AI Glossary

Plain-English definitions for AI marketing, GEO, and agentic commerce terminology.

Agent (AI Agent)

An AI system capable of autonomous action. Unlike a chatbot that just talks, an agent uses tools (web search, code execution, APIs) to perform multi-step tasks to achieve a goal.

Agent Orchestration

The management, coordination, and logic layer (e.g., using frameworks like CrewAI, AutoGen, or LangGraph) that directs how LLMs, tools, memory, and prompts interact within an agentic workflow.

Agent Payments Protocol (AP2)

A secure transaction standard designed to facilitate tokenized financial approvals, allowing AI agents to complete checkouts autonomously within pre-authorized budget limits and authorization constraints.

Agentic Browsing Score

A metric introduced in Google Lighthouse 13.3 (PageSpeed Insights) that evaluates a website's usability for AI agents based on accessibility trees, DOM structure, and presence of AI-specific resources.

Agentic Checkout

The end-to-end checkout flow tailored specifically for machine-to-machine transactions, allowing AI agents to autonomously manage cart state, verify dynamic pricing, and initiate tokenized payments.

Agentic Commerce

E-commerce transactions where autonomous AI agents act as buyers—searching catalogs, selecting items, managing carts, and executing checkouts on behalf of human users.

Agent-Readiness Optimization (ARO)

The process of optimizing a website's semantic structure, APIs, and manifests (e.g., llms.txt, webmcp.json) to make it fully navigable and transactional for autonomous AI agents.

Agentic Swarm (Multi-Agent System)

A cooperative framework where multiple specialized AI agents work together, communicating and passing tasks to each other to solve complex, multi-stage challenges.

Agentic Workflow

A system where AI agents operate autonomously to complete multi-step tasks. Unlike a standard chatbot that answers one question, an agentic workflow plans, executes, critiques, and iterates on its own work to achieve a complex goal (e.g., "Write and deploy a full website").

AGI (Artificial General Intelligence)

Theoretical AI that possesses the ability to understand, learn, and apply knowledge across a wide variety of tasks, matching or exceeding human cognitive abilities in every domain. The "Holy Grail" of AI research.

AI Overview (AIO)

Google's search feature that displays AI-generated summaries at the top of search results, synthesizing information from multiple websites to provide immediate answers to user queries.

Alignment

The field of ensuring that AI systems goals and behaviors match human values and intentions, preventing unintended harmful consequences.

Answer Engine

A conversational search platform or application (such as Perplexity, Gemini, or ChatGPT) that provides synthesized direct answers to user queries rather than a list of hyperlink search results.

Answer Engine Optimization (AEO)

The practice of structuring and optimizing website content specifically to be consumed, compiled, and presented as the direct answer in AI-powered search engines.

API (Application Programming Interface)

The method by which different software programs talk to each other. In AI, this is how developers access models like GPT-4 or Gemini without hosting them locally.

Attention Mechanism

The mathematical breakthrough behind Transformers. It allows a model to "pay attention" to different parts of a sentence simultaneously to understand context, rather than reading sequentially word-by-word.

Auto-GPT

An early open-source experiment in autonomous AI agents that attempts to achieve a user-defined goal by breaking it into sub-tasks and executing them in a loop.

Bard (now Gemini)

Google's conversational AI chatbot, now rebranded as Gemini, designed to compete with ChatGPT using Google's own LaMDA and PaLM models.

Bias

Prejudice in favor of or against one thing, person, or group compared with another, usually in a way considered to be unfair, which AI models can inherit from their training data.

Black Box

A system whose inputs and outputs are observable, but whose internal workings are opaque. Most deep learning models are black boxes because we cannot easily explain *why* they made a specific decision.

Chain-of-Thought (CoT)

A prompting technique that encourages the AI to explain its step-by-step reasoning before arriving at a final answer. This significantly improves accuracy on complex logic and math problems.

Chatbot

A software application designed to simulate conversation with human users, especially over the internet.

Claude

A family of Large Language Models developed by Anthropic, known for their large context windows and focus on safety (Constitutional AI).

Computer Use

The capability of AI agents to interact with graphical user interfaces (GUIs) just like a human, executing commands, clicking elements, and typing inside applications by viewing the screen.

Computer Vision

A field of AI that enables computers to derive meaningful information from digital images, videos and other visual inputs.

Context Window

The "short-term memory" of an AI. It dictates how much data (documents, code, chat history) the model can process at once. Current frontier models compete heavily on context length and retrieval quality.

Copilot

An AI assistant that works alongside a human user, typically in a specific application like coding (GitHub Copilot) or office work (Microsoft 365 Copilot).

DALL-E

OpenAI's AI system that can create realistic images and art from a description in natural language.

Deep Learning

A subset of machine learning based on artificial neural networks with representation learning. It allows models to learn from vast amounts of data.

