## Overview Patch is a hyperlocal news platform that has been operating for nearly 15 years, providing community-focused news coverage across the United States. In this podcast interview, CEO Warren St. John discusses how the company has leveraged AI to dramatically scale its operations from serving approximately 1,100 communities to over 30,000, while maintaining editorial quality and reader trust. The case study offers valuable insights into how AI can be deployed in production for content curation and summarization at scale, particularly in domains where trust, accuracy, and local relevance are paramount. The company's mission centers on creating a sustainable business model for local and hyperlocal journalism—a sector that has struggled financially since the shift to digital and the rise of social media platforms. Unlike nonprofit local news organizations that rely on donations, Patch operates as a bootstrapped, profitable business that must generate its own revenue through advertising and other products. ## The Problem: Scaling Hyperlocal News Economically Before AI, Patch faced what St. John calls "the hyperlocal conundrum": the better you are at hyperlocal coverage, the fewer people it's relevant to. Expanding to new communities was expensive because each new community required investment in audience building, trust establishment, and eventually sales—with no guarantee of profitability. The traditional model required hiring editors, building audience from zero, and waiting for advertisers to trust the new publication enough to buy ads. Additionally, local news consumers face a fragmented information landscape. Relevant information might be scattered across school websites, police department Facebook feeds, nonprofit blogs, local newspapers (often paywalled), and various other sources. Readers want to stay informed about their communities but don't have time to assemble this information themselves. ## The AI Solution: Curated Aggregation, Not Generation A critical distinction in Patch's approach is that they explicitly do not use general-purpose LLMs like ChatGPT to generate newsletters from scratch. St. John emphasizes this point multiple times during the interview, noting that simply asking an AI to "build a newsletter for this town" would result in significant problems with accuracy, relevance, and trust. Instead, Patch's AI system operates on a curated corpus of hyperlocal data that the company has vetted and trusted. The data sources include: - News headlines from trusted local and regional publishers - Community calendars where people add their own events - Social media sources including Nextdoor's training stories API and X (Twitter) - Various other local information sources The AI's role is to process, summarize, and format this pre-vetted information rather than to generate content from scratch. This approach addresses several critical LLMOps challenges: **Geographic Disambiguation**: St. John notes that approximately one-third of US communities share a name with another community in a different state. He mentions a specific user complaint from Greenport, New York—noting there are actually two Greenports in New York alone. The system must correctly associate news with the right geographic location, which requires careful data pipeline design rather than relying on LLM inference. **Content Bundling and Deduplication**: For major news stories, there might be 15 different articles from various sources—from the New York Times to local nonprofit blogs. The system performs preprocessing to bundle related stories and ensure the newsletter provides breadth of coverage rather than repetitive links to the same story. **Editorial Voice Preservation**: The AI newsletters maintain the sensibility and voice of human-produced Patch content. St. John describes this as "plain spoken, straightforward, capable of humor and lightness"—not formal or stuffy. The newsletters even include a daily "corny riddle" (question at top, answer at bottom) that originated as an internal joke but became a beloved feature that readers complained about when briefly removed. ## Newsletter Content Structure Each AI-generated newsletter contains several distinct sections: - **News Headlines**: Brief summaries with links back to originating publishers - **Weather**: Local weather information - **Local Events**: Upcoming events in the community (identified as particularly valued by readers) - **Nearby News**: Widened geographic scope covering adjacent communities - **Nearby Events**: Events in surrounding areas - **Chatter Section**: Curated social media content showing local online conversation ## Production Operations and Feedback Loops Patch has implemented several mechanisms for monitoring and improving their AI newsletter system in production: **Granular Feedback Collection**: Each news story includes thumbs up/thumbs down buttons for user feedback on relevance. Additionally, a newsletter-level feedback mechanism asks readers how "caught up" they feel on community happenings after reading. **Performance Metrics**: Over a six-month lookback period, the AI newsletters achieved a 93.6% thumbs-up rating. This high satisfaction rate suggests the curation approach is working effectively. **Qualitative Feedback Analysis**: When users provide thumbs up/down feedback, they can explain their reasoning in their own words. Patch uses word clouds and ChatGPT summarization to analyze these responses. The most common complaints are: - "Not much happening" (a function of community activity levels, not the AI) - Frustration with paywalled source links (which Patch maintains intentionally to support local journalism ecosystem) **Editorial Overlay**: In states like Georgia, a breaking news editor sits on top of the AI newsletters and can publish to any community or statewide. In high-density states like Connecticut and New Jersey where Patch has large news teams, human reporters provide additional hyperlocal coverage on top of the AI-curated base. ## Infrastructure and Organizational Approach St. John provides interesting insights into how Patch approached building this AI initiative from an organizational perspective: **Parallel Development**: Drawing on Clayton Christensen's disruption theory (Patch's chairman Charlie Hail studied under Christensen and was a co-founder of jobs-to-be-done product development), the company kept the AI initiative separate from the main organization during development. This prevented the "antibodies" of existing systems from reshaping the innovation to look like current products. **Separate Tooling**: Rather than forcing the new initiative into Patch's existing Salesforce instance, they set up a separate HubSpot account to handle the different data formats and workflows. St. John emphasizes that these seemingly minor details about systems integration are actually critical to enabling truly innovative work within established organizations. **Gradual Internal Communication**: The company deliberately avoided making a big internal announcement about the AI initiative early on, recognizing that sharing news about uncertain experiments mainly stresses employees out. Only when positive signals emerged did they bring the broader team into the loop. ## Trust and Differentiation A recurring theme is the importance of trust in local news, and how AI raises the stakes. St. John notes the proliferation of "pink slime" politically-motivated local news sites where you often can't determine who publishes them or even if they're based in the United States. Against this backdrop, Patch's 15-year track record in established communities provides a trust advantage. For new communities, the AI newsletters serve as a low-cost way to begin building trust—showing up consistently, demonstrating political neutrality, and providing accurate information over time. St. John compares this trust-building process to how people evaluate new neighbors: "you just need to kind of get to know them and see are they who they say they are." ## Business Model Considerations The AI newsletters don't just expand reach—they fundamentally change the economics of entering new markets. Previously, opening in a new community meant: - Zero audience on day one - Full editor costs - Long ramp to advertising revenue - Advertiser skepticism about new publications With AI newsletters, Patch can enter communities at much lower cost, build audience and trust gradually, and add human editorial resources in a more sustainable, layered approach as markets warrant. The company has also developed revenue products that work with hyperlocal's inherent limitations. Rather than CPM-based advertising (problematic when your audience for a single community might be small), they offer promoted calendar events and featured business offerings that leverage their geographic segmentation capabilities. ## Lessons and Limitations St. John acknowledges several limitations and open questions: - The AI approach isn't suited for all content types—video about small towns, for example, can't generate enough views to pay for production costs - Some communities simply don't have much happening, and no AI can fix that - Reader frustration with paywalled content is a persistent issue, though Patch views this as important for the broader local journalism ecosystem - The question of whether general-purpose AI (like ChatGPT or AI search summaries) will eventually do this job better remains open, though St. John believes trust and curation will remain differentiators The interview also touches on "peak newsletter" concerns—whether newsletter fatigue will set in as the market becomes saturated. St. John's response focuses on the underlying value proposition rather than the delivery mechanism: if a newsletter genuinely makes people's lives better by keeping them informed efficiently, it will survive regardless of format trends.