Where LLMs pull from (and what it means for Generative Engine Optimization)

When the C-suite asks, “How do we show up in AI answers?” one of the key questions behind it is: where are tools like ChatGPT, Gemini and Claude getting their information about your brand?

They’re not just reading your homepage. They draw on the wider information environment around you: news coverage, analyst notes, review sites, reference pages, community threads and your own documentation.

Understanding where LLMs pull from is the foundation of Generative Engine Optimization (GEO). Once you know which sources carry the most weight, you can be more deliberate about where you earn coverage, how you structure content, and how you respond when an answer is wrong or incomplete.

AI trusts what PR already powers

We see one pattern keep showing up: AI visibility is overwhelmingly driven by non-paid, earned sources.

So before we get into the full source mix, one thing is clear: LLMs lean most on the stories, commentary and validation that PR teams help bring to life.

The core source stack: where LLMs actually pull from

Most web-enabled LLMs draw from a similar set of source types, even if they weigh them differently. The main categories look like this:

• Journalistic: News sites, and other journalistic coverage
• Third-party corporate/blog (earned): content about you on other companies’ sites
• First-party corporate/blog (owned): your own content
• Press releases: on newswires or your site
• Government / NGO and academic: official or research-heavy sources
• Aggregators / encyclopedic: Wikipedia, reference sites
• Social / UGC: Reddit, forums, social posts

The mix changes by question type and industry, but one theme is consistent: earned and independent sources dominate.

1. Journalism: the backbone of many AI answers

Newsrooms are one of the most important inputs into AI answers. For discovery prompts like:

• “Which brands are leading in [category]?”
• “What are the latest trends in [industry]?”

LLMs lean heavily on reporting, explainers and roundups from trusted outlets. What is AI Reading? charts show journalism sitting comfortably in the 20–30% range of citations at almost any point in time.

2. Earned media beyond newsrooms: blogs, reference and niche outlets

AI doesn’t stop at big news brands. It also pulls from:

• Third-party corporate blogs and industry sites – for specialized analysis or product comparisons
• Aggregators and encyclopedias – for “what is…” and background questions
• Niche outlets by industry – for vertical expertise

The December report’s industry slides show that each sector has its own blend of niche sources—education, healthcare, government, hospitality, tech and travel all lean on different domain-specific outlets.

For PR teams, this is classic media strategy, translated into GEO:

• Earn a mix of top-tier and niche coverage
• Make sure core narratives show up consistently across those sources
• Assume those same outlets will feed into AI answers later

3. Owned channels: essential for facts, but not the whole story

Let’s make a key distinction:

• Discovery questions: “What should I buy?”, “Who are the main players?”, “What’s happening with…?”
• Fact-finding questions: “What’s the warranty?”, “How does pricing work?”, “What are the technical specs?”

Brand-specific fact-finding queries are often dominated by owned content. When someone asks about your policies or documentation, AI tends to go to your channels first—which is exactly what you want.

For discovery questions (where GEO really lives), AI leans far more on earned media and broad coverage instead of your own site.

So the GEO takeaway:

• Owned content is critical for clarity and trust on specifics
• Earned content is critical for getting into the answer in the first place
• You need both working together

4. Press releases and structured signals

Press releases used to be a relatively small part of the AI citation mix. That’s changing.

In our Generative Pulse analysis of three major newswires, citations to press releases increased roughly fivefold over the back half of the year, and release-related content overall grew as well.

The releases that end up cited tend to have:

• clear, factual headlines
• meaningful data points
• structured formatting (sections, bullets)
• objective, straightforward language

A release isn’t just a media asset—it’s also a clean, machine-readable artifact that can help LLMs understand what’s new and what’s true.

5. Social, Reddit and reviews: supporting voices

Social and UGC don’t dominate the source mix, but they do matter in specific contexts.

Analysis of AI answers shows that:

• Platforms like Reddit appear frequently among top domains for certain models and industries
• In many answers that cite Reddit, models also cite at least one journalistic source, grounding community discussion in reported context

Community conversations and review platforms often reinforce the narratives that start in journalism, analyst coverage and owned documentation.

For comms teams, it means GEO isn’t just about newsrooms. It’s also about monitoring and participating in the communities your audience trusts and aligning what appears in reviews and forums with the story you’re telling in earned and owned channels.

What this means for your Generative Engine Optimization strategy

Pulling this together into a few core moves:

Keep going: Level up your Generative Engine Optimization strategy

If you want to see the full breakdown of where AI is pulling from—by industry, model and source —the December 2025 edition of What Is AI Reading? is the best next step.

To turn those insights into an ongoing strategy, the Fundamentals of GEO course from Muck Rack Academy walks through how to connect earned, owned and community channels for generative visibility, how to choose the right prompts to monitor and how to report on AI visibility in a way that makes sense to leaders.

👉 Start Fundamentals of GEO to learn how to use what LLMs are reading to shape what they say about your brand.