Concepts & glossary

An affiliate or sponsorship disclosure tells readers that you earn money or other value from a recommendation. The FTC's guidance says disclosures should be clear, close to the recommendation and hard to miss. A lone 'affiliate link' label or a single homepage notice may not be enough. IonBin's wiki pages carry no affiliate links at all.

The harness is everything around a model that makes it usable: tools, file access, memory, permissions, the interface and the loop that runs it. Practitioners including Ethan Mollick and Simon Willison argue the harness explains much of the gap between benchmark scores and real-world results.

An agent skill is a folder with a SKILL.md file: a short name and description, step-by-step instructions, and optional scripts or reference files. The agent loads it only when a task matches. Agent Skills is an open format that Claude and OpenAI Codex both read, so one written procedure can travel between tools.

Agentic commerce means an AI agent finds, compares and pays for something on a person's or company's behalf. Open protocols now exist for it: the Agentic Commerce Protocol (created by Stripe, OpenAI and Meta), Google's Universal Commerce Protocol, and x402 for paying for web content or APIs per request. Sellers need clear prices, scope and checkout that a machine can read.

An AI agent is a model that directs its own steps toward a goal. It plans, uses tools, checks the results against what actually happened and adjusts until the task is done or a limit is hit. A workflow or automation instead follows a fixed path that someone wrote in advance, whether or not the result is right.

AI governance is the set of policies, roles and checks that decide how an organization chooses, uses and monitors AI. For a small business it can be one page: approved tools, data rules, review requirements and an owner. The NIST AI Risk Management Framework offers voluntary structure, and the EU AI Act sets risk-based rules for some uses.

AI Overviews and AI Mode are Google Search features that write AI answers with links to supporting pages. Google says there are no extra technical requirements or special markup to appear: a page must be indexed and eligible for a snippet. Traffic from them is counted in Search Console's Web search type.

Answer engine optimization is the work of making your brand and content easy for AI answer engines to find, understand, trust and cite. It builds on SEO basics (crawlable, clear, helpful pages) and adds attention to what third-party sources say about you. Results are hard to measure precisely because AI answers vary.

An API is a defined way for one piece of software to request data or actions from another. In AI stacks, APIs let automations and agents read from and write to your tools without a person clicking through screens. If a tool has a good API, it can be automated; if not, you may need browser automation.

A background agent runs on the vendor's servers rather than your laptop. You give it a goal, a schedule or a trigger, and it works while you are away, then reports back for review. Claude Code routines, Google's Gemini Spark and scheduled tasks in Google Antigravity work this way. The trade-off: the vendor holds the credentials the agent acts with.

Batch APIs let you submit many AI requests at once and get the results later, in exchange for a lower price. OpenAI and Anthropic both charge 50% of standard prices for batch work, with a 24-hour window. Batches suit work that doesn't need an instant answer, such as tagging a backlog or generating descriptions overnight.

Computer use lets an AI model operate a computer the way a person does: look at screenshots, move the mouse, click and type. It reaches apps that have no API. It is slower and more error-prone than an API, so use it for gaps, in a sandbox, with human approval for anything that matters.

Content credentials are tamper-evident metadata attached to an image, video or audio file that record where it came from and how it was edited, including whether AI was used. They follow the C2PA open standard. They help audiences judge authenticity, but metadata can be stripped, so they are a signal, not proof.

Context rot is the drop in answer quality as a prompt grows long, especially with irrelevant or conflicting material. Studies have found that models use information at the start and end of long inputs better than in the middle, and that performance grows less reliable as input length increases, even on simple tasks. The fix is relevant, organized context.

The context window is how much material a model can consider in one request: your instructions, documents, the conversation so far and its own reply, measured in tokens. Current flagship API models from Anthropic and OpenAI list about 1M tokens. App plans can allow much less, so check the plan you use.

Data governance is how an organization decides who owns data, who can access it, how accurate it must be, and how long it is kept. AI raises the stakes: assistants and agents can read and copy data at machine speed. Good governance makes AI answers more trustworthy and leaks less likely.

An embedding is a list of numbers that represents the meaning of a piece of text, an image or other data. Similar meanings produce similar numbers, so software can find related content by meaning instead of keywords. Embeddings power semantic search, recommendations, clustering and RAG.

