AI search optimization glossary: 36 GEO, AEO and AI visibility terms, explained

This glossary collects 36 terms from AI search optimization, the work of getting a brand into the answers of ChatGPT, Gemini, Perplexity and Google’s AI features. Each term gets a one- or two-sentence definition and, where the definition rests on a fact, a link to the source. The terms are in six groups: what the discipline is called, where AI answers appear, how the AI builds an answer, how to measure, bots and files, and what is off-limits.

Much of this vocabulary is still contested. When a term has more than one meaning, or nobody agrees on it, the definition says so instead of picking a winner. For the full background, start with what is GEO.

What the discipline is called

GEO (Generative Engine Optimization)

The work of getting a brand found, cited and recommended in AI answers. The term comes from the paper “GEO: Generative Engine Optimization”, first published in November 2023 and accepted at KDD 2024, which makes it the only name on this list with a documented academic origin.

AEO (Answer Engine Optimization)

In practice, another name for GEO. Google’s AI optimization guide defines AEO and GEO together, as names for the work of improving visibility in AI search experiences, and concludes that, for Search, both are still SEO.

LLMO (LLM Optimization) and LLM SEO

Two more names for the same work, used mostly when the focus is ChatGPT and other assistants rather than Google Search. Neither has an official definition or a documented origin.

AI SEO

The label people reach for when they come to the topic from SEO. It describes the same work as GEO and AEO, and it matches Google’s own reading: for Google, optimizing for AI search is still SEO. Not everyone in the field agrees it is only that; the debate is summarized in the GEO guide.

AI visibility

The share of answers, across a large set of questions, in which a brand appears. It is the number all the names above are trying to move. A single answer does not measure visibility, because the same question brings different brands each time. The method is in how to measure a brand’s AI visibility.

Where AI answers appear

LLM (large language model)

The system behind ChatGPT, Gemini and Claude. It learns patterns from a huge volume of text and writes the answer word by word, picking a likely next word at each step.

Generative engine (or answer engine)

A search system that, instead of returning a list of links, writes an answer from several sources. ChatGPT search, Perplexity, AI Overviews and AI Mode are all generative engines. It is the “G” in GEO and the “A” in AEO.

AI Overviews

The AI-generated summaries that show up on some Google searches. According to Google’s documentation, they help people get the gist of a topic quickly and act as a jumping-off point, with links to learn more.

AI Mode

Google’s conversational search mode, where people can ask follow-up questions. Google’s documentation says it is most helpful for queries that need further exploration, reasoning or complex comparisons, and that it returns a full AI-powered answer with links to supporting websites.

How the AI builds an answer

Training data

The text the model learned from, collected before it was released. It becomes the model’s “memory”, but it is frozen in time. That is why assistants add web search on top, and why there are separate bots for training and for search (see the bots section below).

RAG (Retrieval-Augmented Generation)

The model retrieves relevant documents before answering and writes the answer from them. The name comes from a 2020 paper by Patrick Lewis and co-authors, which combined the model’s own memory with a document index queried at answer time.

Grounding

Anchoring an answer in external sources consulted at question time, rather than relying only on the model’s memory. Google describes grounding with Google Search as the feature that connects Gemini to real-time web content. RAG is the technique; grounding is the outcome it is after.

Query fan-out

When the AI, instead of a single search, issues several related searches at once to gather more information. Google’s guide defines it as a set of concurrent, related queries generated by the model, and its AI features documentation says AI Overviews and AI Mode may use it, searching across subtopics and data sources. The practical consequence: the page that gets cited may answer a question the user never typed.

Temperature

The setting that controls how much risk the model takes when picking each word: higher temperature gives more varied answers, lower gives more predictable ones. Zero does not guarantee identical answers: in a Thinking Machines test, 1,000 identical requests at temperature zero produced 80 different completions.

Nondeterminism

The property of giving different answers to the same question. According to SparkToro, there is less than a 1 in 100 chance that ChatGPT or Google’s AI returns the same list of brands in two answers. The reasons are in why AI gives a different answer every time.

How to measure

Prompt

The question sent to the AI. For measurement, a prompt is written the way a customer would ask it, without the brand’s name, and is repeated many times to see how often each brand appears.

Mention

When the brand name appears in the answer text, with or without a link. It is what most AI visibility metrics count.

Citation

When the answer links to a page used as a source. Mentions and citations come apart: a brand can be recommended without its site being cited, and a site can be cited in an answer that recommends a competitor. Why that matters is covered in why citations are the new backlinks.

Presence (mention rate)

The share of answers in which the brand appears, for one question or a set of questions. It is the core AI visibility metric, because it settles as the sample grows: in the methodological study FA-MET-01, “How to Measure What an AI Recommends”, published by First Answer in April 2026, with 7,000 runs in the ChatGPT interface, going from 1,000 to 5,000 runs moved none of the top ten brands outside the predicted margin of error.

