AI

Prompt Engineering is the practice of composing text queries to LLM (Claude, GPT) to obtain the desired result.

LLM (Large Language Model) is a large language model such as Claude, GPT, Gemini. A foundation of AI tools used in marketing in 2026.

AI marketing - implementation of neural networks and LLM in marketing production: creatives, texts, analytics, automation.

Claude is an LLM from Anthropic. In 2026, Claude Sonnet 4.6 is the best model for long texts and code generation in my stack.

The AI agent is an LLM that not only responds, but performs a chain of tasks: calls tools, accesses the API, and makes decisions step by step.

RAG is a combination of “search + LLM”: the model first retrieves relevant data from your database, then responds based on it. Fewer hallucinations.

Context window - how much text (in tokens) the model holds at a time: both your prompt and its response. Overfilled - the model “forgets” the beginning.

Tokens are pieces of text that LLM operates with. They calculate both the context window and the cost of the API. Russian text “eats” more tokens.

Hallucination - when LLM confidently produces a false fact: a non-existent figure, link, quote. The main risk of AI content without verification.

Fine-tuning - additional training of a ready-made LLM using your own data to a stable style or format. Expensive; For a marketer, a prompt and examples are often enough.

Multimodality is the ability of a model to work not only with text, but with pictures, video and audio: both understand them and generate them.

System prompt - constant instructions to the model: its role, response style and rules. Set once, affects the entire dialogue.

Zero-shot - you ask without examples. Few-shot - you give the model 2-5 examples of “input → output”, and it copies the pattern. The simplest way to improve quality.

Chain-of-Thought is a technique where the model reasons step by step before answering. On problems with counting and logic, accuracy is noticeably higher.

MCP (Model Context Protocol) is a standard by which AI tools connect to external systems: CRM, databases, API. "USB port" for agents.

Embeddings are a representation of text as a vector of numbers, where texts with similar meanings lie side by side. The basis of semantic search and RAG.

AI Overview - generated answer at the top of the search results. Takes clicks from regular links; SEO's job is to get to its sources.

GEO - optimization of content for citation in AI responses (ChatGPT, Perplexity, AI Overview). A New Layer of SEO: The goal is to become a source, not a link.

AI detection - determining that text or photos were generated by a neural network. Search engines demote “naked” AI content without editing or substantive detail.

n8n is a visual workflow automation builder: you connect services and AI into working scenarios without code. The basis of marketing auto-funnels and AI agents.

Claude Code is an Anthropic terminal AI agent: writes code, edits files, works with git and tools. I use it to build websites and bots.

Veo is Google's generative video model: a video from a text description with sound. Replaces part of video production for advertising and social networks.

Vibe coding is the creation of software by describing a task to an AI agent in natural language, without manual code. This is how a marketer collects landing pages and bots himself.

GPT stands for Generative Pre-trained Transformer and refers to a family of OpenAI language models. GPT models are based on the Transformer architecture.

Gemini is a multimodal LLM from Google DeepMind, built into Search, Workspace and Android.

Perplexity is an AI search engine that provides answers with sources instead of a list of links.

Midjourney is a text-to-image generator used to create marketing visuals.

Stable Diffusion is an open-source image generator that works locally without a subscription fee.

Sora is a video generator from OpenAI: a text prompt turns into a video of up to 20 seconds.

Function Calling is a mechanism that allows LLM to call external APIs and tools within a conversation.

Temperature is a parameter from 0 to 2 that controls the randomness of LLM responses: 0 - deterministic, 2 - chaotic.

Top-p and Top-k are LLM sampling parameters that limit the pool of candidate tokens for generation.

Vector DB is a database for storing embeddings and searching by semantic proximity, the basis of RAG systems.

Semantic search finds documents based on semantic similarity rather than exact matches of keywords.

Prompt Injection is an attack on LLM: malicious instructions are hidden in the data that the model processes.

Jailbreak is a technique for bypassing LLM security restrictions through special prompts or scripts.

Knowledge Cutoff - the date after which LLM has no information about events: training has stopped.

YandexGPT is a Yandex language model built into Alice, Direct, Metrica and B2B products.

GigaChat is a language model from Sberbank with an API accessible in Russia. Personal-data compliance depends on the deployment, configuration and data-processing arrangements.

LangChain - Python/JS framework for building agents and LLM call chains with tools.

LlamaIndex is a framework for RAG: indexing documents, searching for relevant fragments and transferring them to LLM.

Open-source LLM - open language models (Llama 3, Mistral, Qwen) for launching without dependence on the API.

AI Watermark is an invisible mark in AI-generated content to identify machine generation.

AEO - content optimization for AI search engines (Perplexity, ChatGPT) instead of traditional SEO.

Cursor is a code editor with built-in Claude/GPT that allows marketers to write automation without development experience.

Replit Agent is an AI agent in the browser: it writes, runs and deploys code based on the text description of the task.

Synthetic data - AI-generated data for training models when real data is insufficient.

Nano Banana is an image generation model from kie.ai that I use to create blog covers.