What Is a Large Language Model (LLM)?

A large language model (LLM) is an AI system trained on very large amounts of text that writes answers by predicting what words should come next. ChatGPT is built on large language models, and so are the AI features in many search tools. For a health practice, the model is the thing that decides whether your name shows up when a patient asks for help.

How a large language model works

Think of an LLM as a very well-read writer with no filing cabinet. During training, it reads a huge amount of text and learns which words tend to follow which. When you ask a question, it does not look up an answer the way a search engine retrieves a page. It writes the most plausible response, one piece at a time, based on the patterns it absorbed.

It helps to picture three layers. Retrieval picks the material, the model writes the answer, and the product decides what the reader actually sees. A practice can do well in the first two and still be hidden by the third, for example when sources sit behind a small icon. The model's stored knowledge is a fixed snapshot (see training data), so the lever you control day to day is what gets retrieved. I keep screen-layout details out of my advice because products change them often.

That has three consequences for a practice owner.

• It has a memory with a date on it. What a model learned in training is frozen at some past point. A clinic that opened last year may not be in its memory at all.
• It can look things up, but only if the product lets it. The original 2020 research paper on retrieval-augmented generation described the model's stored knowledge as parametric memory, and an outside index it can search as non-parametric memory. The authors noted that updating a model's knowledge and showing where an answer came from were open problems. Adding a searchable index was their proposed fix. You can read how that plays out in retrieval-augmented generation.
• It writes fluently whether or not it knows. OpenAI researchers published a paper in September 2025 arguing that models hallucinate, meaning they state plausible falsehoods, because training and scoring reward guessing over admitting uncertainty. They compare it to a student guessing on an exam. See hallucination.

Different products wrap the same kind of model in different ways. ChatGPT can look things up through ChatGPT search when a question calls for current information, and OpenAI runs a separate crawler, OAI-SearchBot, to surface sites in those results. Perplexity searches the internet in real time as part of how it answers. Google's AI features add a model on top of Google Search. So the question to ask is not "does the AI know my practice" but "does this tool, in this mode, have a way to find me."

That is also why two patients can get opposite answers. One asks a tool that is answering from memory and your practice is absent. Another asks a tool that searches, finds your well-built page and a consistent directory listing, and names you. The model is the same kind of system in both cases. The sources it had access to differ.

For you, the practical takeaway is that a model is a mirror of the web at one moment. It does not hold opinions about your practice. It reflects how your practice is described in the sources it saw or fetched. Improve those, and the reflection changes. That is the core idea behind generative engine optimization.

The word "large" refers to scale, meaning the amount of text used in training and the number of adjustable settings inside the model, which are called parameters. You do not need those details to act on any of this. What matters is the behavior. A model produces the most likely next words, so a clear, repeated, consistent fact about your practice is easier for it to reproduce than a vague or conflicting one.

Example: a large language model at a health practice

This example is illustrative, not a transcript. A functional medicine clinic in a Colorado suburb opened 14 months ago. It has a clean website and a handful of reviews. An owner asks ChatGPT with search switched off, "What is [clinic name] in [suburb], and what do they treat?" The model replies with a confident paragraph about a clinic that offers "comprehensive wellness and hormone care," and it lists a website address. The address does not load.

Nothing malfunctioned. The model did what it is built to do, which is write a plausible answer from patterns.

Real data shows how common this is. Health Biz Scale checked 764 established US health practices (one check per practice, test runs excluded) between May 11 and Aug 20, 2026. We asked an AI model, GPT-4o-mini through OpenAI's API answering from memory with no web search, 2,292 patient-style questions. The model named 3,554 practice websites, and 1,153 of them (32.4%) did not load. Only 6 of the 764 practices (0.8%) were named. About 7 in 10 questions named a city. The ChatGPT app searches the web live and can answer differently, so these numbers describe a memory-only model, not what a patient sees in the app.

Two lessons follow. First, a memory-only model fills gaps with plausible inventions, so a wrong answer is a signal about missing sources and not a reason to panic. Second, even established practices were rarely named from memory. Being found by AI depends on search-enabled tools reading your current pages and listings.

For the clinic, the plan is to run the same question with search on, see which pages the tool cites, and close the gaps. Add a clear "who we are and what we treat" page. Make sure the address and name match across directories. Check that OAI-SearchBot is allowed in robots.txt. Then retest.

How to check yours in 60 seconds

1. Open ChatGPT and ask, "What is [your practice name] in [city], and what do they treat?"
2. Compare the answer to your real services, providers and website.
3. Note anything wrong, missing or invented, including any web address that does not load.
4. Ask the same question in Perplexity and compare.

If a tool offers a switch for web search, test once with it on and once with it off.

What to do about it

1. Treat what the model says as a report card on how the web describes you.
2. Fix the source, not the chatbot. Correct directory entries, listings and your own pages.
3. Publish a plain statement of your specialty, location and clinicians on your site.
4. Make sure search bots can reach your pages, since a blanket AI-bot block can remove you from live answers.
5. Keep patient details out of any AI tool unless a business associate agreement is in place, as HHS requires for disclosures of protected health information to a business associate (see business associate agreement).
6. Recheck monthly, since models and tools update.