This is a glossary of the thirty terms that recur in AEO (Answer Engine Optimization) and GEO (Generative Engine Optimization) work, grouped into core terms, mechanics, infrastructure, and measurement. Every definition is written to stand on its own in one to three sentences — that is the point: a definition that cannot be quoted out of context has not done its job. One warning before you read: both acronyms collide with other fields in search. Bare “GEO” most often means geography, and bare “AEO” is ambiguous enough that autocomplete will not commit — which is why the longer forms appear everywhere in serious writing. The full series these terms belong to starts on the AEO/GEO basics hub page.
Core terms
AEO (Answer Engine Optimization)
AEO is the practice of making your content likely to be selected and cited when an AI system generates an answer to a question. Where SEO competes for a slot in a ranked results list, AEO competes for a sentence — or at least a named mention — inside the answer itself. The term predates the LLM era; its full history is in what is AEO.
GEO (Generative Engine Optimization)
GEO is the research term for improving how often and how favorably your content appears in the responses of a generative engine, introduced in the 2023 paper “GEO: Generative Engine Optimization” (Aggarwal et al., arXiv:2311.09735, accepted to KDD 2024). That single study found passage-level changes can lift response visibility by up to 40% in its benchmark — a real direction, one study, an upper bound. The industry now uses “GEO” loosely as the umbrella name for all AI-visibility work; the distinction is covered in what is GEO.
Generative engine
A generative engine is an AI system that synthesizes information from multiple sources and summarizes it into a generated response, rather than returning a ranked list of links. ChatGPT, Perplexity, Gemini and Claude answer products, and Google’s AI Overviews, are all instances of generative engines.
Answer engine
An answer engine is a system whose output is a direct answer to a question, optionally with supporting citations. The term covers both LLM-based surfaces and their pre-LLM predecessor, the featured-snippet box, which answered queries directly atop the classic results list.
Generative search engine
A generative search engine is a search product whose results are composed prose with citations, rather than (or alongside) a blue-link list. Perplexity is the canonical example: it searches the live web in real time and answers with numbered citations to the original sources.
Mechanics
RAG (Retrieval-Augmented Generation)
RAG is the pattern of retrieving external information at query time and feeding it to the model as grounding for a generated answer. It is the mechanism that lets an AI answer cite pages that did not exist when the model was trained.
Retrieval
Retrieval is the stage of answer generation in which candidate sources are gathered from the model’s training knowledge, from live web search, or both. A page must first be retrieved to be citable — but retrieval does not guarantee citation.
Grounding
Grounding is the use of retrieved external material as the factual basis for a generated answer. Grounding crawlers are the fetchers that build the index a surface grounds live answers on, such as OAI-SearchBot and PerplexityBot.
Citation
A citation is a source reference that a generated answer attaches to a specific claim, usually linking to a URL. Citations are attached to claims, and claims bind to named entities — which is why consistent naming and corroboration matter more than owning a single perfect page.
Retrieval visibility vs citation visibility
Retrieval visibility is how often your content is fetched as a candidate; citation visibility is how often it is actually cited in the final answer. The two are related but distinct — an observational 2026 ChatGPT study (arXiv:2610.15382) found candidates retrieved more often are more likely to be cited, and that external proxies recover only about a quarter of cited URLs, which is why the distinction matters.
Knowledge cutoff
A knowledge cutoff is the date after which a model’s training data stops, meaning its internal knowledge of anything newer is unreliable. It is why AI surfaces search the live web for current questions — and why your stale pages compete against your fresh ones.
Hallucination
A hallucination is a confident, plausible claim in a generated answer that is not supported by any retrieved or trained source. It is the failure mode that made citation and corroboration part of the answer contract: an engine that cannot ground a claim may still assert it.
Infrastructure
GPTBot
GPTBot is OpenAI’s training crawler: its documentation says it “is used to crawl content that may be used in training” OpenAI’s foundation models. Disallowing it in robots.txt opts your content out of training use, with no documented effect on ChatGPT search visibility.
OAI-SearchBot
OAI-SearchBot is OpenAI’s grounding crawler for ChatGPT search, and the bot whose access determines whether your site can appear in ChatGPT search answers. Opting out in robots.txt means you “will not be shown in ChatGPT search answers,” though brand searches may still produce a navigational link; changes take about 24 hours.
ClaudeBot
ClaudeBot is Anthropic’s training crawler, which collects web content that “could potentially contribute to” model training. It is distinct from Claude-SearchBot (grounding) and Claude-User (user-directed fetch) — blocking ClaudeBot does not remove you from Claude search results.
PerplexityBot
PerplexityBot is Perplexity’s grounding crawler, explicitly “not used to crawl content for AI foundation models” per Perplexity’s documentation. Blocked pages may still be indexed as domain, headline, and brief factual summary, and changes take up to 24 hours.
