GEO glossary

Forty terms from AI search, defined short and straight. Each entry is written to make sense without having read the previous one, which is exactly how a model reads it when it retrieves a single fragment.

Engines and surfaces

GEO (Generative Engine Optimization)

The practice of optimising a brand so that generative engines mention and cite it inside the answers they write. Unlike SEO, it does not chase a position in a list of links but a place inside the answer text, and its unit of success is the citation rather than the click.See also:AEO (Answer Engine Optimization) ·Citation ·Generative engine

AEO (Answer Engine Optimization)

Optimising so a search engine returns your content as the direct answer: featured snippets, “people also ask” blocks and voice assistants. It predates GEO and shares its demand for short, self-contained answers, but it plays out inside the search engine rather than inside a model.See also:GEO (Generative Engine Optimization) ·AI Overview

Generative engine

A system that answers a question by writing new text instead of returning a list of links. It combines a language model with a retrieval layer that fetches sources in real time. ChatGPT, Gemini, Perplexity, Copilot, Claude and Google AI Overviews are all generative engines.See also:LLM (large language model) ·GEO (Generative Engine Optimization)

LLM (large language model)

A model trained on vast amounts of text that predicts the most likely continuation of a sequence of words. It does not query a database of facts: what it knows is spread across its parameters, which is why it needs a retrieval layer to answer about current things.See also:Generative engine ·Hallucination ·Context window

AI Overview

A summary generated by Google’s AI that sits above the classic results and cites a handful of linked sources. It triggers mostly on informational and complex queries and pushes the ten blue links down the page, which makes a citation inside the summary worth more than a ranking position.See also:AI Mode ·AEO (Answer Engine Optimization) ·Zero-click

AI Mode

Google Search’s conversational mode, where the answer fully replaces the result list and follow-up questions are allowed. It breaks the query into several internal searches before writing, so a page can end up cited for a sub-query the user never typed.See also:AI Overview ·Query fan-out

Grounding

Anchoring a model’s answer in verifiable external sources rather than letting it come out of its parameters alone. It is the mechanism that makes citation possible: without grounding there is no link to show, and with it the answer’s quality depends on which sources were retrieved.See also:RAG (retrieval-augmented generation) ·Hallucination ·Citation

RAG (retrieval-augmented generation)

A two-step architecture: relevant document fragments are retrieved first, then the model writes the answer using them. It is why GEO works at all: if your content is not retrieved in step one, the model cannot cite you no matter how well written it is.See also:Grounding ·Chunk (retrievable fragment) ·LLM (large language model)

Context window

The maximum amount of text a model can hold at once, counting the question, the retrieved sources and the answer. When the sources do not fit, the system trims, and whatever is left out cannot be cited — which is why the answer should sit early in the document.See also:LLM (large language model) ·Chunk (retrievable fragment)

Hallucination

A statement a model presents as true when it is not, usually because it fills an information gap instead of admitting it. For a brand this is a concrete risk: if your data is ambiguous or contradictory across sources, the model will fill the gap with whatever seems plausible.See also:Grounding ·LLM (large language model) ·Entity disambiguation

Citation and measurement

Citation

A linked reference a generative engine includes in its answer to attribute where a claim came from. It is GEO’s unit of measurement, the equivalent of the click in classic SEO, and not every mention carries one: a brand can be named without its site being linked.See also:Brand mention ·Citation rate ·Grounding

Brand mention

The appearance of a brand name inside the text the model writes, with or without a link. It is measured separately from citation because it does a different job: a mention builds recall even when it brings no traffic, and it usually precedes the citation as a brand gains ground.See also:Citation ·Brand sentiment ·Share of voice (in AI)

Share of voice (in AI)

The percentage of answers that mention a brand out of all answers analysed for a set of prompts. It is computed on mentions, not clicks, and varies widely from engine to engine for the same brand, so it only makes sense read engine by engine and against named competitors.See also:Brand mention ·Tracked prompt ·AI visibility

Tracked prompt

A fixed question sent to several engines on a schedule to record what they answer and whom they cite. It is the equivalent of a keyword in classic rank tracking, with one difference: the answer changes between runs even for an identical question, so you need a time series.See also:Share of voice (in AI) ·Citation rate

Citation rate

The proportion of answers in which a domain appears cited with a link, out of all answers for a set of prompts. It is stricter than share of voice, which counts mentions: a brand can be named a lot and still have a low citation rate.See also:Citation ·Tracked prompt

AI visibility

An aggregate measure of how much a brand appears in generative engines’ answers, combining mentions, citations and position within the text. No engine publishes this figure: it is built by monitoring your own prompts, which is why two tools can report different numbers.See also:Share of voice (in AI) ·AI referral traffic

Zero-click

A search that ends without the user visiting any website because the answer was already on screen. It is the default scenario in generative engines, and it forces you to value the appearance itself — brand, context, recommendation — not just the session that reaches analytics.See also:AI Overview ·AI referral traffic

AI referral traffic

Visits arriving at a site from the link inside a generated answer, identifiable in analytics by domains such as chatgpt.com or perplexity.ai. It is usually low volume and high intent: whoever clicks has already read the answer and comes to verify or to buy.See also:Zero-click ·AI visibility

Brand sentiment

The tone with which a model describes a brand when it names it: recommendation, reservation, or plain listing. It matters as much as frequency, because a mention paired with a recurring caveat does more damage than absence, and it usually comes from reviews or forums the model reads.See also:Brand mention ·Hallucination

