Publishers fight Big Tech with small local language models

As 2025 closes, referrals from social media and organic search are dead or dying, and generative AI is coming for facts. But 2026 may grant publishers an opportunity Silicon Valley has persistently ignored: local knowledge.

Journalism and Big Tech have long been frenemies. For 15 years, Facebook and its peers have wielded immense market power behind polite smiles and self-serving terms. But the wheel of progress turns, and generative AI has recently disrupted news publishers and tech platforms alike. The AI bubble may soon pop, but conversational interfaces powered by large language models (LLMs) are here to stay, and with them, an opportunity for publishers to break free from the grip of the tech titans.

The key is the Model Context Protocol (MCP), an open-source project from Anthropic that allows generative AI tools to interact with more traditional software systems via any standard application programming interface (API). Barely a year old, MCP has seen rapid market adoption and the support of major platforms from Azure to WhatsApp.

The magic is that an MCP server is a dictionary, translating GenAI requests into actions that the API of an external service can provide. In effect, it makes LLMs infinitely extensible via seamless integration with any digital tool available on the internet.

Software developers have been the first to adopt the tool, and can “create a Jira ticket for a WordPress site, build it in a GitHub repo, register a domain on AWS, and deploy the app to EC2” on command. That prompt is an oversimplification, but not an exaggeration.

For news consumers, it could mean asking Siri, Google, or ChatGPT for the latest news and seeing updates from their preferred local or regional news sources. Or: “What’s being built on Elm Street?”, or “When is the farmers market open?”, or any other question tied to specifically local interests. This everyday information is invaluable to the community, but its commercial value is tied to the local proximity and so rarely appears in the large datasets that feed search indexes or train LLMs.

But think of a local newsroom as a human LLM. Journalists collect, organize, and publish select details across a vast array of local topics. Beyond decades of news archives, our digital shelves include event calendars, obituaries, verified lists of local people, places, and institutions, civic meeting agendas and minutes, election results, building permits, restaurant inspections, local ordinances, development projects, and more.

Right now, the value of this information remains largely untapped on our own sites, and readers rarely come to us for it — they’re on other platforms when they ask the questions our local data and reporting might answer. Either individually or in regional collaborations, newsrooms should create knowledge bases — structured repositories of information — trained on local reporting and local data, available to the community through freemium or subscription products. “Subscribe to our website and get access to our local knowledge base — now also available on your favorite chatbot or search engine.”

These local services will run on small language models. SLMs are cheaper to build, easier to maintain, and grounded in a narrowly defined domain, making them far less prone to factual improvisation than LLMs. By design, SLMs are only economically viable at a local level, giving large tech platforms little incentive to compete in the space. What they will have is an incentive to provide their users access to this layer of local intelligence — so long as the administrative and financial demands are reasonable.

And that is the power of open standards. MCP can be thought of as RSS for LLMs: a lightweight, universal way for any model or chatbot to discover, connect to, and use local structured knowledge without bespoke integrations, contract negotiations, or exclusive partnerships. Signup can be automated. Payments (if any) become small, predictable, and standardized. This lowers the barriers for publishers and platforms, and gives readers the choice to enrich their chatbot with trusted local intelligence.

If publishers embrace small language models and open standards, they may regain some control over how local knowledge is collected, delivered, and valued. For decades, news organizations have tried to win while playing by Big Tech’s rules, but MCP and SLMs give them something new in the digital era: a home field advantage. The platforms own the pipes, but publishers can own the intelligence that matters most to our communities.

Local knowledge is journalism’s superpower. Newsrooms that invest in structured data, local SLMs, and MCP-enabled delivery will define a new, durable model for digital journalism, free from platform dependency and focused on accurate and trusted information about the places people actually live.