Diffusion Model

A class of generative models primarily used to create images. They work by adding noise (static) to an image until it is unrecognizable, and then learning to reverse the process to reconstruct a clear image from pure noise.

Embeddings

Numerical representations of text, images, or audio as vectors (lists of numbers). Embeddings capture semantic meaning, allowing computers to understand that "King" is close to "Queen" but far from "Apple" in mathematical space.

Emergent Behavior

Capabilities that an AI model exhibits which were not explicitly programmed or anticipated by its creators, often appearing only when models reach a certain scale.

Factual Density

The ratio of unique, high-value, and verifiable facts to overall word count in a piece of content, serving as a primary ranking signal for Generative Engine Optimization (GEO).

Few-Shot Learning

A machine learning approach where the model is given a small number of examples (shots) to learn how to perform a task, rather than thousands.

Fine-tuning

The process of taking a pre-trained base model (like GPT-4) and training it further on a smaller, specific dataset to specialize it for a particular task, such as medical diagnosis or coding.

Foundation Model

A large-scale AI model trained on a vast amount of data (often unlabeled) that can be adapted (e.g., fine-tuned) to a wide range of downstream tasks.

Function Calling (Tool Use)

An LLM capability where the model outputs structured arguments (typically JSON) to invoke external tools, APIs, or custom code, bridging the gap between text and action.

GAN (Generative Adversarial Network)

A machine learning model architecture involved two neural networks contesting with each other in a game. One creates content, the other judges it. Largely replaced by Diffusion models for image generation.

Gemini

Google's most capable and general model, built to be multimodal from the ground up. It comes in Nano, Flash, Pro, and Ultra sizes.

Generative AI

Artificial intelligence capable of generating new content (text, images, audio, video) in response to prompts, rather than just analyzing existing data.

Generative Engine Optimization (GEO)

The process of optimizing website content's structure, factual density, and semantic markup to increase the likelihood of being cited and recommended by AI answer engines like Perplexity, Gemini, and ChatGPT.

GPU (Graphics Processing Unit)

Electronic circuit designed to manipulate and alter memory to accelerate the creation of images in a frame buffer. Crucial for training and running AI models due to their parallel processing power.

GPT (Generative Pre-trained Transformer)

A series of large language models developed by OpenAI. They are the engine behind ChatGPT.

GraphRAG

An advanced Retrieval-Augmented Generation technique that uses a structured knowledge graph to link related concepts, enabling LLMs to understand complex semantic relationships across large datasets.

Grounding

The process of connecting LLM outputs to verifiable sources of real-world information (like Google Search results) to reduce hallucinations.

Hallucination

When an AI confidently generates false or nonsensical information. This happens because LLMs predict the next likely word based on patterns, not on a verified database of facts.

Hugging Face

A platform and community that hosts open-source machine learning models, datasets, and demos, often described as the "GitHub of AI".

Human-in-the-Loop (HITL)

A system design pattern that requires human approval, intervention, or feedback at critical checkpoints before an autonomous AI agent executes high-stakes or irreversible actions.

Inference

The process of running a trained machine learning model to make a prediction or generate content. It is the "using" phase, distinct from the "training" phase.

LaMDA

Language Model for Dialogue Applications, a family of conversational neural language models developed by Google.

Latent Space

An abstract multi-dimensional space where an AI organizes data. Similar concepts are grouped closer together.

Llama (Large Language Model Meta AI)

Meta's family of highly capable, open-weight large language models that are widely used by developers to build and host custom AI applications locally or in private clouds.

LLM (Large Language Model)

A deep learning algorithm that can recognize, summarize, translate, predict, and generate text and other content based on knowledge gained from massive datasets.

llms.txt

A proposed standard text file located at a website's root directory (/llms.txt) that provides a concise, machine-readable summary, navigation guidelines, and rules for LLM crawlers and AI agents.

LoRA (Low-Rank Adaptation)

A technique for fine-tuning large models efficiently by freezing the pre-trained model weights and injecting validatable rank decomposition matrices into each layer of the Transformer architecture.

Machine Learning (ML)

A subset of AI that focuses on building systems that learn—or improve performance—based on the data they consume.

Midjourney

An independent research lab and the name of their generative artificial intelligence program that creates images from natural language descriptions.

Mistral

A French AI company and their open-weights models (like Mistral 7B, Mixtral 8x7B) known for high performance and efficiency.

Mixture of Experts (MoE)

A model architecture that uses multiple specialized sub-networks ("experts") to answer a query, activating only the relevant ones. This makes massive models more efficient (used in GPT-4 and Mixtral).

Model Collapse

A theoretical degradation that occurs when AI models are trained on AI-generated data. Without fresh human data, models may lose nuance, creativity, and accuracy over successive generations.

