GEO & AEO Glossary: AI Search Terms, Defined

Search no longer ends with a list of links. Tools like ChatGPT, Perplexity, and Google AI Overviews now read the web on a user’s behalf and return a written answer. That shift has created a new vocabulary, and many of its terms are still used loosely or inconsistently.

This glossary defines the terms that matter for Generative Engine Optimization (GEO) and Answer Engine Optimization (AEO). Each entry opens with a plain definition you can read on its own, followed by extra detail only where it helps. Where the industry disagrees on a term, or where a common definition is wrong, the entry says so.

How to Use This Glossary

Terms are grouped into eight sections, from core disciplines to measurement and agentic search. Each term works as a standalone entry, so you can jump straight to the one you need. Traditional SEO terms are left out unless their meaning changes in an AI search context.

GEO vs AEO vs LLMO vs AIO: What’s the Difference?

These four terms overlap heavily and are often used as synonyms. The table below shows how they are most commonly distinguished, although none of them has a single agreed definition.

Term	Main focus	Typical surfaces	Common measure of success
GEO	Being used and cited in AI-generated answers	ChatGPT, Perplexity, Gemini, AI Overviews, AI Mode	Citation rate, share of model
AEO	Being the extracted direct answer to a question	Featured snippets, voice assistants, AI answers	Answer placements, citations
LLMO	How language models understand and describe a brand	LLM training data and live retrieval	Mention accuracy, brand sentiment
AIO	Either all AI search work, or Google AI Overviews only	Varies by who is using the term	Varies

Section 1: Core Disciplines

Generative Engine Optimization (GEO)

GEO stands for Generative Engine Optimization. It is the practice of improving how often and how accurately a brand, website, or piece of content appears in answers written by AI systems such as ChatGPT, Perplexity, Gemini, and Google AI Overviews. The term was popularized by a 2023 research paper from Princeton, Georgia Tech, the Allen Institute for AI, and IIT Delhi.

Core Principles

• Retrievability: AI systems that search the web can only use pages their crawlers can access and read.
• Extractability: Content organized into clear, self-contained passages is easier for AI systems to lift into an answer.
• Credibility: AI systems tend to favor sources that other reputable sources agree with and mention.
• Brand presence beyond your site: What reviews, news coverage, and directories say about a brand shapes how AI describes it.

GEO can improve the chances of being cited, but no technique guarantees that an AI system will mention or cite any source.

Answer Engine Optimization (AEO)

AEO stands for Answer Engine Optimization. It is the practice of structuring content so that a search or AI system can pull out a clear, direct answer to a specific question. It began with featured snippets and voice assistants and now also applies to AI-generated answers.

AEO vs GEO

The two terms are often used as synonyms, and the line between them is not fixed. AEO usually focuses on the page itself, making individual answers easy to extract. GEO is usually broader and includes a brand’s presence across the wider web, how AI systems retrieve information, and how often a brand is cited over time.

Large Language Model Optimization (LLMO)

LLMO stands for Large Language Model Optimization. It is the practice of influencing how large language models understand, remember, and describe a brand, both through the data they were trained on and the pages they retrieve live. It overlaps heavily with GEO, and many practitioners treat the two as the same discipline.

LLMO vs GEO

When a distinction is made, LLMO puts more weight on the model’s overall understanding of a brand as an entity. GEO puts more weight on appearing in specific AI search answers and being cited as a source.

AIO

AIO is an acronym with two common meanings in search marketing. It can stand for AI Optimization, a broad label for any work aimed at AI-driven search, or it can refer to Google AI Overviews, the AI-generated summaries shown in Google results. Because the acronym is used both ways, it is worth confirming which meaning is intended in any strategy or report.

AI SEO

AI SEO is a broad, informal term for search optimization aimed at AI-powered search experiences. It is often used as an umbrella that covers GEO, AEO, and LLMO. The term is also sometimes used to mean using AI tools to do traditional SEO work, which is a different thing entirely.

Section 2: Answer Engines & AI Search Surfaces

Answer Engine

An answer engine is a search system that responds to a question with a direct answer instead of a list of links. It gathers information from its training data, live web pages, or both, and presents a summarized response that often cites its sources. The term is used loosely and has no single agreed definition.

Common Forms of Answer Engines

• AI assistants: Chat tools such as ChatGPT and Claude that can answer questions directly, with or without web search.
• AI-native search tools: Products such as Perplexity, built around cited answers from the start.
• AI features in search engines: Google AI Overviews and AI Mode, which add generated answers to a traditional search engine.
• Voice assistants: Tools such as Siri and Alexa, which read out a single spoken answer.

These forms overlap, and many products fit more than one.

