The NSY AI & web glossary

By Cédric Barme · Founder of NSY

The terms that come up in every web + AI project, explained without jargon by someone who implements them — not just talks about them. Each definition fits in a few sentences; for concrete cases, see the FAQ or the AI website offering.

Artificial intelligence

LLM (Large Language Model)

A large language model: an AI trained on vast text corpora, able to understand and produce natural language. Claude (Anthropic), GPT (OpenAI) and Mistral are the most widely used in business.

Token

The unit a LLM splits text into — roughly ¾ of an English word. API billing and context-window size are measured in tokens: mastering this unit means mastering your costs.

Context window

The amount of text a LLM can "keep in mind" within one exchange. It determines the size of documents that can be analysed in one pass — a major selection criterion between models.

Prompt

The instruction given to the model. Prompt quality (context, constraints, examples, expected format) often does more for the result than the choice of model itself.

Hallucination

A wrong answer delivered with confidence. Unavoidable in free generation, it is controlled through RAG (answers grounded in your documents), guardrails and human review on high-stakes uses.

Embeddings

Representing a text as a numeric vector that captures its meaning. Two semantically close texts have close vectors — the building block of semantic search.

Semantic search

Searching by meaning rather than exact keywords: "parental leave" also retrieves documents about "birth" or "long-term absence". Built on embeddings.

RAG (Retrieval-Augmented Generation)

The architecture pattern that makes an assistant reliable: relevant documents are retrieved first (semantic search), then the LLM answers from them, with sources — instead of inventing.

AI agent

A LLM that does more than answer: it acts through tools (querying a database, filling a form, calling an API), across multiple steps, under guardrails. See Claude integration.

Fine-tuning

Partially retraining a model on your data to specialise its behaviour. Often unnecessary: a good prompt plus RAG covers most needs, at a fraction of the cost.

MCP (Model Context Protocol)

An open standard (initiated by Anthropic) describing how a LLM connects to external tools and data sources — a universal port between models and enterprise systems.

No-LLM chatbot

A rule-based intent engine: free to run, instant, no API key and no data leaving the site. Right when the question domain is bounded — this site's AI assistant keeps it as an offline fallback, for when the LLM (Mistral) is unavailable.

Web & search

SEO (Search Engine Optimization)

Optimising for classic search engines: content, technical health (speed, mobile), internal linking and authority. The foundation — without it, nothing else works.

Semantic SEO

Structuring meaning beyond keywords: named entities, structured data, consistency of facts across pages. It is what lets engines — and AIs — understand who you are.

GEO (Generative Engine Optimization)

Getting cited by generative engines (ChatGPT, Claude, Gemini, Perplexity). The levers: attributable, dated content, cross-checkable entities (Wikidata, official registries), llms.txt files, conversational FAQs.

LLMO (LLM Optimization)

GEO's cousin: optimising how LLMs understand your site during training or browsing — allowing their crawlers, providing structured context, avoiding content reachable only through JavaScript.

llms.txt

A file at a site's root (llmstxt.org convention) summarising for AIs who you are, what you do and where to find the essentials. NSY's: nsy.fr/llms.txt.

JSON-LD / structured data

Invisible markup (schema.org) describing your pages in machine language: organisation, people, services, prices, FAQ. The raw material of Google's Knowledge Graph and of assistant answers.

hreflang

Tags declaring the language versions of a page (fr/en…) so every visitor — and every engine — gets the right language, without being counted as duplicate content.

Core Web Vitals

Google's metrics measuring real experience: display speed (LCP), responsiveness (INP), visual stability (CLS). They weigh on rankings — and above all on your visitors' patience.

AI crawler

An AI player's exploration robot (GPTBot, ClaudeBot, PerplexityBot…). Allowing them in robots.txt is the entry condition for existing in assistant answers — blocking them writes you out of their corpora.

Edge hosting

Serving your site from servers distributed close to visitors (Cloudflare, Vercel): minimal latency everywhere, higher resilience — relevant as soon as your audience spans more than one region.

Single sign-on (SSO)

One account for several applications: one holds identity, the others consult it. At PRV Concept, the forum account opens the shop through a single-use encrypted token — see One site, one forum, one shop, one account.

Child theme / child style

A layer that inherits from a piece of software (phpBB, WordPress) and replaces only what it is asked to, usually the look. The software updates as normal and the look is maintained separately: that is what makes it possible to modernise a forum or a shop without migrating them.

Community knowledge base

A forum where every question has become a topic filed in a section, with a title and a stable address: the search, Google and AI assistants find the answer years later, where a social-network post sinks within three days. See The same question, every three months.

Forum ↔ social network bridge

An automation that prepares, for each new forum topic, a post for the Facebook page or Instagram, and copies the comments left on the networks back into the original topic. The networks remain where people talk, the forum where people find. At PRV Concept, nothing goes out without human approval.

Monitoring

Knowing that a production system is well — or unwell — before its users notice. Four questions, at any scale: what is abnormal, who gets told, what do we keep, what do we do with an alert that lied. A false alert is qualified, not switched off. See From terabytes to megabytes.

Observability

The ability to understand a system's internal state from what it emits: metrics, logs and traces. Monitoring says there is a problem; observability lets you understand which one. Typical tools at large accounts: Dynatrace (execution; Davis, its causal AI, detects without thresholds and walks up to the root cause) and ELK (logs, business KPIs).

Video & 3D

Motion design

The animation of shapes, text and logos in the service of a message: channel intro, ident, lower thirds. What sets it apart from a "generated" video: every frame is controlled, and the logo stays exact from the first shot to the last. See the motion design for YouTube offer.

Storyboard

The plan of the video, shot by shot, with the exact on-screen text and the timing. Written like a score, in bars, because the music sets the rhythm: every cut lands on a beat, every impact on an accent.

Animatic

A low-resolution version of the animation, without final materials or lighting, used to judge the rhythm before rendering. Fixing things at this stage costs minutes; after the render, hours.

3D render

The computation of the final image from a three-dimensional scene — geometry, materials, light. Blender's Cycles engine simulates light physically; a script-driven render replays identically when a colour or a text changes.

Extrusion and chamfer

Giving volume to a flat shape — a logo — by pulling it into depth (extrusion), then softening its edges (chamfer) so they catch the light. The basis of a logo in volume, see the 3D animation offer.

16:9 and 9:16

The two picture formats: landscape for YouTube, portrait for Shorts, Reels and stories. A 9:16 video is recomposed — the elements are repositioned in the vertical frame —, never cropped: cutting the sides of a 16:9 loses the subject.

Loudness

The measure of a video's perceived volume, in LUFS. Each platform brings the sound to its own target: delivering at the right level avoids a crushed or too-quiet result once online.

Interactive 3D model

A three-dimensional object visitors rotate inside the page, with no extension to install (model-viewer, compressed GLB file). Optimised, it weighs a few hundred kilobytes and does not slow the site down.