When ChatGPT mentions your company in a response, what actually caused that to happen?
It's not a rhetorical question. The answer has real implications for anyone thinking about brand visibility in the age of AI. And the honest answer is: it depends on which of two very different mechanisms brought your brand into that response.
The first path: what the model learned during training
Large language models are trained on massive text datasets — articles, forums, documentation, news, reviews, product pages, academic papers. During training, the model adjusts its internal parameters to learn language patterns and, implicitly, associations between concepts, entities, and brands.
When your company appears repeatedly in relevant contexts — described consistently, associated with specific categories, mentioned across credible sources — that leaves a trace in the model's weights. There's no database entry saying "Brand X is known for Y." It's more like sediment: enough consistent signal, and the model "knows" that your brand and certain concepts belong together.
The limitation of this path is significant: the window closes. Once a model is trained, the knowledge cut-off freezes everything. Products launched after that date, rebrands, acquisitions, crises resolved — the model simply doesn't know. Depending on the model, that cut-off might be over a year in the past.
You also can't directly control what a model learned from its training data. What you can do is shape the sources that training datasets tend to absorb: high-authority media coverage, Wikipedia, technical forums with strong indexing histories, open data repositories. These sources carry disproportionate weight. A well-maintained Wikipedia page often matters more than a hundred posts on your own blog.
The second path: real-time retrieval (RAG)
Retrieval-Augmented Generation — RAG — is the mechanism by which a model fetches external information at query time, before generating a response. Instead of relying only on what it "learned," the system accesses current sources: web pages, indexed documents, structured databases.
Google's Gemini has native access to the Google search index. ChatGPT with web browsing enabled retrieves pages in real time. Enterprise systems often run RAG over internal knowledge bases. The point is: there's a retrieval layer happening before — or during — response generation, and what gets retrieved directly shapes what the model says.
For brands, this changes the calculus. A company absent from training data can still appear in a response if it has indexed, relevant, and well-structured content available for retrieval. Conversely, a brand that was well-represented in training data can lose ground to a competitor with more current, better-organized content.
Real-time retrieval tends to favor content that answers specific questions directly, pages with clean technical structure, sources with established authority and citation patterns, and structured data that makes fact extraction easier.
How the two paths interact
The training data / RAG distinction is analytically useful, but in practice, many models use both in the same response. What was learned during training influences how the model selects and interprets what it retrieves. What gets retrieved can confirm, contradict, or fill gaps in what the model already "knew."
This creates some nuanced outcomes. A model might surface your brand from training data but pull in an outdated detail. Or it might retrieve your current page and interpret it through the lens of what it previously learned about your industry. The final response is a synthesis — and that synthesis is what the user reads.
What you can actually influence
For training data, the lever is long-term editorial presence. Being consistently cited in relevant, high-authority publications. Maintaining accurate, up-to-date entries in reference sources like Wikipedia. Generating coverage in outlets that appear in the datasets large model providers tend to use. This work compounds slowly — each credible mention adds to the signal — but the payoff timeline is measured in months, not weeks.
For RAG, the lever shifts. Here, what matters is that your content is findable, structured, and genuinely useful at retrieval time. That means content that answers real questions directly (not just marketing copy), schema markup that helps models extract named entities and facts, solid site performance, and presence in sources that grounding-enabled models query. Changes here can show up in AI responses within days or weeks.
The practical implication: both tracks need to run in parallel, but with different timelines and different definitions of success.
The gap in most GEO strategies
When your brand shows up in an AI response, you typically can't tell which path brought it there. The model doesn't annotate its provenance. This creates a real blind spot: you might be investing heavily in new content (optimizing for RAG) while the real problem is that training data carries an outdated or inaccurate picture of your company — or the reverse.
Tracking which version of your brand's story appears in AI responses, how often, in what contexts, and with what factual accuracy is the step most brands are still skipping. Not because it's hard to understand why it matters — but because the infrastructure to do it systematically is relatively new.
Genoma was built to map exactly this kind of presence: which queries surface your brand, what the model says when it does, where there are gaps, and where there are inaccuracies. If you don't yet have that baseline, it's worth establishing one before doubling down on any optimization effort.
The bottom line
Training data and RAG aren't competing mechanisms — they're two parallel vectors operating on different timelines, with different levers, and different implications for your brand strategy. Understanding which one is driving (or limiting) your current AI visibility is the starting point for any GEO effort worth taking seriously.
And that understanding, almost always, starts with measurement.