What Is a Foundation Model?

A foundation model is a large AI model trained on broad, web-scale data that serves as a general-purpose base for many downstream applications. Rather than building one model per task, developers adapt a single foundation model — through fine-tuning, prompting, or retrieval — into chatbots, search engines, coding assistants, and agents. The term was coined by Stanford researchers in the 2021 report "On the Opportunities and Risks of Foundation Models."

Which model families should marketers know?

A handful of families sit under nearly every AI surface where brands appear. Knowing which foundation model powers which engine tells you where one optimization effort propagates.

Why one base model means leverage for GEO

Because engines are adaptations of shared bases, your brand's representation in a foundation model surfaces in many products at once. Improve how the Gemini family understands your category and the gain shows up in AI Overviews, AI Mode, and the Gemini app together. The same logic applies to Llama: its open weights are embedded in thousands of third-party assistants that inherit whatever the base model learned — or never learned — about you.

Two layers to influence, two clocks

Foundation models acquire brand knowledge through pretraining on corpora like Common Crawl, Wikipedia, and licensed datasets — a slow path where changes take effect only at the next training run, often many months out. Retrieval layers bolted onto those models update in days. A sound GEO strategy works both clocks: durable third-party coverage for the parametric layer, crawlable and citable pages for the retrieval layer. Treating them as one channel is the most common measurement mistake in AI visibility programs.