The AI Terms Glossary: 60+ Concepts in Plain Language

The AI terms glossary: LLM, token, RAG, hallucination and 60+ concepts explained in plain language. A practical reference for technical conversations.

Anar RustamliAugust 1, 202613 min read2,419 words

When the terms multiply, the understanding does not always grow. Sometimes a team uses the same word in three different senses and sees it only when a decision goes wrong. This glossary gathers into one place, in plain language, the concepts you meet most often in artificial intelligence, automation, SEO, GEO, analytics and marketing. This glossary exists precisely to remove that confusion among the AI terms.

The definitions are kept short. The goal is not memorising the word; it is pinning down what you are talking about in a document, a meeting and a tool choice. A term can take a different shade per platform. So checking a changing feature or rule at the primary source before the final decision remains your job.

How to work with the glossary?

Find the word you seek with the browser's find-on-page function.
Compare the concept in the definition with your own real scenario.
Re-check a changing platform rule at the primary source.
Define a term once in the project document and then use it in the same sense.

The AI and digital terms

Artificial intelligence: A computer system performing tasks that usually demand human thinking: recognition, prediction, text creation and decision support.
Machine learning: A method where the program learns relationships from example data — without each rule being written separately — and gives results for new input.
Deep learning: The machine learning field that learns features from complex data like images, sound and text through multi-layer neural networks.
Neural network: A computational model that processes inputs through weights and recognises patterns by adjusting those weights during training.
Generative AI: An artificial intelligence system able to create new text, images, sound, video or code based on the patterns it learned.
LLM: A large language model trained on a vast body of text, modelling language sequences to understand and generate text.
Token: The unit — a word, word fragment, punctuation mark or symbol — the AI model uses when processing text.
Prompt: The input instruction telling a generative AI system the task, the context, the boundaries and the expected output format.
Context window: The limit of the total token volume the model can consider at once while creating one answer.
Embedding: A text, image or other object's representation converted into a vector of numbers for computing semantic similarity.
Vector: A mathematical representation expressing an object's chosen features as a sequence of numbers and enabling comparison.
Vector database: A database storing embeddings, optimised for quickly finding the objects semantically closest to a query.
RAG: The retrieval-augmented generation approach where, before the model answers, the relevant information is found in an external source and added to the prompt.
Fine-tuning: The re-training process of adapting a pre-trained model's behaviour to a specific task or style with additional examples.
Foundation model: A large base model trained on broad data and later adaptable to many different tasks.
Transformer: The architecture that processes the relationships in a text sequence via the attention mechanism, and that underlies the modern large language models.
Attention: The mechanism computing which parts of the input relate most to each other while the model creates the output.
Inference: The stage where a fully trained model accepts new input and creates a prediction, classification or generative answer.
Training: The learning process in which the model's parameters are computed and updated over example data to reduce the error.
Dataset: An orderly collected set of text, image, number and other samples for training, validation, testing or analysis.
Model: The mathematical system storing the learned relationships and computations for getting a result from input data.
Parameter: An internal numeric value the model learns during training, determining how it reacts to input.
Temperature: The parameter governing variety in the generative model's choice among probabilities; a high value usually creates a more variable answer.
Top-p: It regulates the model taking the next token only from the most likely options whose combined probability reaches the given threshold.
Hallucination: The case of an AI model creating information that sounds convincing but is absent from the source, wrong or invented.
Guardrail: The technical rule, filter and control layer applied to restrict an AI system's risky inputs and outputs.
Alignment: The direction of matching an AI system's behaviour with human goals, safety requirements and accepted values.
Bias: A tendency arising from the data, the selection or the model's build that creates systematic prejudice for some groups and outcomes.
Explainability: The degree to which a human can understand and assess on what grounds the model reached its result.
Multimodal AI: The AI approach that accepts and processes more than one data type — text, image, sound and video — in the same system.
Computer vision: The field studying computers recognising the objects, text, motion and visual relationships in images and video.
