AI search has added a new term to our work almost every week.
I have worked in SEO for more than 10 years. Yet while building RadarKit and studying millions of AI answers, I still found myself stopping to check what a new term meant and how it was different from the last one.
AEO, GEO, LLMO, RAG, grounding, query fan-out and MCP often get used as if everyone already knows them. Most people do not. Even experts do not always agree on every definition.
So I made this glossary for SEO teams, marketers and founders who want plain answers.
The interest is not only coming from our industry bubble. DataForSEO currently estimates about 4,400 monthly US searches for generative engine optimization, 2,400 for answer engine optimization and 1,300 for AI search optimization. People are trying to understand this shift while it is happening.
AEO
AEO means Answer Engine Optimization.
It is the work of helping a brand, product or page appear in answers from systems such as ChatGPT, Perplexity, Gemini, Microsoft Copilot and Google AI Overviews.
SEO mainly aims to earn a place on a search results page. AEO aims to earn a place inside the answer itself. That place could be a brand mention, a linked citation, a product card or a local recommendation.
Example: someone asks an AI assistant for the best CRM for a small agency. If it names your CRM, your AEO work has helped you become part of the answer.
GEO
GEO means Generative Engine Optimization.
It is often used as another name for AEO. Some people make a small difference between them:
- AEO can include any answer engine, including voice search and featured snippets.
- GEO focuses more on answers created by generative AI.
In daily work, most teams use AEO, GEO, AI SEO and LLM optimization for very similar work. I prefer AEO because it explains the goal clearly: be part of the answer.
LLMO
LLMO means Large Language Model Optimization.
It is another term for improving how a brand or its content appears in LLM-based products. It puts more focus on the model, while AEO puts more focus on the answer.
The label matters less than the work. You still need clear facts, strong entities, trusted sources and useful content.
LLM
LLM means Large Language Model.
An LLM is the model that reads and creates language. GPT, Gemini and Claude model families are examples.
ChatGPT is an application that uses models. The app and the model are not the same thing. A simple way to see it is that ChatGPT is the product you use, while the LLM is the system doing much of the language work behind it.
Generative AI
Generative AI is AI that creates new output from patterns learned from data. It can create:
- Text
- Images
- Audio
- Video
- Code
An LLM is one type of generative AI model. Not every generative AI system is an LLM.
AI answer engine
An answer engine gives the user a direct answer instead of only showing a list of links.
ChatGPT, Perplexity, Gemini and Copilot can all act as answer engines. Google Search also acts like one when it shows an AI Overview or an AI Mode answer.
This changes what search visibility means. A page can rank well on Google but still be missing from the final AI answer.
Prompt
A prompt is the question or instruction a user gives an AI system.
In AEO tracking, prompts work a little like keywords, but they are not identical. Prompts are often longer, more detailed and shaped by context.
A keyword may be CRM software. A prompt may be Which CRM is best for a five-person agency that needs simple reporting?
One prompt can also lead to several hidden searches. That brings us to query fan-out.
Query fan-out
Query fan-out happens when an AI system turns one prompt into several related searches before it gives an answer.
For example, a user may ask whether a company is a good investment. The AI system may search for its latest earnings, recent news, analyst views and debt before forming one reply.
We see this directly in RadarKit. A prompt such as query fanout tracking tool has produced hidden searches around SEO, search visibility, databases and observability. The wording changes by model and by day.
This is why optimizing one exact phrase is not enough. Your content needs to cover the full problem and its related questions.
Retrieval
Retrieval is the step where an AI system finds outside information before answering.
The information may come from a web search, a private database, uploaded files or another connected source. Retrieval helps the system use facts beyond the model’s training data.
RAG
RAG means Retrieval-Augmented Generation.
With RAG, the system first retrieves useful information and then gives that information to the model to create an answer.
A basic question may be answered from model knowledge. A question about today’s stock price, weather or news needs fresh data. RAG is one common way to supply it.
RAG does not promise a correct answer. The system can retrieve a poor source, miss a key detail or read the source badly. It still needs good retrieval and good reasoning.
Grounding
Grounding means tying an AI answer to outside evidence.
The evidence may be a webpage, a company database, a product feed or a document. A grounded answer should be easier to check because its claims come from a known source.
RAG is a method used for grounding. The two terms are related, but they are not the same. Grounding is the goal; RAG is one way to reach it.
Training data
Training data is the large body of content used to teach a model before it is released.
It shapes what the model knows and how it responds. It is different from live retrieval. A model does not search its training data like a normal search index every time a user asks something.
For marketers, this means new content may help through live search and citations long before it could ever affect a future model training run.