Position

The order in which the brand appears in the answer. It is far less stable than presence: in the SparkToro study, City of Hope hospital appeared in 69 of 71 ChatGPT answers but was mentioned first in only 25 of them.

Share of voice

The slice of a category’s mentions that goes to each brand, across a set of answers. It comes from the same count as presence and carries the same requirements: a large sample, a stated environment and a margin of error.

Margin of error and confidence interval

The margin of error says how far a sampled number may be from the true value; the confidence interval is the resulting range. With 1,000 answers, a brand measured at 50% sits, at 95% confidence, between roughly 47% and 53%, the margin of up to ±3.1 points worked out in FA-MET-01. Differences smaller than the margin are noise, not change.

API versus interface

The two ways to ask an AI the same question: through the API, the programmatic access developers use, or through the interface consumers use. They can give quite different results: in FA-MET-02, “API or Interface: That Is the Question”, from the same series, the same question asked 1,000 times in each environment significantly changed 12 of the 15 most relevant brands. So every measurement needs to say where it was taken; the step-by-step is in the measurement guide.

Earned media

What third parties publish about a brand: press, reviews, forums, videos. A study by Chen et al., from September 2025, found a “systematic and overwhelming” bias toward earned media in AI search, over brand-owned and social content.

Bots and files

AI crawler

A bot an AI company uses to visit pages. There are three jobs: collecting content to train models, indexing pages for the assistant’s search, and fetching a page on the spot when a user asks. OpenAI, Anthropic and Perplexity document separate bots for each job, which can be blocked one by one. Google has no separate AI crawler: it uses a robots.txt token, Google-Extended.

robots.txt

The file at the root of a site with access rules for bots. It is where each AI crawler is allowed or blocked by name. Search and training bots generally follow it; user-triggered fetchers do not always: OpenAI says robots.txt rules “may not apply” to ChatGPT-User, and Perplexity says Perplexity-User generally ignores them.

GPTBot

OpenAI’s crawler for content that may be used to train its generative AI models, according to OpenAI’s crawler documentation. Blocking GPTBot does not remove a site from ChatGPT search.

OAI-SearchBot

OpenAI’s bot for surfacing websites in ChatGPT’s search features. OpenAI points to this bot, not GPTBot, for managing search opt-outs in robots.txt. Each setting is independent: you can allow OAI-SearchBot and block GPTBot.

ChatGPT-User

The OpenAI agent that visits a page when a user asks ChatGPT or a custom GPT a question. According to OpenAI, it does not crawl the web automatically, and because it acts on a user’s request, robots.txt rules may not apply to it.

Google-Extended

A robots.txt token, not a separate crawler, that controls whether content Google crawls can be used to train future Gemini models and for grounding in the Gemini app and Vertex AI. According to Google’s documentation, it does not affect a site’s inclusion in Google Search and is not a ranking signal.

PerplexityBot

Perplexity’s crawler for surfacing and linking websites in its search results. According to Perplexity, it is not used to collect content for training AI foundation models.

ClaudeBot and Claude-SearchBot

Anthropic’s bots for Claude. According to Anthropic, ClaudeBot collects content that may go into training, Claude-SearchBot indexes pages to improve search results, and Claude-User fetches a page when a user asks. All three honor robots.txt and can be blocked separately.

llms.txt

A proposed Markdown file at the root of a site, with a summary and a list of links meant to help AI models and agents use the site. It was proposed by Jeremy Howard on September 3, 2024. Google’s guide says Search does not use these files, and keeping one neither helps nor hurts visibility.

Schema markup (structured data)

Code on a page that describes its content in a standard way (product, article, FAQ, organization), using the schema.org vocabulary founded by Google, Microsoft, Yahoo and Yandex. According to Google’s guide, structured data is not required for AI search and there is no special schema for it; it still matters for rich results.

What is off-limits, and what is asked for

AI answer spam (black hat GEO)

Trying to manipulate AI answers with tactics such as bought mentions, planted content or pages at scale. Since May 15, 2026, Google has made it explicit that its spam policies apply to generative AI responses in Search, and its spam policies page now lists attempts to manipulate those responses.

Non-commodity content

Google’s term for content that adds something beyond common knowledge: an expert’s or a practitioner’s take. Its AI optimization guide asks for this kind of content instead of tactics such as chopping text into tiny chunks for AI, which it says is not needed.

Where to go next

The full definition of GEO, where the term came from and what changes from SEO are in what is GEO. How to turn the measurement terms in this glossary into a number is in how to measure a brand’s AI visibility.