The terms, without the jargon.

Two phrases from the trade press. They describe one change from two angles.

The agentic web

The shift underway where an AI sits between a person and the internet and does the looking for them. The person asks; the AI reads the pages, decides what matters, and comes back with an answer. Fewer people are browsing their way to a decision, and more are being handed one.What changes: you are no longer being chosen from a list. You are being chosen, or not, by something that read the list on the reader's behalf.

The Great Decoupling

The name the search industry gives to what this does to website traffic. Being found and being visited have come apart. A practice can hold its search position, or improve it, and still watch visits fall, because the answer was delivered before anyone needed to click.Why it matters: a traffic decline used to mean something had gone wrong on your site. Now it can mean nothing went wrong at all, which changes what the fix is.

Zero-click

The older, narrower word for the same thing. A search where the person got what they needed on the results page and never visited a website.

Three acronyms that get used interchangeably and are worth telling apart.

SEO search engine optimization

Making sure your website shows up in a list of blue links when someone searches Google. The work is about ranking on a results page.Like getting your practice listed in the phone book, as high up the page as possible.

AEO answer engine optimization

Making sure you are the answer when someone asks a question and gets a paragraph back instead of a list of links. No ranking, no page of results. One answer, and either you are in it or you are not.Like being the name the person behind the desk says when a stranger asks who to see.

GEO generative engine optimization

The same idea as AEO, aimed at the systems that write their answers from scratch rather than picking from a list. In practice the two words describe the same work.Worth knowing: when someone uses all three acronyms in one sentence, they are usually describing one job.

AI Overview

The paragraph Google now puts above the search results, written by its AI. Most people read it and stop there.

Citation

Whether the AI names where it got something and links to it. Being mentioned is good. Being mentioned and linked is better, because the reader can click through to you.

Crawl and index

Crawling is a machine visiting your website and reading it. Indexing is that machine filing what it read so it can be used later. A page that was never crawled cannot appear anywhere else.

Schema also: structured data

Invisible labels in the code of your website that tell a machine what each thing on the page is. That this string of numbers is a phone number, that this person is a physician, that this is a service you offer.Like the nutrition label on a package. A person reads the front; a machine reads the label.

Entity

The difference between a word and a thing. These systems can know a practice name as a string of letters, or as a specific clinic in a specific city with named practitioners and a known specialty. The second is what you want, and it takes work.

Model

The thing itself. A very large piece of software that has been shown an enormous amount of writing and has learned the patterns in it. ChatGPT, Claude and Gemini are products; the model is the engine inside.

LLM large language model

The technical name for that kind of model. It works by predicting what words come next, very well and very fast. That is most of what is happening.

Training

The one-time process of showing a model an enormous body of text so it learns those patterns. It happened in the past, it is expensive, and it is not happening while you are talking to it.

Hallucination

When the model states something false with complete confidence. It is not lying. It has a gap and fills it with whatever fits the pattern, because producing a plausible answer is what it was built to do.Why it matters to a practice: a confident wrong answer about your credentials or your services does more harm than not appearing at all.

Prompt

What you type in. The question or instruction.

Context window

How much the system can hold in mind at one time. Large, but not unlimited. Past the edge, earlier parts of a long conversation stop being available to it.Like a desk. A big one, but a desk. What falls off the edge is gone.

Token

How these systems count text. Roughly three quarters of a word. It matters because usage is billed by the token.

Inference

The moment the model produces an answer. "At inference time" simply means "while it is answering."

Agent

A model that has been given a job, a set of tools, and permission to take several steps on its own to finish it. An ordinary chat answers your question. An agent goes and does the task.Like the difference between asking a colleague a question and asking a colleague to go handle something. The second involves decisions made while you are not watching.

Tools

The specific things an agent is allowed to use. Search the web. Read a file. Send an email. The list is the boundary, which is where the safety comes from.

Guardrails

The rules set around an agent before it runs: what it may touch, what stays out of reach, and when it stops and asks a person.

Human in the loop

A design where everything headed for a patient, a chart or the outside world is read by a person first. For anything clinical this is the design itself.

Automation

A fixed sequence that runs the same way every time. Worth separating from an agent, which decides its own steps. Much of what a practice needs is automation, and it is cheaper and safer.

Three ways this is done, each suited to a different job.

Context

Pasting the relevant material into the conversation each time. Simple, immediate, and it disappears when the conversation ends.

RAG retrieval augmented generation

Giving the system a filing cabinet it consults at the moment of answering. It looks up your actual documents, answers from them, and can show you which page it used. Your material stays in the cabinet rather than being absorbed into the model, and you can take it back.This is usually the right answer for a practice, because it is auditable and reversible.

Fine-tuning

Adjusting the model itself by showing it many examples, so it adopts a way of reasoning or a house style. It changes how the system responds, not what facts it can look up. Expensive, slower to correct, and harder to explain to a regulator.

On-premises also: self-hosted

The system runs on hardware you control rather than on someone else's servers. More work to set up. The reason to do it is that your records stay in the building.

Plain descriptions only. What a particular case requires is a question for the practice's counsel.

PHI protected health information

Anything in a record that identifies a patient or connects a patient to their care. Its handling is governed by law.

De-identified

Records with the identifying details removed, so the clinical content remains and the person cannot be recognized. Doing it properly is a discipline, not a find-and-replace.

BAA business associate agreement

The written agreement that has to be in place before anyone outside the practice handles patient information. It comes before the work, not alongside it.

Audit trail

A record of who accessed what, when, and what the system did. It is what answers the questions asked after something goes wrong.

EMR or EHR

The system holding your patient records. The two acronyms are used interchangeably in conversation.

What can be improved is the accuracy and structure of the information these systems draw on. Whether a system recommends, cites or names a particular practice rests with the system, so that is the part no one can promise.