Google AI Overviews

Google AI Overviews are AI-generated summaries that appear at the top of some Google search results. They are produced by Google’s Gemini models and include links to the web pages used as sources. Google does not show them for every query, and which queries trigger them changes over time.

AI Overviews vs Featured Snippets

A featured snippet quotes a passage from a single page. An AI Overview writes a new summary that can draw on several pages at once and link to more than one of them.

Google AI Mode

Google AI Mode is a separate, conversational search experience inside Google Search that answers questions with AI-generated responses. It is designed for longer or more complex questions and supports follow-up questions in the same session. Google has said it uses query fan-out, which means it runs many related searches behind the scenes before writing its answer.

AI Mode vs AI Overviews

AI Overviews appear above the regular results on a standard search page. AI Mode is a separate view where the AI answer replaces the regular results page, and users can keep the conversation going.

Search Generative Experience (SGE)

SGE stands for Search Generative Experience. It was the name of Google’s experimental AI search feature, tested through Search Labs starting in 2023. It was later replaced by AI Overviews, so the term now appears mainly in older articles and research.

ChatGPT Search

ChatGPT Search is the web search capability built into ChatGPT, OpenAI’s AI assistant. When a question needs current information, ChatGPT can search the web, read relevant pages, and write an answer with links to its sources. OpenAI has not published full details of which search indexes it relies on, and this may change over time.

Perplexity

Perplexity is an AI-powered answer engine that searches the web and responds with a written answer, usually including numbered citations linked to its sources. Because it shows sources consistently, many GEO practitioners use it as a reference point when checking which sites get cited for a topic.

Gemini

Gemini is Google’s family of large language models and the name of its standalone AI assistant app. The same model family also powers AI Overviews and AI Mode in Google Search. As a result, Gemini’s behavior matters across several Google surfaces, not just the Gemini app.

Microsoft Copilot

Microsoft Copilot is Microsoft’s AI assistant, available in Windows, the Edge browser, Bing, and Microsoft 365 apps. It uses Bing’s search results to ground its answers in current web content. This makes a site’s visibility in Bing relevant to how it appears in Copilot answers.

Claude

Claude is the AI assistant developed by Anthropic. When web search is turned on, Claude can look up current information and cite the pages it used. Without web search, it answers from what it learned during training.

AI Search Engine Results Page (AI SERP)

An AI SERP is a search results page that includes AI-generated content alongside, or in place of, traditional links. On Google, this can mean an AI Overview above the regular results, while on Perplexity the entire page is an AI answer. The term is informal and does not have one fixed meaning.

Featured Snippet

A featured snippet is a highlighted box at the top of Google search results that shows a short passage taken from one web page to answer a query directly. It is the original target of AEO. Content written to win featured snippets, with a clear question and a short direct answer, is also well suited to being used in AI-generated answers.

People Also Ask (PAA)

People Also Ask is a Google search feature that shows a list of related questions, each of which expands to reveal a short answer from a web page. For AEO, PAA questions are a useful source of real questions that people ask about a topic.

Voice Assistant Answers

Voice assistant answers are spoken responses given by tools such as Siri, Alexa, and Google Assistant. Unlike a screen of results, a voice assistant usually gives one answer, which makes being that single answer the goal. Voice search was one of the earliest drivers of AEO.

Section 3: How AI Search Works

Large Language Model (LLM)

LLM stands for Large Language Model. It is an AI model trained on very large amounts of text to understand and generate human language by predicting which words are likely to come next. GPT, Gemini, Claude, and Llama are families of LLMs, and they are the systems that write the answers in AI search.

Retrieval-Augmented Generation (RAG)

RAG stands for Retrieval-Augmented Generation. It is a method where an AI system first retrieves relevant documents or passages from an outside source, such as a search index, and then uses them to write its answer. Most AI search products rely on some form of RAG to give current answers and cite sources.

How It Works

• Retrieve: The system searches an index for passages related to the question.
• Augment: The most relevant passages are added to the information the model works with.
• Generate: The model writes an answer based on those passages, often linking to where they came from.

For GEO, RAG is the main reason a page can be cited even if the model never saw it during training.

Grounding

Grounding is the practice of tying an AI model’s answer to specific, verifiable source material rather than relying only on what it learned in training. In AI search, this usually means retrieving web pages at the time of the question and basing the answer on them. Grounding reduces errors but does not eliminate them.

Grounding vs RAG

The terms are closely related and often used together. RAG describes the technical method of retrieving and then generating. Grounding describes the goal, which is an answer anchored to real sources. Google uses the phrase “grounding with Google Search” for this capability in its AI products.

Inference-Time Retrieval

Inference-time retrieval is when an AI system fetches information at the moment it answers a question, rather than relying only on its training data. “Inference” is the technical term for a model producing an answer. This is how AI search tools can discuss events and pages published after their training ended.