Query fan-out

The engine’s breakdown of one question into several internal searches before it writes the answer. It explains why a page can be cited for queries that never show up in Search Console: the query that earned the citation was generated by the system, not typed by the user.See also:AI Mode ·RAG (retrieval-augmented generation)

Entity and content

Entity

A real-world thing — a company, a person, a product — that a system recognises as a single object rather than a string of text. Being a clear entity to the model is the precondition for any citation: if it does not know who you are, it cannot recommend you.See also:Knowledge graph ·Entity disambiguation ·Structured data (schema.org)

Knowledge graph

A database that stores entities and the relations between them: who founded what, which product belongs to which company, where it is based. Engines use it to disambiguate names and to know what can safely be said about a brand before recommending it.See also:Entity ·Structured data (schema.org)

Entity disambiguation

Deciding which of several same-named entities a text refers to. It is a real problem for brands with a common or shared name: if the model confuses you with another company, your mentions are credited to it and no measurement tool flags it as an error.See also:Entity ·Knowledge graph ·Hallucination

Structured data (schema.org)

Markup that declares, in a machine-readable format, what each thing on a page is: an organisation, a product, an author, a frequently asked question. It does not guarantee a citation, but it removes ambiguity — and a model prefers to cite what it can attribute without risk.See also:Entity ·Knowledge graph

Chunk (retrievable fragment)

The piece a document is split into so it can be retrieved on its own, usually a few paragraphs. The model does not read your whole page: it reads isolated chunks. That is why a paragraph opening with “as we saw above” is useless out of context and never gets cited.See also:RAG (retrieval-augmented generation) ·Citable content ·Context window

Citable content

Text written so it can be extracted and attributed without ambiguity: the claim comes first, stands on its own, and carries its figure, its date and its source. It is not the same as good content; an excellent article argued as a chain is hard to quote in one paragraph.See also:Chunk (retrievable fragment) ·Original data ·Citation

E-E-A-T

Experience, expertise, authoritativeness and trust: the four criteria Google uses to judge who stands behind a piece of content. In GEO they translate into concrete signals — an identifiable author with a track record, an update date, cited sources — that make a model willing to attribute something to you.See also:Entity ·Original data

Pillar page

A page that covers a whole topic end to end and from which every piece that develops it hangs by a link. In GEO it does two jobs: it concentrates the topic’s internal authority and gives the model a single document holding the definition, the method and the examples.See also:Topic cluster ·Citable content

Topic cluster

A set of pieces about the same topic, linked to each other and to their pillar page. What tells an engine a brand owns a subject is not one excellent stray page but a coherent set in which each piece answers a different question without repeating the others.See also:Pillar page ·Entity

Original data

A figure only whoever measured it can publish, because it comes from their own product, clients or research. It is the hardest advantage to copy in GEO: a competitor can rewrite your definition in an afternoon, but cannot invent your sample or your measurement date.See also:Citable content ·E-E-A-T

Access and technical foundations

GPTBot

OpenAI’s crawler that collects web content. It is controlled from robots.txt through its own user-agent, and blocking it has a consequence worth understanding first: it lowers the odds of your content being retrieved and cited in ChatGPT’s answers.See also:robots.txt ·PerplexityBot ·Google-Extended

PerplexityBot

Perplexity’s crawler. It matters more than others because Perplexity shows its sources on screen alongside the answer: it is the engine where a citation is visible, clickable and turns into a visit, so blocking it gives up the most direct traffic in the generative ecosystem.See also:GPTBot ·robots.txt

Google-Extended

A robots.txt control that decides whether Google may use your content to train and feed Gemini, without affecting your indexing in ordinary search. It is the lever that lets you stay in the classic index while staying out of Google’s AI, or the other way round.See also:GPTBot ·AI Overview ·robots.txt

ClaudeBot

Anthropic’s crawler for Claude. It is declared in robots.txt like the others. Worth checking separately, because many inherited robots templates block any unknown user-agent by default, and the result is an exclusion nobody consciously decided.See also:GPTBot ·robots.txt

llms.txt

A Markdown file at the domain root that tells a model what the site is and which its useful pages are. It is a community convention, not a standard any engine promises to read: cheap to maintain, and no substitute for having well-structured content.See also:robots.txt ·Pillar page

robots.txt

A file that tells each crawler what it may walk through. In GEO it is the first check of any audit, because a block inherited years ago may be keeping AI bots out with nobody having decided it. It does not prevent indexing: the noindex tag does that.See also:GPTBot ·Google-Extended ·Crawl budget

Server-side rendering (SSR)

Delivering complete HTML from the server instead of building it in the browser with JavaScript. It is decisive in GEO because most AI crawlers do not execute JavaScript: whatever only appears after hydration simply does not exist for them.See also:Chunk (retrievable fragment) ·Crawl budget

Canonical

A tag that points to the good version of a page when several carry the same content. It concentrates signals on a single URL, and that decides which address ends up cited: without it, authority splits across duplicates and none stands out enough.See also:hreflang ·Crawl budget

hreflang

A declaration that links the versions of the same page across languages. In GEO it prevents a specific problem: an engine citing your English version to someone who asked in Spanish, or treating translations as duplicate content instead of one and the same piece.See also:Canonical ·Entity disambiguation

Crawl budget

The number of pages a crawler is willing to walk through on a site in a given period. It is spent on whatever it finds, not on what matters to you: hundreds of near-identical pages eat the budget the few you actually wanted retrieved needed.See also:robots.txt ·Canonical ·Server-side rendering (SSR)