NLP: The natural language processing field concerned with computers processing, analysing and generating human language.
Speech-to-text: The speech recognition technology automatically converting spoken sound into written text.
Text-to-speech: The technology converting written text into synthetic human speech, used in voice interfaces.
OCR: The technology converting the printed or handwritten writing in images and scanned documents into machine-processable text.
API: The interface allowing two programs to exchange data and functions through pre-defined request and response rules.
Webhook: The method of one system sending another an automatic HTTP notification when a certain event happens.
SDK: The set of code libraries, tools, samples and documents easing app building for a certain platform.
Agent: A system observing its environment, choosing the next step toward a goal and able to execute certain actions.
AI agent: An AI system that uses a language model and tools to plan, gather information and carry out permitted operations in stages.
Workflow: The defined sequence of tasks, decision points, responsible people and data flow for creating one result.
Automation: Executing repeating work with a clear rule through software and integrations, with minimal manual intervention.
Trigger: The time, form submission, status change or other event condition starting an automated flow.
Action: The step the system performs after the trigger fires: sending a message, creating a record or updating data.
Integration: The setup of separate programs exchanging data and operations, working as parts of a single process.
CRM: The system managing customer and prospect data, the communication, the sales stages and the next steps.
ERP: The platform uniting core operations like finance, procurement, stock, production and human resources under one data system.
CDP: The platform merging first-party customer data from various channels into a single profile and readying it for activation.
Chatbot: A program accepting user questions via text or voice and answering by rules, a knowledge base or AI.
Copilot: An AI assistant giving suggestion, text and operation support inside the program a person uses, without replacing their work entirely.
Custom GPT: A ChatGPT-based assistant adapted to a certain use scenario with special instructions, files and tools.
Open-source model: A model whose weights or source code are provided under an open licence for use, inspection and adaptation.
Proprietary model: A model whose weights and internals are kept closed by its owner, used usually via a service or licence.
Cloud AI: The AI solution where the computation and model services run on remote cloud infrastructure and are used over the network.
Edge AI: The approach of executing the AI computation close to the data source — on the phone, the camera or a local device.
AGI: The still-debated AI concept presumed able to learn and perform many different intellectual tasks generally, like a human.
AEO: The optimisation approach for content to be selectable by search and answer systems as the direct, precise answer to a user's question.
GEO: The optimisation done so that a brand and content are correctly understood, used and cited in generative AI answers.
SEO: Improving a site's technical, content and trust signals so it gets found in search engines for the relevant queries.
SERP: The results page where the search engine shows the organic results, ads, maps and other blocks in reply to a query.
AI Overview: Google's feature showing, for some queries, a source-grounded generative summary above the search results.
Structured data: The markup announcing a page's objects and relationships to the search engine in a standard, machine-readable format.
Schema markup: The structured data code describing articles, products, people and other objects using the Schema.org vocabulary.
Entity: A person, organisation, place, product or concept object that a search and knowledge system can recognise distinctly.
Knowledge graph: The data model storing objects and the relationships between them as a graph, joining facts in context.
Citation: The reference information showing which source a claim or answer rests on.
Source attribution: Clearly identifying the information's author, publication and original source and tying the result to that source.
Zero-click search: The case of the user getting the answer on the results page and completing the search without visiting any website.
Search intent: The information, comparison, site-navigation or purchase goal the user wants to achieve by typing the query.
Keyword: The word or phrase people use in search that connects a content topic with query demand.
Topic cluster: The content structure covering one main topic through a pillar page and narrower articles tied to it with internal links.
Pillar page: The main, comprehensive page explaining a broad topic systematically and linking out to the sub-topics.
Internal link: The link given from one page to another within the same domain, showing the connection to the user and the search engine.
Backlink: The link another domain gives to a page on your site, able to create discovery, trust and traffic signals.
Canonical URL: The marker announcing to the search engine the preferred main URL for indexing identical or very similar content.