Query Fan-Out

Query fan-out is a technique where an AI search system breaks one question into several related sub-questions and searches for each of them separately. The system then combines what it finds into a single answer. Google has said its AI Mode uses this approach for complex questions.

How It Works

• Breaking down the question: The system identifies the different parts of the question. A query like “best CRM for a small law firm” might become sub-queries about pricing, legal-specific features, and user reviews.
• Running the sub-queries: Each sub-query is searched separately, often at the same time.
• Combining the results: The system selects the most relevant passages from all the results and writes one answer from them.

Because the sub-queries are generated by the system, a page can be retrieved for a question the user never typed.

Passage Retrieval

Passage retrieval is when a search or AI system finds and uses a specific section of a page instead of treating the whole page as one unit. Most AI search systems work at this level. As a result, one clear, relevant paragraph can be used in an answer even if the rest of the page covers other topics.

Chunking

Chunking is the process of splitting a long document into smaller pieces, called chunks, so an AI system can store, search, and retrieve them separately. The term describes what the AI system does to content, and it is also used to describe writing content in self-contained sections that survive being split up.

Types of Chunking

• Fixed-size chunking: Text is cut into pieces of a set length, which can split a sentence or idea in half.
• Structure-based chunking: Text is split at natural boundaries such as headings or paragraphs.
• Semantic chunking: Text is split where the topic changes, based on meaning.

Publishers cannot see or control which method an AI system uses. Writing each section so it makes sense on its own helps regardless of the method.

Vector Embedding

A vector embedding is a list of numbers that represents the meaning of a piece of text. Texts with similar meanings get similar numbers, even when they use different words. AI search systems convert both questions and content into embeddings so they can find passages that match what a user means, not just the words they typed.

Vector Search

Vector search is a way of finding information by comparing vector embeddings to see which pieces of content are closest in meaning to a query. It is often run on a vector database, which is built to store and compare embeddings quickly. Many AI search systems combine vector search with traditional keyword search.

Reranking

Reranking is a second step in retrieval where an initial set of results is sorted again using a more precise model. The first pass quickly gathers many possible matches, and the reranker then judges which ones are most relevant to the question. Only the top results after reranking are usually passed to the model to write the answer.

Semantic Search

Semantic search is search that matches the meaning and intent behind a query rather than only its exact words. It allows a page about “reducing restaurant food waste” to be found for a question like “how can my kitchen throw away less food.” AI search systems rely heavily on semantic search, often through vector embeddings.

Context Window

A context window is the maximum amount of text an AI model can take into account at one time, measured in tokens. It includes the user’s question, any retrieved content, and the conversation so far. Context windows have grown considerably, but systems still have to choose which retrieved passages to include.

Token

A token is the basic unit of text an AI model processes, usually a whole word, part of a word, or a punctuation mark. Model limits and costs are counted in tokens rather than words. In English, a token is on average a little shorter than a word.

Token Efficiency

Token efficiency is an informal term for how much useful information a piece of content delivers relative to its length. The idea is that concise, information-dense passages are easier for AI systems to retrieve and use than padded ones. The term has no standard definition, and there is limited public evidence on how much it affects citation.

Knowledge Cutoff

A knowledge cutoff is the date after which an AI model’s training data ends. The model has no built-in knowledge of events or changes after that date. It can only discuss newer information if it retrieves it from the web at the time of the question.

Latent Knowledge

Latent knowledge is the information an AI model has absorbed into its internal parameters during training and can recall without searching. It is why a model can often describe well-known brands accurately even with web search turned off. Brands with little coverage in training data tend to have weaker or less accurate latent knowledge.

Latent Knowledge vs Retrieved Knowledge

Latent knowledge comes from training and can be outdated. Retrieved knowledge comes from pages fetched at the time of the question and can be current. Many AI answers mix both.

Training Data Influence

Training data influence is the degree to which a particular source or piece of content shaped what an AI model learned during training. Widely published and frequently referenced information tends to have more influence than content that appears in only one place. Publishers cannot see or measure this influence directly.

Hallucination

A hallucination is an AI-generated statement that sounds plausible but is false or unsupported. For a brand, this could be wrong opening hours, a product that does not exist, or an incorrect price. Hallucinations can happen even when an AI system retrieves accurate sources, although grounding tends to reduce them.

Section 4: Entities & Authority Signals

Entity

An entity is a distinct, identifiable thing, such as a person, business, product, place, or concept, that search engines and AI systems can recognize and connect to other things. A restaurant, its head chef, and the neighborhood it is in are all separate entities with relationships between them. Search and AI systems increasingly organize information around entities rather than keywords alone.

