# Fine-tuning

Category: [AI Search & GEO](https://seo-geo-glossary.com/en/glossary/category/ai-search-geo/) › [AI & LLM basics](https://seo-geo-glossary.com/en/glossary/category/ai-search-geo/ai-llm-basics/)

**Fine-tuning** in AI means continuing the training of an already pretrained model on additional, task-specific examples. The result is a new version of the model that handles a particular task, vocabulary or style better than the base model. The word is also used in physics and theology, which this entry does not cover.

## How does fine-tuning work?

IBM describes fine-tuning as adapting a pretrained model with a smaller, task-specific dataset and as a subset of transfer learning, which reuses existing model knowledge instead of training from scratch. Common approaches are:

- **Full fine-tuning:** updates all parameters. It is effective but needs a lot of computing power.
- **Parameter-efficient fine-tuning (PEFT), for example LoRA:** changes only a small part of the parameters, which reduces the effort.
- **Instruction tuning:** teaches a model to follow instructions using pairs of prompts and desired responses.
- **RLHF:** human feedback steers qualities such as helpfulness or tone.

IBM also lists risks. A model can forget earlier knowledge (catastrophic forgetting) and can overfit on small datasets.

## Fine-tuning, prompting and RAG

Prompting gives the model instructions and context with every request. Retrieval-augmented generation (RAG) fetches matching documents at the time of the request. Fine-tuning changes the model itself. Databricks and Progress both name data that changes often and the need for source citations as reasons to prefer RAG. Fine-tuning fits when a consistent style or a specialized domain is required. The two can be combined. According to OpenAI's documentation (as of October 2026), OpenAI is winding down its fine-tuning platform and new users cannot access it.

## What does it mean for SEO and AI visibility?

A website cannot be added to someone else's model through fine-tuning. Providers or companies with their own data run it. For websites, the routes into AI answers matter more, namely training data and retrieval at the time of the request (assessment).

## Example

A support team fine-tunes a model on several hundred past replies so it answers in the house tone and a fixed ticket format. Prices and delivery times would go stale inside the model, so the system passes them in per request via RAG.

## FAQ

**What does fine-tuning mean?**
Training an existing model further on specific examples so that it fits one task or style.

**Is fine-tuning the same as training?**
It is a second training phase on an already trained model, usually with much less data than the original training.

## Related terms

- [Large language model (LLM)](https://seo-geo-glossary.com/en/glossary/large-language-model/index.md)
- [Training data](https://seo-geo-glossary.com/en/glossary/training-data/index.md)
- [Retrieval-augmented generation (RAG)](https://seo-geo-glossary.com/en/glossary/retrieval-augmented-generation/index.md)
- [Embeddings](https://seo-geo-glossary.com/en/glossary/embeddings/index.md)

## Sources

- [What is fine-tuning? - IBM Think](https://www.ibm.com/think/topics/fine-tuning)
- [What is fine-tuning? - Databricks](https://www.databricks.com/blog/what-is-fine-tuning)
- [RAG vs fine-tuning - Progress (secondary source)](https://www.progress.com/blogs/rag-vs-fine-tuning-choosing-right-ai-strategy-your-data)
- [Model optimization (fine-tuning) - OpenAI Docs](https://developers.openai.com/api/docs/guides/model-optimization)

---

Last updated: October 7, 2026
URL: https://seo-geo-glossary.com/en/glossary/fine-tuning/