So this is organized differently. Four groups: terms that describe how the technology actually works, terms you will hear from vendors, terms about whether AI systems can find your business, and terms that are mostly noise.
If you want the underlying explanation of what these systems are rather than what they are called, read the basics piece first. This one is a reference for when somebody uses a word at you.
Group one: how it actually works
These describe real things and are worth understanding.
Large language model, or LLM. The underlying system. Trained on enormous amounts of text, it predicts likely continuations. Everything else in this list sits on top of it.
Generative AI. Any system that produces new output rather than classifying existing things. Writing, images, audio, video. The broader category that LLMs belong to.
Prompt. What you type. Prompt engineering is the practice of writing better ones, which mostly means being specific about audience, format, and constraints.
Context window. How much text the model can hold in view at once, including your conversation so far and any documents you paste. Exceed it and the earliest material silently drops out. This is why long conversations start to feel like the model forgot something. It did.
Token. The unit text is broken into, roughly three quarters of a word. Pricing is usually per token, which is why long documents cost more to process.
Hallucination. A confident, fluent, wrong answer. Not a malfunction. A direct consequence of a system that predicts plausible text rather than retrieving verified facts.
Training data. What the model learned from. You cannot see it, and neither can the vendor in any specific sense.
Knowledge cutoff. The date training stopped. Ask about anything after it and, without live search, the model will either say so or invent something.
Multimodal. Handles more than one kind of input or output. Reads an image and writes about it, or turns text into audio. Most leading models now do this.
Group two: what vendors will say
These are real techniques. They are also where the sales pressure lives.
Fine tuning. Additional training on your specific material. Genuinely useful at scale, expensive, and almost never what a small business needs. If a vendor proposes it, ask what problem it solves that a well written prompt does not.
Retrieval augmented generation, or RAG. Connecting a model to a specific body of documents so it answers from those rather than from training alone. This is what powers "chat with your documents" features. It is a legitimate technique and it is what most people actually want when they think they want fine tuning.
Agent, or agentic AI. A system that takes a sequence of actions rather than answering one question. It might search the web, run code, read files, and call other software to complete a task. Real and increasingly capable. Also the word most likely to be attached to something that is a normal chatbot.
API. The connection point that lets one piece of software talk to another. Relevant to you only in that it determines whether a tool can be connected to your other tools.
Embeddings and vector search. The mechanism underneath RAG, which finds material by meaning rather than exact keyword match. You will never configure this. It is useful to recognize the word so you know somebody is describing plumbing rather than a product.
Whatever the acronym, ask: what specific task does this complete, and what happens when it gets that task wrong? A vendor with a real product answers both immediately. A vendor selling vocabulary will redirect to capability in the abstract. The second question matters more than the first, because a system that acts on your behalf can be wrong in ways a chatbot cannot.
Group three: whether AI can find you
This is the group that actually affects revenue for a local business.
AI Overview. The AI generated answer Google now shows above the traditional results on many searches. They appear on roughly a quarter of searches and substantially reduce clicks to the listings below.
GEO, AEO, LLMO. Generative engine optimization, answer engine optimization, large language model optimization. Here is the honest position: these are three names for the same goal, which is getting cited when an AI system answers a question your customer asked. The industry has not settled on one word and probably will not. Anybody presenting one of them as a distinct discipline requiring a separate retainer is selling you a naming convention.
Citation. Being named as a source inside an AI answer. This is the thing that matters. Research suggests branded mentions across the web correlate with appearing in AI Overviews far more strongly than backlink counts do, which is a meaningful shift from traditional search.
Schema markup. Machine readable labels in your website code that state plainly what each piece of content is. A review, a business address, an FAQ answer, a service. Old technology, newly important, because it removes ambiguity for a system deciding whether to cite you.
Structured data. The broader category schema belongs to. Information formatted so a machine can parse it without guessing.
llms.txt. A file at the root of a website summarizing what it contains, intended for AI systems. Adoption is not universal and its value is not yet proven. Cheap to add, so worth doing, but not urgent.
Group four: mostly noise
Terms you will encounter that carry less than they appear to.
- AI powered. Applied to everything, including software using a rule that was written in 1998. It tells you nothing. Ask what the feature does
- Machine learning. A real field, and technically the parent of everything here, but in marketing copy it is usually a synonym for "we do math"
- AGI. Artificial general intelligence, meaning human level ability across all domains. It does not exist, definitions vary wildly, and it has no bearing on any decision you will make this year
- Neural network, deep learning, transformer. Accurate descriptions of the underlying architecture. You do not need them. If a vendor leads with them, they are demonstrating expertise rather than explaining a benefit
- Proprietary acronyms. If you cannot find a term used by anybody other than the company saying it, that is what it is
The three that actually matter
If you retain nothing else from this.
- 01Hallucination, because it means you verify anything with consequences before it leaves your hands
- 02Agent versus chatbot, because one answers and the other acts, and buying the wrong one wastes money in different ways
- 03Citation, because AI systems now sit between your customers and your business, and whether you get named is a thing you can influence
Everything else you can look up when somebody says it.
How to use this
The vocabulary will keep changing. Some terms here will be obsolete within a year and others will appear that nobody has coined yet.
What will not change is the underlying test. When somebody uses a word you do not recognize, ask them to explain what it does in a sentence with no acronyms in it. Anybody who understands their own product can do this immediately.
The inability to do it is the most reliable signal available, and it has never once failed to be informative.