Named Entity Recognition (NER)

NER stands for Named Entity Recognition. It is a natural language processing technique that finds names in text and sorts them into categories such as people, organizations, products, and locations. When a brand name is spelled and used consistently, it is easier for these systems to identify it correctly and not confuse it with something else.

Entity Salience

Entity salience is a measure of how central an entity is to a piece of text. A brand that is the main subject of an article has high salience, while a brand mentioned once in passing has low salience. Google’s Natural Language API reports a salience score, although how any AI search system weighs salience is not publicly documented.

Entity Optimization

Entity optimization is the practice of making sure a brand and its related people, products, and locations are described accurately and consistently across the web. The goal is to help search engines and AI systems identify the brand clearly and connect it to the right facts.

Core Concepts

• Consistent naming: The same name, address, and core details appear across the website, business profiles, and directories.
• Structured data: Schema markup on the website states key facts in a machine-readable way.
• Third-party presence: Profiles on trusted sources, such as Wikidata, industry directories, and review platforms, confirm the brand exists and what it does.
• Clear relationships: Content makes explicit how the brand relates to its people, products, and locations.

Knowledge Graph

A knowledge graph is a database that stores entities and the relationships between them. Google’s Knowledge Graph powers knowledge panels in search results, and public sources such as Wikidata work in a similar way. Being represented accurately in these sources helps search and AI systems describe a brand with the correct facts.

Earned Media

Earned media is coverage of a brand that the brand did not publish or pay for, such as news articles, independent reviews, and mentions in industry roundups. In GEO, it matters because AI systems often draw on third-party sources when describing or recommending brands. Some studies have reported that AI search answers cite earned media more often than brand-owned pages, although results vary by platform and query type.

Source Diversity

Source diversity is the range of different websites and types of publication that mention or cite a brand. A brand discussed across news sites, review platforms, forums, and industry publications has higher source diversity than one mentioned mainly on its own website. Broad, independent coverage is generally treated as a stronger signal of credibility than repeated mentions from a single source.

Corroboration

Corroboration is when several independent sources state the same fact or claim. AI systems that combine information from multiple pages are more likely to repeat a claim that many sources agree on. A claim found in only one place, or contradicted elsewhere, is less likely to be presented as fact.

Co-Citation

Co-citation is when two entities are mentioned together by the same third-party source. If many articles about Toronto brunch spots mention the same cafe, AI systems may learn to associate that cafe with the topic. The term comes from academic citation analysis and is used in SEO to describe this kind of association.

Co-Citation vs Co-Occurrence

Co-citation usually refers to two entities being referenced by the same source. Co-occurrence usually refers to two words or entities appearing near each other in text. The terms are often used interchangeably in SEO writing.

Information Gain

Information gain is the amount of new, useful information a piece of content adds compared with what is already available on the topic. Original research, first-hand experience, and unique data all add information gain. In SEO and GEO, the term is used informally, and there is no public way to measure how any search or AI system scores it.

Big Brand Bias

Big brand bias is the tendency of AI systems to mention or recommend large, well-known brands more often than smaller ones. It likely reflects how much more often big brands appear in training data and on the web. For smaller businesses, it means that being specific, for example by location or specialty, can matter more than competing on broad queries.

Recency Bias

Recency bias is the tendency of AI search systems to favor recently published or updated content, especially for questions where current information matters. It is most noticeable for topics like prices, opening hours, events, and news. Updating a page’s date without changing its content is not the same as keeping it current.

Position Bias

Position bias is the tendency of AI models to give more weight to information based on where it appears, such as at the start of a list or document. Research has found that models can overlook information placed in the middle of long inputs. For measurement, it also means that a brand listed first in an AI answer is not necessarily the one the model prefers most.

Section 5: Content for AI Citation

Answer-First Content

Answer-first content is writing that states the direct answer to a question at the start of a page or section, then adds detail afterward. It mirrors how journalists put the most important information first. For AI systems that pull individual passages, a clear answer near the top of a section is easier to find and use.

Q&A Structuring

Q&A structuring is the practice of organizing content as questions and answers, usually with the question as a heading and a direct answer in the first sentence below it. It matches how people phrase questions to AI assistants and makes each answer easy to locate. Questions work best when they reflect how real users ask them, for example as seen in People Also Ask boxes or customer enquiries.

Direct Answer

A direct answer is a short, self-contained statement that fully answers a specific question without needing surrounding context. A sentence like “Most Toronto patios open in late April” works as a direct answer, while “It depends, as mentioned above” does not. AI systems often use direct answers word for word or close to it, sometimes without the text around them.

Answer Density

Answer density is an informal term for how much of a page is made up of clear, usable answers compared with introductions, filler, or repetition. A page with high answer density reaches its main points quickly. There is no standard way to measure it, and it describes a writing goal rather than a known ranking factor.