Indexing: The process of the search engine processing a discovered page and admitting it into its search index.
Crawling: The process of the search bot finding pages through links and the sitemap and reading their content and resources.
Sitemap: The XML file presenting the site's important URLs, and sometimes their update information, to the search engines.
Robots.txt: The root-level rules file telling the search bots which of the site's paths not to crawl.
Core Web Vitals: Google's user experience indicators measuring a page's loading, visual stability and interaction responsiveness.
Conversion: The user completing a pre-defined valuable action such as an enquiry, purchase, registration or download.
Lead: A potential customer who has shown interest in a product or service and, through a contact or behaviour signal, creates a future sales opportunity.
Funnel: The marketing and sales model showing the stages a person passes from first acquaintance through evaluation to the decision.
Landing page: The opening page built for one campaign, audience and goal, steering the user toward a concrete next action.
CTA: The button, link or short call to action telling the user which step to take now.
CRO: Optimising the experience and the page so that more users complete the target action, without raising the site's traffic.
A/B test: The method of comparing two variants' effect on the same core indicator by splitting them across an audience in a controlled way.
Retargeting: The approach of showing relevant ads again to people who previously interacted with the site or content.
Attribution: The model showing by which rule the conversion and revenue get distributed among the channels and campaigns the user touched.
KPI: The key performance indicator with a pre-agreed definition, measuring the progress toward a goal and influencing the decision.
ROI: The profitability measure showing the ratio of the net benefit obtained to the investment made.
CAC: The customer acquisition cost, calculated by dividing a period's sales and marketing spend by the number of new customers won.
LTV: The expected total value or contribution margin a customer will create over their relationship with the business.
Churn: The loss indicator showing the share of customers, users or revenue leaving the service in a given period.
Retention: The indicator of customers and users continuing their product use or buying relationship after a given period.
Personalization: Adapting the message, offer and experience to the existing information and behaviour signals about a person.
Segmentation: Splitting the audience into distinct groups by need, behaviour, value and other meaningful criteria.
First-party data: The data a business collects directly and within consent through its own site, app, sales and customer relationship.
Cookie: A small piece of data stored in the browser and used for purposes like the session, preferences, measurement and advertising.
Consent: The free and withdrawable agreement a person gives, understanding for what purpose their data will be collected and used.
Data privacy: The principles and practice of collecting and protecting personal data lawfully, transparently, purpose-fit and at minimal volume.
2FA: The account security method demanding, beyond the password, a second proof such as a phone, an app or a physical key.
Phishing: The fraud that, posing as a trusted institution or person, tries to obtain the user's password, payment and other sensitive data.
Encryption: Converting data by mathematical means into a form only the party with the matching key can read.
Backup: The verified spare copy kept apart from the main system for restoration at data loss and malfunction.
Prompt injection: The attack type where a user or external content tries to make the AI system break its earlier instructions and act without permission.
Human-in-the-loop: The oversight model where a human checks the result, handles the exception and gives the final approval on the critical decision in an AI process.

When does a definition turn into benefit?

Knowing the term does not automatically improve the decision. Saying "we're building RAG" is easy. Saying which source will be searched, whom the result serves and where a wrong answer will be stopped is harder. The benefit arises in the answers to those hard questions.

When meeting a new concept, split it into four: the input data, the work done, the output and the responsible person. Then pick one real example. If the word explains that example more clearly, it is useful. If it merely makes the presentation look more technical, it has solved nothing yet.

Where to verify the source?

Continuing the topic

A good glossary does not prolong the conversation.

It helps use the same word in the same sense.

Author

Anar Rustamli

I'm Anar Rustamli - a strategist, entrepreneur, and AI adoption leader working at the edge of growth, technology, and human thinking. Since 2016, my work has focused on helping businesses evolve in a rapidly changing digital landscape. I design growth systems, AI-powered workflows, and strategic frameworks that align performance with purpose. I believe real growth happens when strategy, data, and human insight work together - and my mission is to help businesses adopt AI in a way that strengthens both their results and their identity.