ETL (extract, transform, load) is the process of pulling data from source systems, cleaning and reshaping it, and loading it somewhere useful, such as a spreadsheet or warehouse. It is the plumbing behind reliable reporting. AI can now handle messy transform steps, like categorizing free text.

Evals are repeatable tests that check whether an AI system does a task well enough, using your own examples and a clear pass bar. They turn 'it seems fine' into evidence, and tell you whether a new model, prompt or vendor is actually better. A few dozen real examples is a practical starting point.

Fine-tuning further trains an existing model on your own examples so it adopts a specific format, tone or narrow skill. It isn't the best way to teach a model new facts. Retrieval and good prompts usually do that better. Vendor options are shrinking: OpenAI is winding down self-serve fine-tuning and Mistral marks its fine-tuning API as deprecated. Google and open-weight models still support it.

Generative AI is AI that creates new content, such as text, images, audio, video or code, in response to a prompt. Large language models are the best-known kind. For a business, it is a drafting and analysis engine, not a source of truth: outputs need checking.

Guardrails are the rules and checks around an AI system that keep it within safe, allowed behavior: input filters, output checks, permission limits and human approvals. They do not make a model perfect; they limit the damage when it is wrong or manipulated.

A hallucination is when an AI model states something false or made-up as if it were true, such as a fake statistic, quote or citation. It happens because models generate plausible text, not verified facts. Reduce it by giving sources, allowing 'I don't know', and checking claims before use.

Human-in-the-loop means a person reviews, approves or corrects an AI system's work at defined points, especially before actions that are costly or hard to undo. It is the main control for using agents safely in a business. The design question is where a human adds the most judgment for the least delay.

Incrementality testing measures what your marketing actually caused. You hold back a random group of people or regions from the ads, compare outcomes with those who saw them, and count only the difference. It is the honest check on attribution dashboards and on AI-written reports that confuse correlation with impact.

Inference is the act of running a trained AI model to get an output, as opposed to training it. Every chat reply, summary or classification is an inference call, and inference is what you pay for in AI APIs. Cost and speed depend on model size, tokens and hardware.

A knowledge cutoff is the date after which a model has no built-in knowledge. Anything newer must come from search, files or connectors at question time. Current vendor-stated cutoffs run from April to June 2026, so a model can be months out of date on prices, products and news.

A large language model is an AI system trained on very large amounts of text to predict the next piece of text. That simple skill lets it draft, summarize, translate, classify and work through problems. Claude, GPT, Gemini and Grok are examples of LLM families. ChatGPT and similar apps are products built on top of them.

llms.txt is a proposed Markdown file, usually at a site's root, that gives AI agents a short summary of the site and links to its most useful pages. It's a community proposal, now in a second version, not an adopted standard. Google says you don't need AI text files to appear in its AI features, so treat it as optional.

MCP is an open standard for connecting AI applications to outside systems: files, databases, business apps and workflows. A tool that publishes an MCP server can be used by any MCP-compatible assistant, such as Claude, ChatGPT, VS Code or Cursor, without a custom integration for each one.

Vendors sell each model family in tiers, from a slow, expensive flagship to fast, cheap models for high-volume work. On current published input prices the gap is large: about 100× between OpenAI's GPT-6 Astra and GPT-6 Luna. Choosing the lightest tier that does the job is usually the biggest cost lever.

A multi-model aggregator is a service that gives access to many AI models from different vendors through one account, interface or API. Examples include Gab AI, Perplexity and API routers such as OpenRouter. The business benefit is choice and flexibility; the cost is an extra party in your data path.

Multimodal AI can take in or produce more than one kind of data, such as text, images, audio and video, in the same system. In practice, you can show a model a screenshot, chart or PDF page and ask about it, or have it reply with speech. Most frontier model families are now multimodal to some degree.

OAuth is the standard that lets you give one app limited access to your account in another app without sharing your password, as when you click 'Connect Google'. It is how most AI assistants and automation hubs connect to business tools. Review the permissions you grant, and revoke what you no longer use.

Open-weight models publish their trained parameters so anyone can download and run them, subject to a license. Examples include Mistral Small and Large, Microsoft's Phi family and some Qwen models. They let you host AI yourself for privacy or cost reasons, but you take on security, updates and evaluation.

