Apple Podcasts  ·  Spotify  ·  Buzzsprout

Paper A

Machine marketing: rethinking the customer in the age of generative AI

Marko Sarstedt, Susanne Adler, Monika Imschloss — 2026 — Journal of Marketing Analytics

peer reviewed journal article  ·  open access  ·  deep dive

Key findings

• AI systems are crossing a line: they used to just help people shop (think ChatGPT giving advice), but now they can actually make purchases on their own — searching, comparing, and buying without asking the human each time. The authors call these systems 'AI customers.'
• An AI customer is not just a smarter search engine. It follows its own decision logic, which can be different from how a human would decide. So a brand that is great at convincing people might still lose to a competitor if an AI agent rates it lower based on its own criteria — like data structure, pricing transparency, or how machine-readable its product pages are.
• The authors argue that marketing as a field needs a new branch — 'machine marketing' — specifically focused on understanding and influencing how AI agents make decisions, the same way marketing today focuses on understanding and influencing human consumers.
• Consumers are handing over more and more of their buying decisions to AI — choosing what sources to consult, which retailer to use, even completing checkout. This means the AI in between the brand and the buyer becomes a gatekeeper that brands need to actively think about reaching.

Marketing implications

• Your next customer might be an AI bot, not a human. Start auditing your product pages and pricing pages: are they easy for an AI to read and compare? Structured data, clear specs, and transparent pricing help AI agents evaluate you favorably — the same way SEO helped Google find you.
• If you're in e-commerce, test how your brand appears when someone asks ChatGPT or Gemini to recommend products in your category. If the AI doesn't mention you, or mentions you unfavorably, that's a gap worth fixing — just like a bad Google ranking.
• Pay attention to 'generative engine optimization' (GEO) as a real discipline. The paper flags this as something brands will need to invest in — optimizing not just for human readers or search engines, but for how AI agents interpret and rank your offerings.

Paper B

LLM-OSDA: An Optimal-Stopping Dynamic Auction for Native Advertising in Multi-Turn LLM Conversations

Yan Fang, Jialin Chen, Chun Gan, Hang Yu et al. — 2026 — arXiv (Cornell University)

preprint  ·   ·  read now

Key findings

• Deciding when to show an ad during a conversation matters as much as deciding which ad to show. Show it too early (when the user's intent is unclear) and it's wasted; wait too long and the user leaves. The system learns to pick the right moment.
• The new auction system (LLM-OSDA) earned 11% more ad revenue than the best existing approach that uses a fixed moment to show ads — while keeping user drop-off rates about the same.
• Advertisers who bid more don't just have a better chance of winning; they can also influence which turn in the conversation their ad appears in. The system accounts for this so advertisers still have no financial incentive to lie about what a click is worth to them.
• A small neural network (StopNet) can learn to approximate the ideal timing decision, and the authors show mathematically that any errors it makes are confined to borderline cases — so in practice it behaves close to the theoretical ideal.

Marketing implications

• If you run or build a chatbot-based product and want to add ads, inserting them at a fixed point in every conversation (e.g., always after the second message) leaves money on the table — timing the ad to when the user shows clear commercial intent could increase ad revenue by double digits.
• For ad platforms exploring conversational AI placements, this paper gives a concrete blueprint (with open-source code) for building an auction that decides timing and winner simultaneously, rather than treating them as separate problems.
• Advertisers buying placements in AI chatbots should expect that future auction systems may charge them based not just on click probability, but on the option value of the conversation moment — meaning a high bid could buy a better-timed slot, not just a higher-priority slot.

Paper C

PILA: Plug-and-Play Insertion for LLM-native Advertising

Zhaowei Zhang, Yuhan Fu, Yihang Zhang, Xiaohan Liu et al. — 2026 — arXiv (Cornell University)

preprint  ·   ·  test this week

Key findings

• PILA can add sponsored content to AI chatbot answers after the answer is already written — like a lightweight add-on — without touching the chatbot itself. This means it works with any AI service, including closed ones like ChatGPT, where you can't change the underlying model.
• Compared to other methods that try to insert ads while the AI is still writing its answer, PILA beat prompt-only approaches by about 34%, sampling-based methods by about 47%, and fine-tuning approaches by about 8% on a combined score of response quality plus ad effectiveness.
• When PILA was plugged into seven major commercial AI models as a bolt-on module, it improved their combined user-quality and ad performance scores by roughly 17–18% without changing those models at all.
• PILA includes a dial that lets operators choose how prominently ads appear — turning it up makes ads more visible (better for advertisers) but slightly reduces how natural the response feels to users, giving platform operators a concrete lever to tune revenue vs. user experience.

Marketing implications

• If you run a chatbot or AI assistant and want to monetize it with sponsored content, PILA's approach shows it's technically feasible to add ads after the AI writes its answer — you don't need to rebuild your AI from scratch. Look for similar 'post-generation rewriting' tools when evaluating ad monetization vendors.
• When evaluating AI advertising tools, ask vendors whether their solution modifies the underlying model or works as an external layer. The external layer approach (like PILA) is safer for preserving your chatbot's quality and works even with closed APIs like OpenAI.
• Be aware that a tunable 'ad intensity' setting is now a real product feature — if a platform offers this, higher intensity means more prominent ads but a potentially worse experience for your users. Test where that dial should sit before rolling out to your full audience.

Apple Podcasts  ·  Spotify  ·  Buzzsprout