AI crawler
An AI crawler is an automated bot that visits webpages. Different bots can have different jobs.
The simplest useful split is:
- Search bots collect pages for AI search and citations.
- User agents visit a page because a user asked an AI product to open or use it.
- Training bots collect content that may be used to improve future models.
Do not treat every AI bot as the same. A site owner may want to allow search access while making a different choice about model training.
robots.txt
robots.txt is a file that tells automated crawlers which parts of a site they may access.
It can give separate rules to search bots, training bots and other crawlers. It is a control file, not a ranking tool. Allowing a bot does not mean the bot will cite you, and blocking one bot does not block every AI system.
AI citation
An AI citation is a page or domain linked or named as a source for an AI answer.
Citations matter because they show which pages helped ground the response. They can also send referral traffic and build trust.
A citation is not the same as a brand mention. Your page can be used as a source without your brand being recommended. Your brand can also be named without a link to your site.
Brand mention
A brand mention happens when an AI answer names a company, product or person.
The mention may be linked or unlinked. It may also be positive, mixed or negative.
For most companies, mentions are the clearest AEO outcome. If a buyer asks for a product and the AI names your brand, you have entered the buyer’s shortlist.
AI visibility
AI visibility measures how often a brand appears across a chosen set of prompts and AI platforms.
There is no single score used by every tool. One platform may count any mention. Another may give more weight to position, sentiment or prompt value.
Always check what sits under the score. The raw answers, prompts, locations and models matter more than one big number.
AI share of voice compares your brand mentions with competitor mentions across the same tracked prompts.
If five brands receive 100 total mentions and your brand receives 20, your share of voice is 20% for that dataset.
It is useful for tracking market presence, but the prompt set must be fair. A poor or very small prompt list can give a misleading result.
Average position
Average position shows where a brand tends to appear when an AI answer gives a ranked or ordered list.
Being mentioned first is often more useful than being named near the bottom. Still, AI answers are not always clean ranked lists, so position should be read with mentions, sentiment and full answer text.
Sentiment
Sentiment shows whether the answer talks about a brand in a positive, mixed or negative way.
Two brands can have the same visibility but very different results. One may be praised as the best choice. The other may appear only in a warning or list of limits.
This is why mention count alone is not enough.
Hallucination
A hallucination is an answer that sounds sure but includes false or unsupported information.
It can happen when the model relies on weak memory, joins unrelated facts or reads a source badly. Good grounding lowers the risk, but it does not remove it.
Brands should track wrong claims about pricing, features, founders and policies. Correcting clear source pages can help AI systems find better facts later.
Entity
An entity is a known person, company, place, product or idea.
The words Apple can mean a fruit or a company. Entity understanding helps a search or AI system know which one the page means.
Strong entity signals come from clear names, consistent facts, useful About pages, trusted mentions and links between related people, products and companies.
Named Entity Recognition
Named Entity Recognition, or NER, is the process of finding and labeling entities inside text.
For example, a system can label Sanjay Singh as a person, RadarKit as an organization, Mumbai as a place and $29 as a price.
This turns free text into structured information that software can compare and use.
Semantic search
Semantic search tries to understand meaning and intent, not only exact words.
A user can ask for an affordable tool for a small team without using the phrase low-cost software. A semantic system can still understand that both ideas are related.
For content teams, this means topic depth and clear meaning matter more than repeating the same keyword.
Passage retrieval
Passage retrieval finds the most useful section of a page instead of treating the whole page as one block.
A long guide may contain one paragraph that gives the best answer. A search or AI system can retrieve that passage even if the rest of the page covers other topics.
Clear headings and focused sections make useful passages easier for people and machines to find.
Chunking
Chunking means splitting content into smaller pieces that can be stored, searched and passed to a model.
It is common in RAG systems. A long document may be split into sections, then only the most relevant sections are sent to the model.
For public web content, do not force every paragraph into a fixed formula. Write clear sections that answer one real question well. Good structure already does much of the work.
Structured data and schema markup
Structured data is machine-readable information added to a webpage. Schema markup is the common vocabulary used to describe it.
It can tell search systems that a page contains a product, price, review, person, organization, event or FAQ.
Schema does not force an AI system to cite a page. It reduces confusion by making important facts easier to identify.
Knowledge graph
A knowledge graph stores entities and the links between them.
It may connect a founder to a company, a company to a product and a product to a category. Search engines use these links to understand what something is and how it relates to the wider world.
Clear and consistent entity information across your site and trusted third-party sources can support this understanding.
Markdown
Markdown is a simple text format that uses symbols for headings, lists and links.
