AI Overviews Are Publishers
2026-07-17 · 5 min read · Janaina Maia
Germany just did something that every product team building AI outputs should pay attention to. A court ruled that Google's AI Overviews and Perplexity are not search engines. They are media companies. The content they generate — summaries, answers, syntheses drawn from multiple sources — is not pointing users to information. It is producing information. And that makes the platform responsible for what it produces.
This is the first ruling of its kind. A German court looked at an AI-generated summary, asked the question nobody in Silicon Valley wants asked — "who said this?" — and answered: "you did." The platform. The company whose model wrote the words.
If you build any product that uses AI to generate, summarise, or synthesise content that users see, this ruling is about you.
What the ruling actually says
The court classified AI Overviews and Perplexity's outputs as media content under German press law. This is not a fine or a restriction. It is a classification. It means these outputs are treated the same way as a newspaper article: the platform is the publisher, the content is editorial, and the platform is liable for falsehoods.
The distinction the court drew is precise and important. A search engine points you to content someone else wrote. An AI Overview takes content from multiple sources, processes it through a model, and produces new text that did not exist before. The court said: that is not search. That is publishing. When the AI says something false, it is not a link to a false source. It is the platform saying something false.
Perplexity was included in the ruling for the same reason. It does not just retrieve and rank. It generates. And generating, the court said, carries editorial responsibility.
Why this matters for product design
Most coverage of this ruling has focused on the legal implications. Those matter. But the product design implications matter more, because they arrive faster and affect more teams.
Every AI output surface is now a publication surface. If your product generates summaries, answers, recommendations, or syntheses that users see, a court in the EU's largest economy just classified those outputs as your content. You are the publisher. You are liable for accuracy. This is not a future regulation. It is a current legal classification in a major market.
Attribution is no longer optional — it is structural. The ruling implicitly creates an expectation: if you generate content, you need to know what you based it on, and you need to be able to show that chain. Right now, most AI products cannot do this reliably. The models synthesise from vast training data, and the output does not come with a bibliography. The court just made that opacity a liability.
The "we just link" defence is over. Google has historically argued that search results are pointers, not publications. That defence worked when the product was ten blue links. It does not work when the product is a generated paragraph that replaces the need to click through to the source. The court drew the line exactly where product people should draw it: if you write it, you own it.
The provenance problem is now a legal problem
I wrote last week that content provenance is a product surface. Cloudflare's Pay Per Use model requires AI products to trace which content contributed to which output. Germany's ruling makes the same demand from the other direction: if you generate content, you need to know what went into it, because you are responsible for what comes out.
Right now, most AI products cannot answer a simple question: "what sources contributed to this specific paragraph?" The models are probabilistic, not archival. They blend thousands of training documents into a statistical output that sounds confident but cannot cite its own reasoning with precision.
This is no longer just a trust problem. It is a legal problem. A publisher that cannot trace its sources cannot defend itself against a defamation claim, a factual error, or a regulatory inquiry. And in Germany, at least, that publisher is now classified as a media company whether it wants to be or not.
What is a media company, anyway?
A media company is an organisation that produces and distributes content for public consumption. Newspapers, broadcasters, and publishers are media companies because they take information, process it through editorial judgment, and produce something new that their audience reads. They are responsible for what they publish because they chose what to say, how to say it, and what to leave out.
The court's reasoning is straightforward: when an AI model takes information from multiple sources and produces a summary, it is exercising editorial judgment. It selects which facts to include, how to frame them, and what to omit. That is what editors do. The fact that the "editor" is a statistical model trained on billions of documents does not change the nature of the output. The output is new content, produced by the platform, for public consumption. That makes the platform a publisher.
This is not a radical idea. It is what every journalist already knows: you are responsible for what you put your name on. The court just said: the model's output has your name on it, even if the name is just a logo in the corner.
What product teams should do
- Build provenance into your output surfaces. If your product generates text that users see, you need to be able to trace which sources contributed to that text. This is not a nice-to-have. It is a legal requirement in at least one major market, and it will spread. Start with citation links, source attribution, and confidence scores. If you cannot show where the information came from, you cannot defend its accuracy.
- Classify your outputs correctly. Stop calling AI-generated summaries "search results." They are not. They are generated content. The language you use internally and externally matters, because courts and regulators are listening. If your marketing says "we generate answers" and your legal team says "we just link," you have a problem that no disclaimer can fix.
- Audit your liability surface. Walk through every place your product generates content that users see. For each one, ask: if this output were false, who would be responsible? If the answer is "nobody" or "the training data," you have a gap. The German court just said the answer must be "the platform." Make sure your product can meet that standard.
- Design for correction, not just generation. Every media company has a corrections process. If you are now a media company, you need one too. When an AI output is wrong, users need a way to report it, the platform needs a way to correct it, and the correction needs to be visible. Right now, most AI products have no mechanism for this. That is a publishing problem, and it needs a publishing solution.
My take
The AI industry has spent years insisting that it is not a media company. It is a platform. A tool. A search engine. A neutral intermediary that merely connects users to information. The German court just said: no, you are not. You write things. You publish them. You are responsible for them.
This is not a punishment. It is a classification that matches reality. When Google's AI Overview says "the Eiffel Tower is in London" — which it has done — that is not a search result pointing to a wrong page. It is Google saying something false. The same logic applies to any product that generates content its users rely on.
The product teams that will handle this well are the ones that stop arguing about whether they are media companies and start building the infrastructure that media companies need: provenance trails, corrections processes, editorial standards, and accountability surfaces. You do not get to choose whether you are a publisher. You only get to choose whether you are a good one.