POAS measures gross profit returned per unit of ad spend, where ROAS measures revenue. Smart-bidding systems maximize whatever value you send them, so feeding revenue can scale low-margin sales. Break-even ROAS is 1 divided by gross margin: at a 40% margin you need 2.5x just to cover the product cost.

Prompt caching lets an AI provider reuse the processed start of a prompt that repeats across requests, such as long instructions or a reference document. Cached input is billed at a lower rate and returns faster. Put stable content first and variable content last to benefit.

Prompt injection is an attack where text hidden in an email, web page, document or tool result tells an AI system to ignore its instructions and do something else, such as leak data. It matters most when agents can read untrusted content and take actions. OWASP ranks it first (LLM01) in its 2025 Top 10 for LLM applications.

Rate limits cap how many requests or tokens you can send to an AI API in a period, such as per minute or per day. They protect the provider's capacity and usually rise as your account spends more. Hitting them causes errors, so automations need retries and backoff.

Reasoning models spend extra computation thinking through a problem before answering. They are stronger at multi-step math, planning, code and analysis, but slower and more expensive per answer. Use them for hard problems and use faster models for routine drafting.

RAG means fetching the most relevant passages from a larger library and giving only those to the model along with the question. It lets an assistant answer from material far bigger than its context window. The trade-off is that a retrieval step, not you, decides what the model gets to see.

Shadow AI is staff using AI tools at work without IT or management approval, such as pasting client data into a personal chatbot account. It usually signals real demand, not bad intent. The fix is an approved toolset and a clear policy, not a blanket ban that pushes use further out of sight.

Structured data is code on a web page that labels its content in a standard vocabulary, usually Schema.org, so search engines understand it: this is a product, this is its price, this is an FAQ. It can make pages eligible for rich results. Google says it must match the visible text and is not required for AI features.

Structured outputs make a model reply in an exact data format, such as JSON that matches a schema you define. Other software can then read the answer reliably, so AI can fill spreadsheets, CRM fields or forms without manual cleanup. OpenAI, Anthropic and Google document schema-based options. Some vendors offer only looser JSON modes.

Synthetic data is artificially generated data that mimics the patterns of real data. Teams use it to test systems, train models or share examples without exposing real customer records. It is only as good as the process that made it, and it can carry over the biases of its source.

A system prompt is the standing set of instructions a product gives a model before your message arrives: its role, rules, tone and the tools it may use. It shapes every answer. In your own setups, it is where your process belongs. In vendors' apps, it is usually hidden.

Temperature is a setting that controls how random a model's output is. Low values make answers more focused and repeatable, and higher values make them more varied. It has traditionally been tuned low for extraction and higher for brainstorming. Some newer models are built to run at their default, so check the vendor's guidance.

The lethal trifecta is Simon Willison's name for a dangerous agent setup: access to private data, exposure to untrusted content, and a way to send data out. With all three, a prompt injection hidden in an email or web page can make the agent leak your data. The fix is to remove at least one leg from every workflow.

Tokens are the chunks of text a model reads and writes. Context windows, output limits and API prices are all measured in them. Google says a Gemini token is about four characters, or 60 to 80 English words per 100 tokens. Prices are quoted per million tokens, and output usually costs several times more than input.

Tool use, also called function calling, lets a model ask your software to run a defined function, such as looking up an order or sending an email, and then use the result. It is how a chat model becomes an agent that can act. Your code, not the model, decides whether the call actually runs.

Turn detection is how a voice agent decides that you've finished speaking and it's time to answer, or that you've started talking again and it should stop. Good turn detection makes voice AI feel natural. Bad detection interrupts people or leaves awkward silences. Some speech models now build it in, such as Deepgram's Flux and OpenAI's semantic VAD.

A vector database stores embeddings and finds the ones closest to a query quickly. It is the usual storage layer for semantic search and RAG over large document sets. Small projects may not need one: a few hundred documents can fit in a model's context window or a simple file.

A webhook is a message one system sends to another the moment something happens, such as a payment succeeding or a form being submitted. It is how automations start without constant checking. Stripe, CRMs and form tools all send webhooks that automation hubs and agents can act on.

A zero-click search ends without a click to any website: the answer appears on the results page, or the person searches again. SparkToro's 2024 clickstream study put it at 58.5% of US Google searches. AI Overviews and AI Mode push more of the journey onto Google's page, so being named in the answer matters alongside ranking.