Apple Podcasts  ·  Spotify  ·  Buzzsprout

Paper A

Task To Tech: An Exploration of Generative AI in Tourism Marketing through Student Experiments and Practitioner Interviews

Rifki Rahmanda Putra, Dicky Arsyul Salam — 2026 — Media Wisata

peer reviewed journal article  ·   ·  read now

Key findings

• AI made every marketing task faster — but speed did not always mean better quality. When students used AI, they finished tasks more quickly across the board, but the quality improvement depended heavily on the type of task.
• For writing long blog posts or detailed descriptions, AI helped produce noticeably better results. For short-form content like social media captions, AI made no meaningful difference in quality. For visual design work, AI actually made the outputs worse.
• Tourism professionals said they adopted AI mainly because it matched the kind of tasks they needed to do and because they expected it to improve their performance. Peer pressure, ease of use, and having the right tools and support also played a role.
• How digitally skilled someone is acts as a master switch — people with stronger digital literacy got more out of AI, while those with weaker skills were more likely to produce lower-quality results even when using the same tools.

Marketing implications

• Use AI for drafting long-form content like destination guides, blog posts, or detailed itineraries — this is where it saves time AND improves quality. Skip AI as a design tool for visual assets; instead, have a human designer review or create visuals from scratch.
• Before rolling out AI tools to your team, run a quick skills check. Team members with weak digital literacy need training first — otherwise AI can actually lower the quality of their work.
• If you manage a small tourism marketing team, prioritize AI adoption for text-heavy, research-style tasks (e.g., long emails, landing page copy, travel guides) and be skeptical of AI-generated social captions — they may not perform differently from what your team already produces.

Paper B

Advertising and large language models: a new frontier influencing medical practice

Mason Crossman, Jeffrey Weng, S. Bacchi, Weng Onn Chan — 2026 — Eye

peer reviewed journal article  ·   ·  read now

Key findings

• Unlike a Google search where you see many links and pick what to trust, AI chatbots give you one confident-sounding answer — which makes it much easier for commercial interests to quietly steer what that answer says.
• Someone with enough money could flood the internet with fake reviews, AI-generated articles, and self-promoting content, training AI systems to recommend their clinic or product without patients ever knowing the fix was in.
• OpenAI started testing ads inside ChatGPT in February 2026, and Google already shows shopping ads inside its AI-generated summaries — so AI systems are now officially part of the advertising landscape.
• When patients arrive at a doctor's appointment already convinced by what an AI told them, it affects what treatments they ask for, what they resist, and how much they trust their actual doctor — even if the AI was influenced by commercial content.

Marketing implications

• If you're building a brand or promoting a service, the era of 'LLM SEO' is here — start thinking about whether the content you publish will get synthesized favorably by AI chatbots, not just ranked on Google. The rules are different and less transparent.
• If you run ads in healthcare or any high-trust category, know that patients and customers are increasingly getting their 'research' from a single AI answer. Your reputation in that AI's training data matters more than your ad spend.
• If you manage a brand's reputation, add 'what does ChatGPT say about us?' to your regular monitoring checklist — just like you'd check Google reviews. The AI's answer is now part of your brand perception.

Paper C

Efficient LLM-based Advertising via Model Compression and Parallel Verification

WenXin Dong, Chang Gao, Guanghui Yu, Xuewu Jiao et al. — 2026 — arXiv (Cornell University)

 ·   ·  test this week

Key findings

• By shrinking the AI model's memory footprint and making it process tokens in parallel, the system ran more than 1.8 times faster in real-world tests on Baidu's ad platform — meaning it could handle the same volume of ads in roughly half the time.
• Inference speed improved by over 78% compared to the unoptimized baseline, and the quality of ad recommendations stayed competitive — the system did not lose meaningful accuracy by going faster.
• The key trick for speed: the system figures out the best moment to switch from slow, one-token-at-a-time generation to verifying a whole batch of possible ad tokens at once, using a tree structure of pre-organized ad candidates.
• The model compression (reducing numerical precision from 16-bit to 4-bit numbers while pruning less important parts of the model) shrank the index structure to about 30% of its original size, cutting memory and compute costs substantially.

Marketing implications

• If your ad tech team is exploring LLMs for real-time ad targeting or creative generation, this paper shows that speed is solvable — but only with serious ML engineering investment. Before budgeting for LLM-powered real-time ads, ask your vendor or tech team specifically how they handle inference latency at scale.
• When evaluating AI ad platforms, ask whether they use model compression or quantization in production. A vendor running full-precision LLMs for real-time bidding is likely burning excessive compute costs that will be passed on to you.
• For teams already running generative ad targeting experiments, this paper is a useful benchmark: if your LLM-powered targeting system is not approaching at least 1.5–2× the speed of a naive implementation, there is probably room for optimization before scaling spend.

Apple Podcasts  ·  Spotify  ·  Buzzsprout