Answer Fidelity

Answer fidelity is an informal term for how precisely a piece of content answers the question it targets, without drifting off topic or leaving out key parts. A page with high answer fidelity gives a complete, accurate answer and stops. The term is used by some GEO practitioners and does not have a widely agreed definition.

Citation Anchor

A citation anchor is an informal term for a specific, quotable piece of information that AI systems repeatedly pick up and cite, such as an original statistic, a named framework, or a clear definition. Because it is distinctive and easy to attribute, it can keep earning mentions over time. The term is used by some practitioners and is not an industry standard.

Semantic Coherence

Semantic coherence is the degree to which every part of a page stays focused on the same topic. A page that covers one subject clearly is easier for AI systems to match to relevant questions than one that mixes unrelated topics. It is a general content quality idea rather than a specific, measurable signal.

Topical Completeness

Topical completeness is how thoroughly a piece of content covers the different parts of a subject. A complete guide to opening a cafe might cover permits, equipment, staffing, and costs, rather than only one of those. Because AI systems often answer multi-part questions, content that covers the related sub-questions has more chances to be relevant.

Topical Completeness vs Topical Authority

Topical completeness describes one page or piece of content. Topical authority describes a whole website’s depth and credibility on a subject, built across many related pages.

Answer Graph

An answer graph is an informal term for a connected set of pages on a website that together answer the main questions on a topic, such as definitions, comparisons, how-to guides, and FAQs, all linked to each other. The idea is that for most questions in that topic, one of the pages provides a strong answer. The term is used by some agencies and does not have a standard definition.

Epistemic Precision

Epistemic precision is the practice of stating claims with the right level of certainty, no more and no less than the evidence supports. “Most customers book 2 to 3 days ahead” is more precise than “Customers always book early.” Accurate, well-qualified claims are generally more trustworthy, although there is no public evidence that any AI system scores content on this specifically.

Conversational Query

A conversational query is a search question written or spoken in natural, full-sentence language rather than short keywords. “Where can I get good gluten-free pasta near Union Station?” is conversational, while “gluten free pasta Toronto” is not. People tend to ask AI assistants conversational queries, which are often longer and more specific than typical search engine queries.

Multi-Turn Conversation

A multi-turn conversation is an exchange with an AI assistant that includes several back-and-forth messages on the same topic. Each new answer can take earlier questions and answers into account. For GEO, it means a brand can be introduced in an early answer and then appear again as the user asks follow-up questions.

Section 6: Technical GEO & AI Crawling

AI Crawler

An AI crawler is an automated bot that an AI company uses to fetch web pages, either to collect training data or to retrieve content for answers. Website owners can allow or block most AI crawlers using the robots.txt file. Blocking a crawler can keep content out of a model’s training data, but it can also stop that content from being cited.

Types of AI Crawlers

• Training crawlers: Collect web content to train future AI models. Blocking them affects what future models learn, not live answers.
• Search crawlers: Build an index that an AI search product uses to find and cite pages in its answers.
• User-triggered fetchers: Visit a specific page when a user asks the AI to look something up or open a link. Some providers state that robots.txt rules may not apply to these, because the visit is made on a user’s behalf.

Most major AI companies now run separate bots for these jobs, so each one can be allowed or blocked independently.

OpenAI Crawlers

OpenAI’s crawlers are the bots OpenAI uses to access web content for its models and for ChatGPT. It publishes separate user agents for different purposes.

• GPTBot: Collects content that may be used to train OpenAI’s models.
• OAI-SearchBot: Finds and indexes pages so they can appear as sources in ChatGPT search results.
• ChatGPT-User: Visits pages when a ChatGPT user’s request requires it.

A site can block GPTBot to opt out of training while still allowing OAI-SearchBot so its pages can be cited in ChatGPT search.

Anthropic Crawlers

Anthropic’s crawlers are the bots Anthropic uses to access web content for its Claude models and assistant. Like OpenAI, it separates them by purpose.

• ClaudeBot: Collects content that may be used to train Claude models.
• Claude-SearchBot: Crawls pages to improve the quality of search results in Claude.
• Claude-User: Fetches pages when a Claude user asks a question that requires it.

Perplexity Crawlers

Perplexity’s crawlers are the bots Perplexity uses to find and fetch web pages for its answer engine.

• PerplexityBot: Crawls and indexes pages so they can appear in Perplexity’s search results. Perplexity states it is not used to train AI models.
• Perplexity-User: Fetches pages in response to a user’s question. Perplexity states that this fetcher generally ignores robots.txt, since the request comes from a user.

