## Overview This case study comes from a podcast interview with Dr. Jonathan Roberts, Chief Innovation Officer at Dotdash Meredith, discussing how one of America's largest digital publishers has integrated LLMs into production systems. Dotdash Meredith owns 48 brands including People, InStyle, Investopedia, Travel + Leisure, and Better Homes and Gardens, reaching approximately 30 million people daily through their websites and serving over 10 billion visits annually. The company has positioned itself as an early partner to OpenAI, using this relationship not just for licensing arrangements but for building production AI systems that enhance their advertising business. ## The OpenAI Partnership Structure The partnership between Dotdash and OpenAI consists of three key components that illuminate different aspects of how publishers can engage with LLM providers in production contexts: The first component involves payment for training data usage. While Roberts describes this as a "decent chunk of money," he emphasizes it's not make-or-break revenue for a large publisher. The principle matters more than the amount—establishing that training on publisher content creates an economic relationship. The second component concerns attribution and citations. This deal was structured before retrieval-augmented generation (RAG) became standard in consumer AI products. Roberts notes that just a year prior, the industry assumption was that LLMs would "just know everything" without needing to cite sources. The shift toward real-time content retrieval and citation has fundamentally changed the publisher-AI relationship, creating ongoing value for content providers rather than one-time training compensation. The third component, and arguably the most interesting from an LLMOps perspective, is collaborative product development—specifically the D/Cipher advertising system. ## D/Cipher: Production LLM for Contextual Advertising D/Cipher represents a sophisticated production deployment of LLM technology for understanding user intent without relying on cookies or identity-based tracking. The system underwent an architectural overhaul in fall 2024, switching to an "AI backbone" powered by OpenAI technology. The core problem D/Cipher solves is a recommendation and personalization challenge: how do you serve the right content or advertisement to a specific user without invasive tracking? Traditional contextual advertising matched ads to page topics (mutual fund ads on mutual fund articles), but this approach doesn't scale because only so many users are reading about any specific product category at any moment. D/Cipher uses LLMs to understand "underlying driving concepts" rather than surface-level topics. For example, someone reading about mutual funds likely has attributes like being a low-risk investor interested in long-term planning. These concepts correlate with interests in retirement planning, fiscal stability, and extend beyond finance into lifestyle categories—cruises, gardening, wine, cooking. The LLM enables extraction of these deeper semantic relationships from content. The system combines this semantic understanding with behavioral data from 10+ billion annual visits across all Dotdash properties. Crucially, this isn't cross-site tracking of individuals—it's pattern learning about user interests and behaviors at aggregate level, then applying those patterns to individual content contexts. Roberts claims D/Cipher "outperforms any kind of creepy identity-based targeting" and that the company guarantees this performance. The system now runs in over half of Dotdash's direct advertising deals. ## Extending Beyond Owned Properties A significant production milestone announced just before the interview was extending D/Cipher's capabilities beyond Dotdash's own content to "the broader premium web." The LLM-based approach allows the company to take patterns learned from their properties and extrapolate user intent across other premium publishers—all without using identity or cookies. This represents an interesting productization of LLM capabilities: using the semantic understanding developed on owned content to create an advertising network that can serve other publishers. The value proposition is bringing "premium" to non-Dotdash publishers by sharing targeting intelligence derived from LLM analysis. ## Grounding and Hallucination Mitigation Roberts provides candid commentary on LLM limitations that informed their production approach. He notes that LLMs "hallucinate, which is just a very polite way of saying they lie" and emphasizes this isn't a bug but an inherent feature of systems that produce "the most likely next set of words." When trained on internet content, LLMs give "the average answer," which Roberts notes from his teaching experience is quite different from "the right answer." The key breakthrough for production deployment came when they discovered that grounding LLMs on their own content enabled the system to say "I don't know" when answers weren't present in the source material. Roberts describes this as the moment when "we actually have a chance to make this thing reliable and useful." This grounding approach is central to their commerce and product recommendation workflows. Dotdash maintains extensive internal testing data—"millions of words of research under the hood that never makes it onto the internet just sitting in Google Sheets"—from product testing labs. A proof-of-concept demonstrated that grounding ChatGPT on this internal testing data produced product recommendations "light years better" than Amazon's Rufus because it provided "clear sourced human first-person perspective quantitative and qualitative research at the skew level." ## Business Impact and Content Quality Incentives A significant theme throughout the interview is how LLM-powered contextual understanding realigns incentives toward quality content. In cookie-based advertising, Roberts argues, there's no incentive for quality because audiences are "already priced out from what they tracked from them over the last three weeks." A clickbait article with the same audience as a deeply researched piece would generate the same advertising revenue. The D/Cipher approach changes this dynamic by demonstrating that "better content drives better outcomes." When advertising value is derived from content quality and user intent rather than user tracking, there's economic incentive to invest in quality journalism and research. Roberts credits this approach with driving Dotdash's growth during a challenging period for digital publishing, noting they "had a great year last year" which he acknowledges "is not the narrative around the internet" for publishers generally. ## Observations on AI Search and Future Directions Roberts offers critical observations about AI search products that inform how publishers should think about production AI. He notes that AI chatbots produce a "flattening and concentrating concentration of information"—if two users submit the same query to ChatGPT, they get identical results, whereas Google searches would lead to different pages based on individual choice patterns. This is framed not as an AI limitation but a "failure of product and user experience design." Traditional search's power comes from offering 10 blue links that let users make differentiated choices based on what's in their heads. AI systems making these choices for users removes "informed consideration or choice." From a commerce perspective, Dotdash sees promising signals: they observe some of the highest click-through rates from ChatGPT commerce recommendations to their detailed content pages. This suggests users want to research purchases rather than accept one-answer recommendations, creating ongoing value for detailed publisher content even in AI-mediated experiences. ## Agent Era Considerations Looking ahead, Roberts discusses production implications of AI agents that can take actions rather than just research. The key barrier he identifies is trust—95% accuracy is acceptable when humans cross-check information, but problematic "when it's got your credit card." He positions 2025 as a year for testing and building agent systems, with real use cases likely emerging in 2026. ## Technical Architecture Notes While the interview doesn't provide deep technical details, several architectural elements emerge: The system processes over a billion words of content across more than a million pieces. The OpenAI integration appears to be via API rather than self-hosted models. The switch to "AI backbone" in fall 2024 suggests replacing earlier NLP/ML approaches with LLM-based semantic understanding. The system operates at scale across 10+ billion annual visits with real-time inference for ad targeting. Extension to external publishers suggests some form of transfer learning or semantic mapping that generalizes beyond training data. ## Critical Assessment The interview comes from a publisher executive with clear incentive to present their OpenAI partnership positively. Claims about D/Cipher outperforming identity-based targeting are not independently verified, though the guarantee language suggests they have internal metrics supporting this. The proof-of-concept for AI-powered product recommendations appears to be exactly that—a proof-of-concept rather than deployed production system. The business model for AI-mediated commerce ("who gets paid for it none of that has been figured out yet by anybody") remains unresolved even within this advanced partnership. Nevertheless, the case study demonstrates a sophisticated approach to production LLM deployment: using grounding to mitigate hallucination, building LLM capabilities into existing products rather than creating standalone AI products, and thinking carefully about user experience design rather than deploying chatbots as a default interface.