AI & software glossary

Software that plans and takes multi-step actions toward a goal.

An AI agent is a system built on a language model that can plan a task, call tools or APIs, observe the result, and decide the next step — repeating until a goal is met or it hands control back to a person. Unlike a single prompt-and-response, an agent operates over multiple steps and is given a bounded task, allowed tools, a stopping condition, and an escalation path.

A model trained on text that predicts and generates language.

A large language model is a neural network trained on very large text corpora to predict the next token in a sequence. That objective produces models that can summarise, translate, answer questions, write code, and follow instructions. LLMs do not look facts up by default, so production systems ground them in trusted data and constrain their outputs.

Grounding an LLM's answers in retrieved documents.

Retrieval-augmented generation is a pattern where, before the model answers, the system retrieves relevant passages from your own content — documents, tickets, a database — and includes them in the prompt. The model then answers from that supplied context rather than its training data alone, which improves accuracy, allows citations, and lets the answer stay current as your content changes.

Building software mostly by prompting AI rather than writing code.

Vibe coding describes building an application primarily by describing what you want to an AI tool and accepting its generated code, with little manual review. It is fast for prototypes and demos, but the output often lacks real authentication, input validation, error handling, tested architecture, and load capacity — the work needed before real users depend on it.

Closing the gaps between a working prototype and a real product.

Production hardening is the work of taking an app that runs in a demo and making it safe for real users and traffic: a security and architecture audit, proper authentication and access control, input validation and error handling, a maintainable code structure, and performance and load testing before launch — keeping the parts that already work.

When an AI states something fluent but false.

A hallucination is a confident, well-formed AI output that is not supported by the input or by fact. It happens because language models generate plausible text rather than retrieve verified answers. Retrieval grounding, output constraints, evaluations, and a model that can say “I don't know” all reduce how often hallucinations reach a user.

Repeatable tests that measure whether an AI feature does its job.

Evals are a curated set of representative cases — ordinary, ambiguous, edge, and adversarial — run against an AI feature to measure task completion, factual grounding, and policy compliance. A fast subset runs in pull requests; a broader suite runs before release. Automated graders are calibrated against expert human review so the score reflects real quality.

Controls that keep an AI system inside safe, allowed behaviour.

Guardrails are the deterministic checks and policies around a model: input and output validation, permission checks, forbidden-action lists, required citations, rate and cost limits, and safe fallbacks when confidence is low. They let a probabilistic model operate inside a system the organisation can reason about and audit.

Numeric representations of meaning used for similarity search.

An embedding is a list of numbers that captures the meaning of a piece of text, so that similar meanings sit close together in vector space. Storing embeddings of your content in a vector database lets a system find the passages most relevant to a question — the retrieval step behind RAG and semantic search.

The smallest build that tests the core value with real users.

An MVP is the most focused version of a product that still delivers its central value and can be put in front of real users to learn from. It is scoped to answer the riskiest question quickly and cheaply, and is built to be extended — not thrown away — once the evidence comes back.

Improving visibility in traditional search engines.

SEO is the practice of improving how well a site ranks in search engines such as Google and Bing, through technical quality (fast, crawlable, well-structured pages), relevant content depth, structured data, clean URLs, and authority signals like links and named expertise.

Making content likely to be cited by AI answer engines.

GEO is optimising content so AI systems such as ChatGPT, Perplexity, and Google's AI Overviews are likely to cite, quote, or recommend it. It rewards crawler access for AI bots, factual density, named and credentialled authorship, third-party validation such as reviews, and a clear, consistent picture of what an organisation does.

Structuring content to win direct-answer placements.

AEO is structuring content to win featured snippets, voice answers, and AI answer panels: question-phrased headings, direct 40–60 word answers, FAQ and HowTo structured data, tables and lists, and speakable markup — so a machine can lift a clean answer straight from the page.

Experience, Expertise, Authoritativeness, Trustworthiness.

E-E-A-T is the framework Google's quality raters use to judge content, standing for Experience, Expertise, Authoritativeness, and Trustworthiness. It is demonstrated with named authors and their credentials, first-hand experience of the topic, third-party validation such as reviews and citations, and transparent organisation information.

Machine-readable labels that describe a page's content.

Structured data is markup — usually JSON-LD following the schema.org vocabulary — that labels the entities on a page: an Organization, a Service, an Article and its author, an FAQ, a HowTo, a job posting. It lets search and AI systems understand the page precisely and makes it eligible for rich results.

An open standard for connecting AI models to tools and data.

MCP is an open protocol that lets an AI application connect to external tools, data sources, and services through a common interface, instead of a custom integration for each one. It standardises how a model discovers available tools, calls them, and receives results, making agent systems easier to build and audit.