Google-Extended

Google-Extended is a robots.txt control token that lets site owners decide whether content Google has already crawled can be used to train its Gemini models and to ground answers in Gemini apps and some developer products. It is not a separate crawler, because Google’s regular crawlers do the actual fetching. Blocking Google-Extended does not remove a site from Google Search, and Google has said it does not affect inclusion in AI Overviews.

Google-Extended vs Googlebot

Googlebot crawls pages for Google Search, which includes AI Overviews and AI Mode. Google-Extended only controls a specific set of other AI uses. Many glossaries describe Google-Extended as a crawler, which is inaccurate.

Applebot-Extended

Applebot-Extended is a robots.txt control token that lets site owners opt out of having their content used to train Apple’s foundation models. Like Google-Extended, it does not crawl pages itself. Apple’s regular crawler, Applebot, still fetches pages for features such as Siri and Spotlight.

llms.txt

llms.txt is a proposed standard for a plain-text Markdown file, placed at the root of a website, that gives AI systems a short summary of the site and links to its most important pages. It was proposed in 2024 by Jeremy Howard of Answer.AI. Adoption by website owners has grown, but major AI search providers have not confirmed that they rely on it, and its effect on citations is unproven.

llms.txt vs robots.txt

Robots.txt tells crawlers which pages they may or may not access. llms.txt does not control access. It is only a guide that suggests which content is most useful.

llms-full.txt

llms-full.txt is a companion file to llms.txt that contains the full text of a site’s key pages in a single Markdown document, rather than just links to them. It is most common on software documentation sites. Like llms.txt, it is an informal convention, and there is no confirmed evidence that AI search systems use it.

AI Indexability

AI indexability is whether an AI system’s crawlers can access, read, and understand a page’s content. A page may be blocked by robots.txt, hidden behind a login, or built in a way that crawlers cannot read. If a page is not indexable, it cannot be retrieved or cited.

Common Barriers

• Blocked crawlers: Robots.txt rules, firewalls, or bot protection that stop AI crawlers.
• JavaScript-only content: Content that only appears after JavaScript runs in the browser. Many AI crawlers do not run JavaScript, so server-side rendering, where the server sends complete HTML, is the more reliable approach.
• Content in images or PDFs: Text inside images, or in files that are hard to parse, may be missed.

Structured Data for AI

Structured data is code added to a web page, usually in the JSON-LD format using the Schema.org vocabulary, that labels information in a machine-readable way. It can state facts such as a business’s name, address, opening hours, menu, or reviews explicitly. Search engines use it for rich results, and it can help AI systems read key facts accurately, although no AI platform has confirmed exactly how it uses structured data when writing answers.

Section 7: Measurement & Tracking

AI Visibility Score

An AI visibility score is a single number that summarizes how visible a brand is across AI-generated answers for a chosen set of prompts. It usually combines several signals, such as how often the brand is mentioned, how often it is cited, where it appears in the answer, and the sentiment of the mention. There is no standard formula, and each GEO tool calculates it differently.

How to Interpret It

The score is useful for tracking change over time within one tool. It is not reliable for comparing results between different tools, since each uses its own prompts, platforms, and weightings. Looking at the underlying metrics shows what is actually driving a rise or fall.

Share of Model (SoM)

Share of Model is the percentage of AI-generated answers, within a defined set of prompts, that mention a particular brand compared with its competitors. It is the AI search version of share of voice. Some tools call it AI share of voice, and methods for calculating it vary.

How It’s Calculated

One common method divides the number of answers mentioning the brand by the total number of answers checked. Another divides the brand’s mentions by the total mentions of all tracked brands. Because the two methods give different results, it is worth confirming which one a report uses.

Citation Rate

Citation rate is the percentage of AI-generated answers, within a chosen set of prompts, that cite a specific brand, domain, or page as a source. It shows how often an AI system credits your content when answering questions relevant to your business.

How It’s Calculated

The basic formula divides the number of answers that cite you by the total number of answers checked, then multiplies by 100. If your domain is cited in 12 out of 50 answers, your citation rate is 24%. There is no standard method, so tools differ in whether they count at the domain or page level, and in which prompts and platforms they include.

How to Interpret It

AI answers can change each time the same prompt is run, so a single check is not reliable. Citation rate is more meaningful when measured across repeated runs and tracked over time. It is also best read per platform, since ChatGPT, Perplexity, and Google cite sources differently.

Brand Mention Frequency

Brand mention frequency is how often a brand is named across a set of AI-generated answers. It counts mentions whether or not the answer links to the brand’s website. It is often the first metric tracked, since it shows whether AI systems bring up a brand at all.

AI Mention vs AI Citation

An AI mention is when an AI-generated answer names a brand. An AI citation is when the answer links to, or credits, a specific web page as a source. The two are measured separately because they often differ.

