Glossary
Short explanations of terms the news takes for granted. Written for someone new to the subject.
A
AI Overviews
The generated summary Google places above the results, written from pages it cites — answering the question on the results page instead of sending the reader on.
AI content labeling
Marking content that a model produced, so a reader and a machine can both tell — required in the EU for synthetic media under Article 50 of the AI Act.
Agent
A program that takes a goal, works out the steps itself, and uses tools to carry them out, instead of answering one question and stopping.
Agent Skills
Folders of instructions, scripts, and files that an agent loads only when the task calls for them, so the starting prompt stays small.
B
C
Canonical URL
The address a page declares as its real one, so the same content reachable through several links is counted as a single page.
Compaction
Replacing an older part of a conversation with a summary so the work can go on after the context window fills up.
Container queries
CSS that reacts to the width of the box a component sits in, rather than to the width of the window.
Context window
How much text a model can hold in view at once, counted in tokens: the conversation, the files, the tool results, and the answer being written.
Core Web Vitals
Google's three measures of how a page feels to real visitors: LCP for loading, INP for responsiveness, CLS for stability.
Cumulative Layout Shift (CLS)
How much the page moves under the reader while it loads. Good is 0.1 or less; the usual cause is content that arrives without a reserved place.
D
Diffusion model
The kind of model behind most image and video generators: it starts from noise and removes it step by step until a picture the prompt describes is left.
Distillation
Training a smaller model on the outputs of a larger one, to get most of the quality at a fraction of the cost and latency.
E
Embedding
A list of numbers standing in for a piece of text, arranged so that two texts about the same thing land close together and can be found by meaning instead of by matching words.
Evals
Your own test set for a model-backed feature: fixed inputs, expected outcomes, and a score you can watch across versions.
F
Fine-tuning
Continuing a model's training on your own examples, so it takes on a form or a task without being told about it in every prompt.
H
Hallucination
An answer that is fluent, confident, and wrong — a fact, a citation, or an API that does not exist.
Hook
A command the tool runs by itself at a fixed moment — before a file is written, after a task ends — so a rule holds without anyone remembering it.
Human in the loop
A design where the agent stops and asks before any step that cannot be taken back — sending, publishing, paying, deleting.
Hydration
The step where JavaScript takes over HTML that arrived already rendered, attaching the behaviour the markup alone does not have.
hreflang
A declaration that two pages are the same content in different languages, so a search engine shows each reader the right one.
I
IndexNow
A ping that tells search engines a URL has changed, instead of waiting for the next crawl. One request, one shared key file.
Interaction to Next Paint (INP)
How long the page takes to show a response after a tap or a click, across the whole visit. Good is 200 milliseconds or less.
L
Largest Contentful Paint (LCP)
The moment the largest thing in the viewport finishes drawing — the closest single number to "the page looks loaded". Good is 2.5 seconds or less.
llms.txt
A proposed file at the root of a site that hands a model a clean index of the pages worth reading, in Markdown.
M
Mixture of experts (MoE)
An architecture that holds many parameters but uses only a small share of them for each token, so a large model can answer at the cost of a much smaller one.
Model Context Protocol (MCP)
An open standard for connecting an agent to tools and data: one protocol instead of a separate integration for every service.
Multimodal model
A model that takes more than text as input — images, audio, video, whole documents — and reasons over them in the same conversation.
O
Open-weights model
A model whose trained parameters are published, so anyone can run it on their own hardware — which is not the same as open source.
P
Progressive enhancement
Building so the page works with HTML alone, then adding CSS and JavaScript as improvements rather than as requirements.
Prompt caching
Reusing the already-processed beginning of a prompt across calls, so the unchanged part is billed at a fraction of the usual input price.
Prompt injection
An attack that hides instructions inside content a model reads — a page, a file, a message — so the model follows the attacker instead of the user.
Q
Quantization
Storing a model's weights at lower precision so it fits in less memory — the usual way a model that needs a data-centre card comes to run on a laptop.
R
Reasoning effort
A dial that decides how much thinking a model spends before it answers, traded directly against time and money.
Retrieval-augmented generation (RAG)
Answering from documents handed to the model at question time, found by searching your own data, instead of from what it memorised during training.
robots.txt
A file at the root of a site telling automated visitors which paths they may fetch — and, increasingly, which AI crawlers are allowed at all.
S
Sandbox
A restricted space an agent runs inside, where a command can only reach the files and the network it was explicitly given.
Semantic HTML
Using the element that says what a thing is — button, nav, label, table — instead of a div dressed up to look like it.
Server rendering
Producing a page's HTML before it reaches the browser — at build time, per request, or on a schedule — instead of leaving the browser to assemble it.
Sitemap
A machine-readable list of the addresses a site wants found, with the date each was last changed.
Structured data
Machine-readable facts about a page, written as JSON-LD in the markup: what it is, who wrote it, when, and under what policy.
Structured output
Making the model answer in a shape a program can read — JSON that matches a schema — instead of prose that has to be parsed and hoped over.
Subagent
A second agent an agent starts for one bounded job, with its own context, so the work does not fill the caller's.
System prompt
The instruction that sits above the conversation and applies to every turn in it: who the model is, what it may do, and what it must refuse.
T
Temperature
The setting that decides how much the model is allowed to stray from its most likely next word. Zero gives the steadiest answer; higher values give a different one each time.
Time to First Byte (TTFB)
How long the browser waits before the first byte of the answer arrives. Good is 800 milliseconds or less, and everything after it inherits the delay.
Time to first token (TTFT)
How long a model takes to start answering, as opposed to how fast it writes once it has started.
Token
The unit a model reads, writes, and bills in: in English roughly three quarters of a word, in Serbian noticeably less.
Tool use
The model asks for a named function with arguments, the program runs it and hands back the result, so the model can act instead of only describing.
V
Vector database
A store built to answer one question quickly: which of these million embeddings sit closest to this one.
View Transitions
A browser API that animates the change between two states of a page, or between two pages, without the site having to hold both in memory itself.
W
Web Components
The browser's own way to define a reusable element — custom tags, a shadow tree that keeps styles from leaking, and a template to stamp out.
WebAssembly
A compact instruction format browsers run at near-native speed, so code written in other languages can work on a page without being rewritten in JavaScript.