Google-Extended
Google-Extended is not a crawler at all but a robots.txt token that controls whether content collected by Googlebot may train future versions of Gemini models. Google states it “doesn’t affect a site’s inclusion in Search, nor do we use Google-Extended as a ranking signal.”
llms.txt
llms.txt is a proposed web convention — a markdown file, usually at your site’s root, giving LLM agents a concise machine-readable index of your content with links to per-page markdown versions. The v2 spec at llmstxt.org (modified August 10, 2026) makes no ranking or citation claim; it is a courtesy index for the agents that fetch it.
llms-full.txt
llms-full.txt is a v1-era convention in which a single file concatenates your site’s full text content for agent consumption. It does not appear in the llms.txt v2 spec, so it is optional legacy practice rather than a requirement.
robots.txt
robots.txt is the plain-text file at a site’s root in which a site declares which paths crawlers may and may not fetch; the file is client-side, meaning the crawler reads it and decides, and a block produces no report or notification anywhere. Because it applies to every unnamed agent, a blanket User-agent: * / Disallow: / silently blocks AI crawlers you never named.
JSON-LD
JSON-LD (JavaScript Object Notation for Linked Data) is the standard format for embedding machine-readable metadata about a page — Organization, Product, FAQ and other schema.org types — directly in the page’s HTML. It is the structured-data layer that helps machines bind claims to a named entity with consistent attributes.
Entity resolution
Entity resolution is the process by which an AI system binds a claim in an answer to a specific named entity — your product or company — from the web evidence around it. A 2026 observational ChatGPT study (arXiv:2610.15382) found that external proxy signals (Google Search results and OpenAI Responses API candidates) matched cited URLs better at the hostname and entity levels than at the exact-URL level, which is why consistency of naming, facts and URLs across the web is an infrastructure problem, not a copywriting one.
Measurement
Citation rate
Citation rate is the share of a defined set of tracked questions on which your brand is cited by a given surface. It is the core AEO number because it measures the actual unit of success — a sentence inside an answer — rather than rankings or impressions.
Share of voice
Share of voice in AI answers is your brand’s share of the citations or mentions that appear for a query set, measured against a named competitor set. In a cited ChatGPT conversation — roughly six unique sources on average, per Profound’s ~700,000-conversation study — you are competing for one slot in a small set, which is what makes the denominator matter.
Zero-click
Zero-click describes a search interaction where the user’s query is satisfied without clicking any result, a behavior AI answers intensify by surfacing the answer directly. For a B2B SaaS site it reframes the question from “do we rank” to “are we in the answer, or is the answer complete without us.”
AI Overviews
AI Overviews are Google’s AI-generated, cited summaries shown in Search for qualifying queries; Google states they are shown only when “additive to classic Search,” and that a supporting page must be indexed and eligible with a snippet — no special markup or AI-specific file. They are the AI-answer layer inside the classic results page, distinct from Gemini’s chat answers.
Featured snippet
A featured snippet is the pre-LLM answer box Google placed atop the results list for direct-answer queries — the original answer engine, and the predecessor that shaped the first era of AEO before LLM surfaces arrived.
AI referral traffic
AI referral traffic is the site traffic in analytics that arrives from AI answer surfaces (ChatGPT, Perplexity, Gemini, Claude), visible as referral rows in GA4 with the surface’s domain as the source. It is the click layer of AI visibility — real, measurable, and smaller than the cited-but-not-clicked population.
Generative AI impressions
Generative AI impressions are the views of your pages in Google’s generative AI surfaces as reported in Google Search Console’s generative-AI performance report (released June 2026). An impression is not a citation and not a click; the report measures a third quantity that sits between the two.
Surface
A surface is a distinct AI answer product — ChatGPT, Perplexity, Gemini, Claude, Google AI Overviews — each with its own crawlers, retrieval behavior, and citation patterns. Measuring across surfaces (and, for ChatGPT, across Instant and Thinking modes) is the minimum for a defensible visibility number, since one probe of one surface is one data point.
Where these definitions come from
Every fact-level claim above traces to a verified source: the vendor crawler documentation and Google docs compiled in our maintained LLM crawlers reference, the llmstxt.org v2 spec as audited in llms.txt in 2026, the Aggarwal et al. 2023 paper (arXiv:2311.09735), the 2026 ChatGPT citations study (arXiv:2610.15382, observational), and Profound’s 2026 conversation analysis. The definitions that are ours (citation rate, share of voice, the glossary’s groupings) are working definitions, not industry standards — use them, quote them, and note that we defined them.
The full series: what is AEO, what is GEO, and how AI answers are actually generated — the three posts this glossary supports, all indexed on the AEO/GEO basics hub page.