Key Differences

• Mention without citation: The AI names a brand but cites other sources, such as a review site or news article. This builds awareness but sends no traffic to the brand.
• Citation without mention: The AI uses a brand’s page as a source but does not name the brand in the answer text. This can send traffic without building brand recall.
• Both: The brand is named and its page is cited, which is the strongest outcome.

Inline Citation

An inline citation is a link or source marker placed directly within an AI-generated answer, next to the claim it supports. Perplexity uses numbered inline citations, and other platforms show sources in different ways, such as link cards or a source list. Inline citations are the main way AI answers send referral traffic to websites.

Query Coverage

Query coverage is the share of a defined set of relevant prompts for which a brand appears in AI-generated answers. It measures breadth rather than depth. A brand with high query coverage shows up across many different questions in its category, not just a few.

Competitive Citation Gap

A competitive citation gap is the difference between how often a brand is cited and how often a competitor is cited for the same set of prompts. It highlights the specific questions where competitors are credited as sources and the brand is not. The term is informal and is calculated differently by different tools.

AI Brand Sentiment

AI brand sentiment is the tone in which AI systems describe a brand when they mention it, usually classified as positive, neutral, or negative. Being described as “a popular spot for weekend brunch” is different from being described as “often criticized for slow service.” Sentiment is usually scored automatically by another AI model, which can make its own classification mistakes.

Answer Accuracy

Answer accuracy is how correctly AI-generated answers state verifiable facts about a brand, such as its prices, locations, opening hours, or products. It is measured by comparing AI answers against a list of confirmed facts. A brand can be mentioned often and still be described inaccurately, which is why accuracy is tracked separately from visibility.

Hallucination Rate

Hallucination rate is the share of AI-generated statements or answers that turn out to be factually wrong when checked against a verified source. In GEO, it is used to track how often AI systems get facts about a brand wrong, such as its pricing, products, or leadership.

How It’s Calculated

The rate divides the number of incorrect answers or claims by the total number checked, then multiplies by 100. If 8 out of 40 AI answers about your business contain a false detail, the hallucination rate is 20%. There is no standard method, so results vary depending on whether errors are counted per answer or per individual claim, and on which prompts and platforms are tested.

How to Interpret It

A lower rate means AI systems describe the business more accurately, but a single test can be misleading because AI answers change between runs. The rate is more useful when measured repeatedly over time and broken down by platform and topic, which shows where the errors cluster. Checking is often done by a person or by a second AI model acting as a judge, and automated checks can make mistakes of their own.

Hallucination Rate vs Answer Accuracy

The two metrics measure the same problem from opposite directions. Answer accuracy reports the share of facts an AI system gets right, while hallucination rate reports the share it gets wrong or makes up. Some teams track both, since a missing fact lowers accuracy without counting as a hallucination.

Non-Determinism

Non-determinism is the property of AI systems that the same prompt can produce different answers each time it is asked. The wording, the brands mentioned, and the sources cited can all change between runs. This is why GEO measurement relies on running prompts multiple times and looking at averages rather than single results.

Prompt Tracking

Prompt tracking is the practice of regularly running a fixed set of prompts through AI platforms and recording the answers to monitor a brand’s visibility over time. The fixed set is often called a prompt set. Keeping the prompts consistent is what makes results comparable from one period to the next.

Core Concepts

• Prompt selection: Prompts should reflect questions real customers ask, across informational, comparison, and buying-intent topics.
• Repetition: Each prompt is run several times to account for non-determinism.
• Platform coverage: Results are recorded separately for each AI platform.
• Consistent settings: Location, language, and whether the user is logged in can all change answers, so they are kept the same where possible.

Prompt Volume

Prompt volume is an estimate of how often people ask AI assistants a particular question or type of question. It is the AI search counterpart to keyword search volume. AI platforms do not publish this data, so all prompt volume figures are estimates built by third-party tools from panels, clickstream data, or modelling, and they should be treated as directional.

AI Referral Traffic

AI referral traffic is website visits that come from users clicking links in AI-generated answers. It can often be identified in analytics tools by referrer domains such as chatgpt.com or perplexity.ai. Some AI traffic arrives without a referrer and is recorded as direct traffic, so reported figures usually undercount the true amount.

Zero-Click Search

A zero-click search is a search that ends without the user clicking through to any website, because the answer was shown directly in the results or the AI response. AI Overviews and AI assistants tend to increase zero-click behavior. As a result, visibility and brand mentions in answers matter alongside traffic when judging performance.

LLM-as-a-Judge

LLM-as-a-judge is a method where one AI model is used to evaluate the answers produced by another, for example to score accuracy, sentiment, or whether a brand is mentioned. It makes it practical to review thousands of AI answers that would take too long to check by hand. Judge models have known weaknesses, such as favoring longer answers, so their results are more reliable when spot-checked by people.

