Fine-tuning: what it is and what it isn't

99Visibility Research · Updated October 21, 2026 · 4 min read

Short answer

Fine-tuning is additional training that a model's owner runs on an already trained base model to change how it behaves. It updates the model's weights using a smaller, targeted dataset, such as examples of a task, a tone or a domain. Brands cannot fine-tune the public versions of ChatGPT, Gemini, Claude or Perplexity. Only the companies that run those assistants train and update them.

Key takeaways

Fine-tuning changes a model's weights, so only whoever controls the model can do it.
A developer who fine-tunes a model changes their own copy, not the public assistant everyone else uses.
What a public assistant knows without search is fixed at training time, and no outside brand controls it this quarter.
Your real levers sit at the access stage: crawler permissions, firewall reachability and content that exists in the HTML before JavaScript runs.
Uploading documents, writing a system prompt or pasting text into a chat is context, not fine-tuning.

Definition

Fine-tuning is a second round of training on a model that has already been trained. The owner starts with a general large language model and trains it further on a narrower set of examples. The goal is a model that is better at one job: following instructions, answering in a house style, or handling a specialist domain.

The key word is owner. Fine-tuning edits the model itself, so it requires access to the model and its training process. That puts it out of reach for anyone who only uses the model through an app.

How fine-tuning works

The process follows the same basic steps wherever it runs:

Start with a pretrained base model.
Collect examples of the behavior you want, usually input and output pairs.
Train the model further on those examples, which adjusts its internal weights.
Evaluate the result and release it as a new model version.

Whatever the model learns this way becomes part of its parametric knowledge: baked in, frozen at release, and available with no web access.

That is the difference from context. A document you upload, a system prompt, or a web page an engine retrieves sits in the context window for one conversation. It shapes that answer and then disappears. Fine-tuning changes every later answer the model gives.

Some model providers let developers fine-tune a copy of a model for their own application. That copy belongs to the developer's app. The public assistant is untouched.

Why it matters for AI visibility

A public assistant answers from two places. One is what the model learned in training. The other is what it retrieves from the web at answer time, when it chooses to search. Fine-tuning belongs to the first, and the vendor controls it on its own schedule.

So what reaches the model through training is not a lever you can pull this quarter. The full picture is in our guide to training data versus live retrieval.

What you can control is access, which feeds retrieval:

Per-bot crawler permission. OpenAI's crawler documentation says GPTBot governs training use and OAI-SearchBot governs ChatGPT search, and that each setting is independent of the others.
Content in the raw HTML. Vercel found that ChatGPT and Claude crawlers fetch JavaScript files but do not execute them, so client-rendered content may be invisible to them.
Training controls. Google documents Google-Extended as the control for Gemini training and grounding, and says it does not affect inclusion in Google Search.

There is a limit on the training side too. Nothing published shows that allowing a training crawler improves how a future model describes your brand. The honest label is unmeasured. How brands end up in a model's training set is covered in how a brand gets into training data.

Common misconceptions

"We can fine-tune ChatGPT on our brand." You can't. You can only fine-tune a model you control, and that copy serves your app, not ChatGPT's users.
"Our custom assistant is fine-tuned on our docs." Most tools that answer from your documents retrieve them into context at answer time. That is retrieval, not training.
"Allow GPTBot so ChatGPT knows us." GPTBot governs training only. Search visibility runs through OAI-SearchBot.
"A content tweak will shift what the model believes." A critical survey of 45 GEO studies (arXiv 2607.14035, July 2026) found no technique with a stable, longitudinal, cross-platform effect on discoverability.
"Hidden text can steer the model." Injected or adversarial content is illegitimate. Research on LLM search agents (arXiv 2606.16821) shows manipulated pages can sway answers once retrieved, which makes them a reputational risk to watch for, not a tactic.

One practical corollary: "ChatGPT" is not one model. The app routes users to different models, so check whether a test reflects what people see. See whether tools test the same models as ChatGPT.

Frequently asked questions

Can I fine-tune ChatGPT so it recommends my product?

No. Only OpenAI trains and updates the models behind ChatGPT. A developer can fine-tune a copy of a model for their own application, but that changes nothing about what ChatGPT tells its users. What you can influence is whether ChatGPT search can reach your pages, which depends on OAI-SearchBot permission in robots.txt and whether your host or CDN lets that crawler through.

What is the difference between fine-tuning and retrieval?

Fine-tuning changes the model's weights, so the effect is permanent for that model version and applies to every answer. Retrieval fetches pages at answer time and places them in the context window for one response. Fine-tuning is controlled by the model owner. Retrieval depends partly on whether your pages are crawlable, indexed and readable without JavaScript, which is in your control.

Will a future model update include my latest site changes?

Nobody outside the vendor can say. Training data is gathered before a knowledge cutoff, and any effect arrives a full model generation later, if at all. Nothing published measures whether allowing training crawlers improves brand recall. Changes you make today reach answers sooner through live retrieval, when the assistant searches and your pages are accessible.

How do I see what AI assistants currently say about my brand?

Ask the assistants the questions your buyers would ask, and note whether each answer used web search. Answers given with and without search can differ, so record both where you can. The free checker on this site runs up to three buyer questions and scores the answers, giving you a snapshot to compare against later checks.

Sources

This page is AI-assisted. It was written from a verified evidence base and checked against our editorial policy. Engine behaviour changes often; facts are as of October 21, 2026.