It is lighter than a full webpage and easy for both people and software to read. Many documentation sites offer Markdown versions because agents can use the clean content without extra menus, scripts and layout code.
Markdown is not a replacement for a good HTML website. HTML is still the main format for normal web pages, search, design and accessibility.
Markdown negotiation
Markdown negotiation happens when a client asks a server for a Markdown version of a page.
The request can use an HTTP header such as:
Accept: text/markdownIf the server supports it, it can return the clean Markdown version instead of the normal HTML page. This can save an agent from processing extra page code.
HTTP header
An HTTP header is a small piece of information sent with a web request or response.
Headers can tell a server which content format the client accepts, which language it wants or how content should be cached. They are part of the request and response, not the visible page.
llms.txt
llms.txt is a proposed Markdown file that gives AI tools a short guide to a website and links to its useful content.
It is usually placed at /llms.txt. Think of it as a curated map for an agent, not a replacement for robots.txt or sitemap.xml.
The important word is proposed. Publishing the file does not promise better rankings, more citations or even that major AI systems will read it. Use it as a low-cost support file, mainly when your site has large docs or clean Markdown pages. Do not treat it as an AEO shortcut.
YAML frontmatter
YAML frontmatter is a small metadata block placed at the top of a Markdown file.
It can hold the page title, author, date, tags or content type. Static site tools and other software can read this block without guessing those facts from the article body.
OKF
OKF means Open Knowledge Format.
It is an early idea for arranging website knowledge as linked Markdown files with clear metadata. An index file acts as the entry point, while other files hold topics and update history.
The goal is to give agents a clean set of information without making them parse a full website. Treat OKF as an experiment, not a common web standard or a proven ranking factor.
MCP
MCP means Model Context Protocol.
It is a standard way for AI applications to connect with tools and data.
For example, the RadarKit MCP lets an AI assistant access live visibility, citations, competitors and query fan-out data. Instead of exporting a CSV and uploading it each time, the assistant can request the needed data through the connection.
An MCP server is the bridge. The AI application is the client using that bridge.
API
An API lets one software product request data or actions from another in a defined way.
MCP and APIs overlap, but they are not the same. APIs are a broad software building block. MCP gives AI agents a common method to discover and use tools, and an MCP server may call APIs behind the scenes.
WebMCP
WebMCP is a way for an AI agent to use tools exposed by the webpage open in a browser.
Normal MCP often connects to a backend service. WebMCP works with the current page and browser state. The two can support different parts of the same task.
For example, a backend connection may fetch hotel data, while a browser tool may help complete the booking form on the page.
Agentic browsing
Agentic browsing happens when an AI system uses a browser to complete a goal.
The user gives the goal, and the agent decides the steps. It may open pages, search, click buttons, fill forms, check the result and continue until the task is done.
This makes website structure important. Clear labels, stable forms and working buttons help both people and agents.
Accessibility tree
The accessibility tree is a simpler map of a webpage made for assistive technology such as screen readers.
It describes elements by their role, name and state. A button can be understood as a button even when the visual design is complex.
Browser agents can also use this structure to understand and control a page. A clean accessibility tree makes a site easier for both disabled users and AI agents.
Semantic HTML
Semantic HTML uses page elements based on their meaning.
Examples include main, nav, article, button and clear heading levels. These elements tell browsers, screen readers and agents what each part of the page does.
A real button is easier to understand than a generic box made clickable with scripts. Good semantic HTML is basic web quality, but it also helps agentic browsing.
What this glossary means for your AEO work
Knowing the terms is useful. Connecting them is better.
An AI answer engine receives a prompt. It may create a query fan-out, retrieve pages, select useful passages and ground an answer. It may then mention a brand, cite a source and place products or businesses in an ordered list.
Your job is to make the right facts easy to find, understand and trust across your own site and third-party sources.
The third-party part matters. In one recent RadarKit dataset covering 21,230 source appearances across 12 AI visibility prompts, YouTube appeared 797 times, Zapier 752 times and Reddit 651 times. AI systems were not using only product websites. They were pulling from videos, reviews, communities and trusted roundups.
That leads to a simple AEO plan:
- Publish clear first-party facts on your website.
- Build useful pages around real buyer questions.
- Use strong entities, schema and semantic HTML.
- Earn credible third-party mentions and reviews.
- Create content in formats AI systems already use, including video.
- Track prompts, citations, mentions, position and sentiment.
- Check the full answers instead of trusting one visibility score.
AEO is moving fast, and some terms will change. The core work is less new than it looks: publish reliable information, make it easy to understand and build enough trust that other sources support what you say.
That is the part worth focusing on.