Model Context Protocol (MCP)

An open-source communication protocol developed by Anthropic that standardizes how applications securely provide context, data, and tool actions to large language models (LLMs).

Multimodal

A model capable of processing and generating multiple types of media simultaneously, such as text, images, audio, and video (e.g., Google Gemini, GPT-4o).

Natural Language Processing (NLP)

A branch of AI that helps computers understand, interpret and manipulate human language.

Neural Network

A method in artificial intelligence that teaches computers to process data in a way that is inspired by the human brain.

Open Source AI

AI models where the weights and architecture are made publicly available (e.g., Llama 2, Mistral), allowing anyone to run or modify them.

Overfitting

A modeling error that occurs when a function is too closely fit to a limited set of data points, performing well on training data but poorly on new data.

Parameter

The internal variables (weights) that the model learns during training. Generally, more parameters mean a more capable (and expensive) model.

Prompt Engineering

The art and science of crafting inputs (prompts) to guide Generative AI models to produce the best possible outputs. It involves key context, constraints, and formatting instructions.

Prompt Injection

A security exploit where a malicious user crafts a prompt to trick the AI into ignoring its safety controls or revealing sensitive information.

Quantization

The process of reducing the precision of a model's parameters (e.g., from 16-bit to 4-bit) to make it run faster and use less memory with minimal loss in quality.

RAG (Retrieval-Augmented Generation)

A technique that enhances the accuracy of an AI model by fetching relevant facts from an external knowledge base (like your company docs) before generating an answer. This reduces hallucinations and allows the AI to use private data.

Reasoning Model (o1)

A new class of AI models (like OpenAI o1) designed to "think" before they generate. They spend inference time verifying logic and planning steps, resulting in vastly higher performance on math, coding, and complex logic tasks.

Reinforcement Learning (RL)

An area of machine learning concerned with how software agents ought to take actions in an environment in order to maximize the notion of cumulative reward.

RLHF (Reinforcement Learning from Human Feedback)

A training method where human testers rank different AI responses to steer the model toward safer, more helpful, and more aligned behavior.

Sam Altman

CEO of OpenAI, the company behind ChatGPT and DALL-E.

Search Generative Experience (SGE)

Google's initial experimental search initiative integrating conversational AI directly into search results, which served as the foundation for modern AI Overviews.

Semantic Cache

A system that stores previously generated AI responses based on the semantic meaning of prompts rather than exact keyword matches, significantly reducing costs and latency.

Sentiment Analysis

The use of natural language processing to identify, extract, quantify, and study affective states and subjective information.

Singularity

A hypothetical future point in time at which technological growth becomes uncontrollable and irreversible, resulting in unfathomable changes to human civilization, often associated with the arrival of superintelligence.

Slop

Slang for low-quality, mass-produced AI content (spammy text or uncanny images) that floods the internet. It is the "junk food" of the information age.

Stable Diffusion

A latent text-to-image diffusion model capable of generating photo-realistic images given any text input, developed by Stability AI.

Stripe Agentic Commerce Protocol (ACP)

A payment and transaction standard developed by Stripe that enables secure, programmatically-controlled API checkouts tailored for machine-native purchases by AI agents.

Superalignment

The technical challenge of steering and controlling superintelligent AI systems that are much smarter than humans.

Synthetic Data

Information that's artificially generated rather than created by real-world events. It's used to train models when real data is scarce, expensive, or sensitive.

Temperature

A parameter that controls the randomness of an AI's output. High temperature (e.g., 0.8) makes it more creative and unpredictable; low temperature (e.g., 0.2) makes it deterministic and focused.

Token

The basic unit of text for an LLM. A token can be a character, a word, or part of a word. Roughly 1,000 tokens equal 750 words.

Training Data

The large dataset used to teach an AI model. For LLMs, this is typically a significant portion of the public internet, books, and academic papers.

Transformer

The neural network architecture that makes modern Generative AI possible. Introduced by Google in 2017, it allows models to handle long-range dependencies in text using "self-attention" mechanisms.

Turing Test

A test of a machine's ability to exhibit intelligent behavior equivalent to, or indistinguishable from, that of a human. Proposed by Alan Turing in 1950.

Universal Commerce Protocol (UCP)

A standardized protocol co-developed by Google and Shopify that enables AI shopping assistants to build, synchronize, and checkout a single cart state across multiple different online retail stores.

Vector Database

A specialized database designed to store and search vector embeddings. It allows for ultra-fast semantic search (finding content based on meaning rather than keywords).

WebMCP

A web-focused extension of the Model Context Protocol (MCP) that allows websites to expose a static webmcp.json manifest, enabling AI agents to interact with search, carts, and forms using standard JSON schemas.

Zero-Shot Learning

The ability of a model to perform a task it has never explicitly seen before during training, simply by following instructions in the prompt.