In short
Plain-English definitions of the AI and AI-search terms you will actually meet, each in one sentence, with a longer explanation and an example. No maths, no jargon used to define other jargon. Written for people who work in marketing rather than for engineers.
Key points
A very broad umbrella. A spam filter is AI. So is a self-driving car. When most people say "AI" today they mean one narrow corner of it: generative AI, and usually a chat assistant.
For example: Netflix recommendations, face unlock and ChatGPT are all AI, and they work in completely different ways.
The distinction is creation versus categorisation. An older system looked at an email and decided "spam or not spam". A generative system writes the email.
For example: ChatGPT writing a reply is generative. Your bank flagging a transaction as fraud is not.
This is the engine behind ChatGPT, Claude and Gemini. It is not searching a database of answers. It is generating one, piece by piece, from patterns it learned during training. That single fact explains why it can be fluent and wrong at the same time.
For example: Ask for a book recommendation and it produces a plausible-sounding title, which may or may not be a real book.
Most people think a prompt is the question. It is really the whole briefing. The quality of what comes back tracks the quality of the briefing far more than it tracks the tool.
For example: "Write a headline" is a prompt. So is a page of background, three sample headlines and a word limit - and it produces a far better result.
Less mystical than the name suggests. It is mostly specificity: say who the model should be, what the situation is, what good looks like, and what to avoid. The grander-sounding version of the same skill is context engineering, which means supplying the background rather than just the instruction.
For example: "Write a headline" versus "You are a B2B copywriter. Audience: finance directors. Write six headlines under nine words, no puns."
Models do not see letters or words. Text is chopped into tokens, and both usage limits and pricing are measured in them. It is the reason a model can miscount the letters in a word.
For example: "unbelievable" might be three tokens. A 750-word article is roughly 1,000 tokens.
Think of it as desk space rather than memory. When a long conversation exceeds it, the earliest part falls off the desk, which is why a model can forget what you said an hour ago while remembering the last message perfectly.
For example: Paste a 200-page report into a small context window and it will silently work from only part of it.
Not a bug in the usual sense, and not lying. The system predicts likely text; likely text is usually true, and sometimes is not. Confidence in the wording tells you nothing about accuracy, which is the trap.
For example: Invented statistics, non-existent citations and confidently wrong dates are the classic three.
A model without web access is frozen at its cutoff. Many now search the web to fill the gap, which is a different mechanism with different failure modes.
For example: Ask about a product launched last month and an offline model will either say it does not know, or invent something.
Distinct from the context window, which only covers the current conversation. Memory is persistent and worth reviewing occasionally, because a wrong fact it stored quietly shapes every later answer.
For example: It remembers you work in SEO and stops explaining what a SERP is.
Useful when a person would need to think rather than recall: trade-offs, multi-step calculations, ambiguous decisions. Wasteful for a quick rewrite.
For example: Comparing two job offers across salary, equity and location is a reasoning task. Fixing a typo is not.
A grounded answer can be checked, because it points at where it came from. This is what separates an AI search tool from a chat assistant working from memory.
For example: Perplexity citing four sources is grounded. A model recalling a fact from training is not.
They answer the query on the results page, so fewer people click through. For publishers this is the central problem of the last two years: the same visibility produces less traffic.
For example: Search a how-to question and the boxed answer at the top, before the blue links, is an AI Overview.
Overlapping with SEO but not identical. Ranking wins a position; being cited wins a mention inside the answer. The work is clear structure, verifiable claims and something worth quoting.
For example: Rewriting a page so its answer sits in the first sixty words is a GEO change.
Read the full page →The industry has not settled on one label. GEO, AEO and LLMO describe substantially the same work. Do not spend energy on the distinction; anyone selling it as three services is selling you one.
For example: Structuring a page around real questions people ask, with each answered directly beneath its heading.
The AI-era equivalent of a ranking, and it behaves differently: citations are not stable, the same question can produce different sources each time, and being cited does not guarantee a click.
For example: Asked about a topic, an assistant answers and lists five sources. Being one of them is an AI citation.
GPTBot, ClaudeBot and PerplexityBot are the ones you will see in your logs. They are separate from Googlebot with separate rules, so allowing Google does not allow them.
For example: Blocking GPTBot in robots.txt does nothing to Googlebot, and vice versa.
Read the full page →Cheap to write and harmless. No major AI company has confirmed reading it, so treat it as documentation rather than a visibility tactic.
For example: yoursite.com/llms.txt, listing twenty pages with a sentence each.
Read the full page →It predates AI - weather, definitions and sports scores were always zero-click - but AI answers extend it to far more queries, including ones that used to send real traffic.
For example: Asking a factual question, reading the answer at the top, and closing the tab.
The chat window is the human door into a model; the API is the door for other software. It is how an AI feature gets built into an app, and it is billed per token rather than by monthly subscription.
For example: A helpdesk that drafts replies automatically is calling a model through an API.
The standard way to make an assistant that knows your business. The model supplies the language; your documents supply the facts. It reduces invention without eliminating it.
For example: An internal assistant that answers HR questions by retrieving the actual policy document first.
Before it, every tool needed its own bespoke integration. MCP is the shared plug: build a connector once and any assistant that speaks the protocol can use it.
For example: Connecting an assistant to your analytics, calendar and files without three separate custom builds.
The difference is doing versus answering. An assistant tells you how to do something; an agent attempts it, checks the result and tries again. The word is heavily oversold, and most things marketed as agents are automations with a model inside.
For example: "Research these ten companies and put the results in a spreadsheet", carried out end to end.
The practical safeguard against confident errors. The AI drafts, a person decides. It costs a little speed and prevents most of the failures that end up in the news.
For example: An AI writes the customer reply and saves it as a draft; a person reads it and presses send.
Available in most tools and usually off by default on free tiers. Worth finding once and setting deliberately, particularly if client work goes anywhere near the chat box.
For example: Turning off 'improve the model for everyone' in your account's data controls.