Definition
A foundation model is a large AI model trained on broad, mostly unlabeled data at scale, which can then be adapted to a wide range of tasks through prompting, fine-tuning or connection to tools and data. The term comes from Stanford's Center for Research on Foundation Models, whose August 2021 paper On the Opportunities and Risks of Foundation Models argued that a single pre-trained model was becoming the base on which many different applications are built.
Most foundation models in use today are large language models built on the transformer architecture, and the leading ones are now multimodal, reading text, images and often audio and video. Regulators use a similar concept: the EU AI Act calls them general-purpose AI models and has applied specific transparency and documentation obligations to their providers since August 2, 2025.
As of October 2026, the main model families are:
- OpenAI's GPT-6, in Astra, Sol and Luna tiers, released in September 2026.
- Anthropic's Claude, with Opus 5.5, Sonnet 5.5, Haiku 5.5 and Fable 5.1.
- Google's Gemini, led by Gemini 4 Argon (generally available October 13, 2026) and Gemini 3.8 Flash, plus the open Gemma models.
- xAI's Grok 4.7, released September 21, 2026.
- Open-weight families such as DeepSeek-V4.1-Flash (MIT licensed), Alibaba's Qwen and Mistral's models.
The frontier proprietary models from OpenAI, Anthropic and Google all offer context windows of roughly one million tokens and built-in reasoning with adjustable effort, which a few years ago required separate reasoning models.
A foundation model on its own only knows what was in its training data up to its knowledge cutoff. Products add layers on top: retrieval from a search index, tool use through function calling, system instructions and safety filters. Two products built on the same foundation model can therefore give very different answers. ChatGPT with web search and a third-party app calling the same GPT model without search may describe the same brand differently.
That distinction is useful for GEO. Every AI answer surface, from ChatGPT to Google AI Overviews to Perplexity, runs on one or more foundation models, but visibility depends on two separate things. The first is what the model learned in training, which shapes which brands it associates with a category and how it describes them. The second is what the product retrieves at answer time, which decides which pages get cited. Training influence is slow and indirect: it grows with consistent, widely repeated facts about your brand across sources a crawler can reach. Retrieval influence is faster and more measurable: crawlable pages, clear answers and presence in the sources each platform's search layer favors.
Because model generations now turn over every few months, track AI answers by model and rerun prompts when a new version ships. A change in citation patterns after a model switch is common, and it is easier to diagnose when you know the date the underlying model changed.
Examples of Foundation Models
- A bank builds its internal assistant on a self-hosted DeepSeek-V4.1-Flash model and connects it to policy documents through retrieval, so the foundation model never has to be retrained when policies change.
- An e-commerce company tests the same product-description prompt on GPT-6 Sol, Claude Sonnet 5.5 and Gemini 3.8 Flash, then picks the cheapest model that meets its brand style rules.
- A GEO team notices a jump in citations for its documentation the week ChatGPT switches its paid chat model from GPT-5.6 Sol to GPT-6 Sol, and annotates the date in its reporting.
- A legal-tech startup fine-tunes an open-weight Qwen model on court filings rather than prompting a proprietary model, because it needs to run fully offline for client confidentiality.
