AI & Infrastructure
Foundation Models

Foundation Models

A foundation model is a large AI model pretrained on broad data that can be adapted to many tasks. GPT, Gemini, Claude and Llama are foundation models behind most AI assistants and answer engines.
June 26, 2026
Cihan Geyik
Table of Content

A handful of model families sit under most of the AI answers your buyers read. Google’s AI Overviews, AI Mode and the Gemini app all run on Gemini. Microsoft Copilot added OpenAI’s GPT-5 in August 2025. These shared bases are called foundation models. They explain why answers across products sound alike, and the products built on top explain why they still cite different pages.

Same model, different answers. A foundation model supplies language and general knowledge. Each product adds its own search index and citation rules, and those decide which brands get named.

What is a foundation model?

Stanford researchers introduced the term in 2021 in "On the Opportunities and Risks of Foundation Models". They described models that are trained on broad data at scale and can be adapted to a wide range of downstream tasks. Their examples were BERT, DALL-E and GPT-3.

The name stresses that the model is a base, unfinished until someone builds on it. The paper also flagged a risk it called homogenization. When many products share one model, its defects are inherited by every product adapted from it.

Foundation models vs LLMs

Large language models are foundation models trained mostly on text. The wider class also covers image, audio, video and multimodal models. Every LLM is a foundation model, but an image generator can be a foundation model without being an LLM.

Two related labels show up in news coverage. "Frontier model" means the most capable foundation models at a given time. "Open-weight model" means one whose weights are published, such as Meta’s Llama, so anyone can run or tune it.

How do products build on foundation models?

Pretrain on broad data → Tune for instructions and safety → Add a search index and tools → Set ranking and citation rules → Ship as an answer engine

Teams adapt a base model in four main ways: prompting, fine-tuning, retrieval and tool use. AI search uses all four. The model writes, search systems decide what it reads, and product rules set how many sources appear.

Vendors also customize the model itself. In May 2025, Google said it was bringing "a custom version of Gemini 2.5" into Search for AI Mode and AI Overviews in the US. A year later, Gemini 3.5 Flash became the default model in AI Mode.

Which foundation models power AI search?

Answer engineFoundation model familyLinks per answer
Google AI OverviewsGemini, customized for Search~13.7
Gemini appGemini~12.3
ChatGPTOpenAI GPT models~5.2
Google AI ModeGemini, customized for Search~4.8
Microsoft CopilotOpenAI GPT-5, added August 2025~4.5

Links per answer: Ansvisor tracking data, September 15 to October 14, 2026. Copilot’s model is from Microsoft’s August 7, 2025 announcement, which also describes a router that picks the model per question.

Why does the same model give different answers?

Look at the three Google rows. All run on Gemini, and their link counts range from about 4.8 to 13.7 per answer. ChatGPT and Copilot both use OpenAI models and still differ. So the base model doesn’t set citation behavior. Retrieval depth, answer format and product rules do.

For tracking, one "Gemini score" can’t stand in for AI Overviews and Google AI Mode. Each surface needs its own prompt runs and its own baseline.

What happens when a foundation model is updated?

Model updates reach publishers without notice. A new version may write longer lists, hedge more or prefer other source types. Brands then see visibility move with no change on their own site.

The fix is a before-and-after view. Compare the same prompts for a few weeks on each side of a release, then check which cited domains changed. Our guide on how to measure AI search visibility covers the metrics to compare.

Common pitfalls

  • Treating "ChatGPT" or "Copilot" as one fixed model. Both route questions between models.
  • Reading results from one Google surface as true for all of them.
  • Optimizing for a model’s training snapshot alone. Snapshots age, and retrieval decides what’s current.
  • Ignoring open-weight models. Many in-house assistants and smaller search tools run on them.

How Ansvisor helps

Ansvisor tracks each answer surface on its own, so Gemini, AI Overviews and AI Mode results stay separate even though they share a model family. The Gemini visibility tracker shows prompt-level mentions, citations and competitor presence, with changes over time. It’s part of the Growth and Enterprise plans, next to the AI Overviews and AI Mode trackers.

Foundation Model, Base Models, Pretrained Models, Frontier Models, General-Purpose AI Models

Let AI Search Work for Your Growth

Each prompt you win brings more mentions, more citations and more buyers. Ansvisor keeps finding the next one.

AI Answers Mentions Citations Visibility Score AI Share of Voice AI Traffic Revenue
Win Back Search VisibilityTry Ansvisor 14-day Free →✓ Zero risk. Cancel anytime.

FAQ

Frequently asked questions.

What is a foundation model?

A foundation model is a large AI model trained on broad data at scale and then adapted to many downstream tasks. Stanford researchers coined the term in 2021. GPT, Gemini, Claude, Llama and image models such as DALL-E are common examples.

What is the difference between a foundation model and an LLM?

Large language models are foundation models trained mainly on text. The foundation model category is wider and includes image, audio, video and multimodal models. So GPT and Gemini are both, while an image generator is a foundation model without being an LLM.

How are foundation models adapted for specific tasks?

Teams adapt them with prompting, fine-tuning on extra examples, retrieval of outside documents and tool use. AI search products combine several of these, wrapping the base model with a search index, ranking systems and rules about citations.

Do AI search engines that share a foundation model give the same answers?

No. Google AI Overviews, AI Mode and the Gemini app all run on Gemini models, yet they retrieve and cite differently. In Ansvisor tracking data from September 15 to October 14, 2026, AI Overviews averaged about 13.7 links per answer and AI Mode about 4.8.

What happens to AI visibility when a foundation model is updated?

A new model version can change which sources an engine prefers, how many brands it lists and how it phrases comparisons. Visibility can shift overnight with no change on your site. Tracking the same prompts before and after an update shows what moved.

Help us grow the AI Visibility Glossary

New terms are added regularly.

Help us improve the page or suggest a new term →
About the Author
Cihan Geyik

Cihan Geyik

Co-founder at Ansvisor

Cihan Geyik is the co-founder of Ansvisor, an open-source AI Visibility platform for AI Search. With more than 15 years of experience in digital marketing and growth, he writes about AI visibility, AI search, AEO, GEO, citations, and answer engines. He focuses on helping brands understand and improve their presence across ChatGPT, Gemini, Google AI Overviews, Bing AI Mode and other AI answer engines.

Google Preferred Sources
Get more Ansvisor insights in Google

Add Ansvisor as a preferred source and Google will show you more of our AI search, AEO and GEO guides.