AI Glossary

An AI agent is an AI system that doesn’t just answer questions but plans and carries out multi-step tasks on its own — searching, writing, using tools, and checking its own work to reach a goal. In marketing and sales, agents can research prospects, draft campaigns, or keep a CRM up to date without step-by-step human instruction. Agents are the building blocks of agentic AI; several working together form a swarm.

Agentic AI is AI that works toward a goal on its own: it plans the steps, does the work, uses tools, and adjusts as things change, instead of waiting for the next instruction. It is the difference between a chatbot that answers what you ask and an agent that takes a goal and runs with it. Agentic marketing and sales is simply this applied to growth work.

An AI Overview is the AI-generated answer Google places above the normal search results. It often cites a handful of sources, and being one of them is the new top of the page. Earning those citations is the job of GEO, and it can matter more than a classic ranking.

Bias in AI is a systematic skew in a model’s outputs, caused by imbalances or blind spots in the data it was trained on. For marketers this matters in practice: a biased model might default to certain demographics in ad copy or scoring leads unevenly. Human review is the standard safeguard.

A context window is the amount of text an AI model can “keep in mind” at once — the conversation, documents, and instructions it’s currently working with, measured in tokens. When a task exceeds the context window, the model starts losing track of earlier content, which is why long documents are often split or summarised first.

Deep learning is a type of machine learning that uses many-layered neural networks to learn patterns directly from large amounts of data. It’s the technology behind modern image recognition, speech-to-text, and the large language models that power today’s AI assistants.

An embedding is a numerical representation of a piece of text (or an image) that captures its meaning, so software can measure how similar two pieces of content are. Embeddings power semantic search, content recommendations, and RAG — they’re how an AI finds “the right document” even when the keywords don’t match.

Fine-tuning is the process of training an existing AI model further on your own examples so it performs better at a specific task. A common commercial use: fine-tuning a model on past campaigns so it writes consistently in your brand voice. For most teams, good prompting and RAG get you there faster and cheaper.

GEO is the practice of getting your brand mentioned and cited by AI answer engines like ChatGPT, Perplexity and Google’s AI Overviews. Where SEO aims for a ranking, GEO aims to be the source the AI quotes. As more buyers ask an assistant instead of a search box, it becomes its own discipline.

Guardrails are the rules and limits that keep an agent acting safely and on-brand: what it may and may not do, which claims it can make, and when it must hand off to a human. Good guardrails are what let you give an agent real autonomy without real risk.

A hallucination is when an AI model states something false as if it were fact — an invented statistic, a fake source, a product feature that doesn’t exist. It’s the main reason a human owns the quality of anything that goes out. RAG and clear sourcing reduce it, but do not remove the need to check.

A knowledge graph is a structured map of entities — people, companies, products — and the relationships between them, in a format machines can read. Search engines and AI assistants use knowledge graphs to understand who your company is and what it offers, which makes them a foundation of both SEO and GEO.

A large language model is the kind of AI behind ChatGPT, Claude and Gemini: trained on huge amounts of text to predict and generate language, it can write, summarise, answer and reason over words. It is the engine most marketing AI runs on, and an agent is usually an LLM given tools and a goal.

Machine learning is software that learns patterns from data instead of following hand-written rules. It covers everything from lead scoring and churn prediction to deep learning. In marketing you often use it without noticing: the recommendations, the spam filter, the bid optimiser.

MCP is an open standard for connecting AI models to the tools and data they need — a CRM, an analytics account, a file store — through one consistent interface. It is a big part of why agents can now actually do work in your systems rather than just talk about it.

A neural network is a model loosely inspired by the brain: layers of simple units that pass signals to each other and adjust as they learn. Stack enough layers and you get deep learning, the approach behind today’s language and image models.

A prompt is the instruction you give an AI model: the question, task or context you type in. The quality of the prompt largely decides the quality of the output, which is why prompt engineering is a real skill.

Prompt engineering is the practice of writing instructions for an AI model so it produces consistently good results. It covers everything from phrasing and examples to structure and context — and it’s the single highest-leverage AI skill a marketing or sales team can build.

RAG is a technique where an AI first retrieves relevant documents — your own content, data or knowledge base — and then writes its answer from them, instead of relying only on training. It uses embeddings and often a vector database, and it is the main way to make an AI accurate about your business and reduce hallucination.

SEO is the practice of earning visibility in search results so people find you when they look for what you offer. It still matters, and it now has a sibling: GEO, which does the same job inside AI answers rather than the classic list of links.

Semantic search finds content by meaning rather than exact keywords, so a search for “cut churn” can surface a page titled “reduce customer loss.” It runs on embeddings and is the retrieval half of RAG.

A swarm is several agents working together toward one goal, each handling its part and passing work along — closer to a team than a single tool. Directed by a human, a swarm is how a small team can cover a whole go-to-market surface at once. (It is also where the name Parvi comes from: parvi is Finnish for a flock.)

A token is the small chunk of text — roughly a word or part of one — that an AI model reads and writes in. Models price and limit work by tokens, and the context window is measured in them. It is the unit that quietly drives both cost and capacity.

Unsupervised learning is machine learning on unlabelled data, where the model finds patterns and groupings on its own instead of being shown correct answers. A classic marketing use: customer segmentation, where the algorithm discovers natural clusters in your audience that nobody defined in advance.

A vector database stores embeddings so an AI can search your content by meaning at speed. It is the memory behind RAG and semantic search: the place a business keeps its knowledge in a form an AI can actually use.