before asking them to judge a new one.
• LLM-HYPER was deployed live at Walmart and performed comparably to the main production ranking system — meaning new ads reached performance levels that normally take weeks to achieve, much faster.
• The AI generates ranking weights before the ad even goes live, so there is no slowdown at real-time serving — the expensive AI computation happens offline in advance, keeping ad serving fast.

Marketing implications

• If you run a large ad platform or marketplace and new campaigns always underperform in the first days or weeks, this approach shows it is possible to short-circuit that lag by using AI to read the ad creative and predict who should see it — before a single person has clicked.
• If you are building or evaluating ad tech tools, look for products that use multimodal AI (reading both images and text together) for cold-start ranking — this paper shows that combination meaningfully outperforms text-only or image-only approaches.
• If you manage seasonal campaigns (e.g., Black Friday ads that go live with zero history), the key practical lesson is: generate your AI-predicted ranking weights before launch day, not after — so the ad hits the ground running.

Paper C

Opening AI: A study of transparency's impact on brand authenticity and trust in visual advertising

Skjerven, Gina, Vindfallet, Linn Carine — 2024

dissertation  ·   ·  read now

Key findings

• When a brand uses AI-generated images in ads but says nothing about it, consumers see the brand as less genuine — and that damaged sense of genuineness then lowers how much they trust the brand.
• Telling consumers upfront that an ad was made with AI does NOT automatically increase trust compared to a baseline — but it does prevent the trust damage that comes from hiding AI use.
• Transparency acts as a damage-control tool, not a trust booster: brands that disclose AI use end up about as trusted as those who never used AI at all.
• Norwegian consumers react negatively when they find out a brand used AI in ads without saying so — describing it as misleading — suggesting disclosure may become a baseline expectation rather than a differentiator.

Marketing implications

• If you're using AI to create images or visuals in your ads, add a simple disclosure — something like 'This image was created with AI.' Not disclosing it risks damaging how real and trustworthy your brand seems if consumers find out.
• Don't expect an AI disclosure label to make your brand look better than competitors — it won't win you extra trust points. Think of it as brand insurance: it protects what you've already built.
• If you manage brand guidelines or ad approval checklists, add 'AI tool used? Disclosed?' as a standard checkpoint before any AI-assisted creative goes live.

Apple Podcasts  ·  Spotify  ·  Buzzsprout