Section 8: Agentic Search & Emerging Terms

AI Agent

An AI agent is an AI system that can carry out multi-step tasks on a user’s behalf, such as searching, comparing options, filling in forms, or making a booking, rather than only answering a question. Agents decide which steps to take and use tools like web browsers or connected apps to complete them. Most consumer agents are still early, and many require the user to approve important actions.

AI Agent vs AI Assistant

An AI assistant mainly responds to questions within a conversation. An AI agent takes actions to complete a goal. Many products now do both, so the line between them is increasingly blurred.

Agentic Search

Agentic search is a form of AI search where an AI agent plans and carries out several searches on its own, reviews the results, and decides what to look up next before giving a final answer. It is used for research-style questions that one search cannot answer. Features often labelled deep research are a common example.

Agentic Search vs Query Fan-Out

Query fan-out runs a batch of related searches at once and combines the results. Agentic search works in steps, where what the agent finds in one search shapes the next.

Agentic AEO

Agentic AEO is a newer term with two different meanings. Some vendors use it to mean using AI agents to automate AEO work, such as monitoring AI visibility and suggesting fixes. Others use it to mean optimizing content and websites so that AI agents can find, understand, and act on them. The term has no widely agreed definition, so its meaning depends on who is using it.

Model Context Protocol (MCP)

MCP stands for Model Context Protocol. It is an open standard, introduced by Anthropic in 2024, that gives AI applications a common way to connect to outside tools and data sources. Instead of building a custom connection for every service, a developer can connect an AI app to any system that supports MCP.

How It Works

• MCP server: A service that offers tools or data, such as a booking system, a product catalog, or a database.
• MCP client: The part of an AI application that connects to MCP servers and passes requests between them and the model.

For businesses, MCP is one way AI agents may eventually interact with services directly, for example to check availability or place an order, rather than reading a web page.

NLWeb

NLWeb is an open project introduced by Microsoft in 2025 that helps websites offer a natural-language interface, so people and AI agents can ask questions of a site’s content directly. It builds on data many sites already publish, such as Schema.org markup and RSS feeds. Each NLWeb site can also act as an MCP server, which makes its content available to AI agents. Adoption is still early.

Agentic Commerce

Agentic commerce is shopping in which an AI agent helps a user research, choose, and buy products or services, sometimes completing the purchase inside an AI assistant. Several protocols have been proposed to make this possible, including OpenAI’s Agentic Commerce Protocol and Google’s Agent Payments Protocol, both announced in 2025. The standards and their adoption are still developing.

Prompt Injection

Prompt injection is an attack in which hidden or disguised instructions are placed in content that an AI system reads, in an attempt to change what the AI says or does. For example, a web page could contain hidden text telling an AI to recommend one business over others. AI companies actively defend against prompt injection, and attempting it to manipulate AI answers is a form of spam that can backfire.

Frequently Asked Questions

Is AEO the same as GEO?

AEO and GEO overlap heavily, and many people use the terms interchangeably. When they are distinguished, AEO focuses on making individual answers on a page easy to extract, while GEO covers a brand’s wider presence across AI-generated answers, including third-party mentions and citation tracking.

Does GEO replace SEO?

No. Many AI search systems retrieve content from search indexes, so pages that are crawlable, well structured, and trusted in traditional search are also better placed to be used in AI answers. GEO builds on SEO rather than replacing it.

Which AI crawlers should I allow in robots.txt?

It depends on your goals. Allowing search crawlers such as OAI-SearchBot, Claude-SearchBot, and PerplexityBot lets your pages be found and cited in those AI search products. Training crawlers such as GPTBot and ClaudeBot can be blocked separately if you do not want your content used to train future models.

How do you measure GEO performance?

GEO performance is usually measured by running a fixed set of relevant prompts across AI platforms, repeating them over time, and tracking metrics such as citation rate, share of model, brand mention frequency, and answer accuracy. AI referral traffic in analytics adds a view of the visits these answers send.

Does llms.txt help with AI visibility?

There is no confirmed evidence that it does. Major AI search providers have not stated that they rely on llms.txt, so it is best treated as a low-cost experiment rather than a core tactic.

Can any technique guarantee a citation in AI answers?

No. AI systems decide what to cite based on many factors that publishers cannot see or control, and answers can change each time a question is asked. Good practice improves the chances of being used and described accurately, but it cannot guarantee a result.

Final Thoughts

The language of AI search is still settling, and several terms in this glossary will likely shift in meaning as platforms change. The fundamentals are more stable: content that crawlers can access, answers that stand on their own, accurate information about your business across the web, and consistent measurement over time. This glossary is reviewed and updated as the field develops.