AEO and GEO glossary

The vocabulary of AI search, defined so each entry stands on its own. Several of these words are used loosely enough in this industry to hide a defect — branded query and brand recall in particular are how a visibility report can show a good number for a brand that is invisible. Where a term is contested, the entry says so.

The names for this work

Four labels for one practice, and the one place the acronym means something else entirely.

GEO: Generative Engine Optimization: the practice of making a brand findable, extractable and corroborated so generative answer engines name it. Coined in a 2023 academic paper, now the common industry label.
AEO: Answer Engine Optimization: the same practice named from the buyer's side — optimizing to be the answer rather than a link. In digital marketing GEO and AEO are used interchangeably. In customs and logistics, AEO means Authorised Economic Operator, an unrelated trade authorisation.
LLMO: LLM Optimization. A third name for the same practice. The proliferation of acronyms is a sign of an unsettled category, not of three different disciplines.
AI SEO: A marketing term covering both using AI to do SEO and optimizing for AI answers. Ambiguous; we avoid it.
Answer engine: A system that returns a composed answer instead of a list of links: ChatGPT, Claude, Gemini, Perplexity, Copilot, Google AI Overviews and AI Mode.

How an engine builds an answer

The three steps, in order. A failure at any one makes the next two irrelevant.

Grounding: Attaching a model's answer to retrieved documents so it can cite them. An ungrounded answer comes from training weights and cites nothing; a grounded answer names sources you can check.
Retrieval: The step where an engine searches the live web and pulls candidate pages before writing an answer. If your page is not retrieved, nothing else about it matters.
Extraction: The step where a model lifts a span of text out of a retrieved page. A sentence that only makes sense in the context of the paragraph around it is hard to extract.
Corroboration: A model's preference for a claim that appears on more than one independent source. This is why third-party presence beats a longer page on your own site.
Chunk: The unit a retrieval system stores and returns — usually a heading and the text under it. More headings means more entry points into a page.

What is being measured

Three outcomes that are routinely collapsed into one score, and are not the same thing.

Mention: The brand appearing anywhere in an answer, in any framing.
Citation: A link an engine attaches to its answer. Being cited is not the same as being recommended: a page can be cited as evidence for a competitor's superiority.
Recommendation: The brand named as a suggested option for the asker. The only one of the three that a buyer acts on, and the one worth measuring.
Source graph: The set of domains a model panel actually read while answering. A domain cited by several model families is the panel's shared view rather than one model's quirk.
Model panel: A set of model families queried with the same questions on the same day, so a result is a panel observation instead of one vendor's output. Ours is OpenAI, Anthropic, Google, xAI and Perplexity.
Share of voice: The proportion of answers to a question set in which a brand appears, relative to competitors. Meaningless without the question set and the date attached.

How a measurement lies

Every entry here is a way a visibility report can show a good number for an invisible brand.

Branded query: A question containing the brand name. A model handed a brand almost always recognises it, so a branded query measures recall, not visibility. Scoring one is how a visibility report flatters its buyer.
Discovery query: A question a buyer asks before they know your brand exists. It never contains your brand name. These are the only queries that measure visibility.
Vendor-seeking query: A discovery query whose answer is expected to name products or suppliers — "best X for Y", "tools to do Z". A how-to query returns a tutorial that names nobody, so it cannot measure vendor visibility.
Brand recall: Whether a model can describe a brand it was handed by name. Reported separately from visibility, never folded into a score.
Personalisation: An assistant's use of a signed-in account's memory and history. It makes a signed-in check a measurement of the account, not the brand.
Temporary chat: A signed-out or memory-disabled session. The only client-side way to see something close to what a stranger sees.
Locale drift: A brand being present in one language and absent in another for the same question. Measuring in one language and reporting it as global is a category error.

Crawlers and files

The tokens are not interchangeable. Blocking the wrong one costs visibility you meant to keep.

GPTBot: OpenAI's training crawler. Blocking it withholds content from training, not from ChatGPT's live answers.
OAI-SearchBot: OpenAI's index crawler for ChatGPT search. Blocking it removes you from ChatGPT's search results.
ChatGPT-User: OpenAI's live fetcher, sent when one person's question needs one page right now. Its user agent contains no "bot" token, so most analytics bot filters miss it.
Google-Extended: The token that governs whether Google may use a page for Gemini and Vertex grounding. Separate from Googlebot: allowing search does not allow Gemini.
llms.txt: A proposed plain-text file at /llms.txt describing a site for language models. Widely published, and measured as widely ignored: of 137,210 domains carrying one, 97% received zero requests for it.

Entities and markup

How a machine decides that two records describe the same company.

Entity: A thing a knowledge system can identify and hold facts about. A brand becomes an entity when independent sources describe it consistently.
sameAs: The schema.org property linking a site's Organization to its profiles elsewhere. The cheapest way to tell a machine that two records are the same company.
FAQPage: The schema.org type marking question-and-answer pairs. Present on 35–86% of pages across the sites AI assistants recommend in this category.
Zero-click: A search that ends without a visit because the answer was rendered in place. The reason citation share matters even when traffic does not move.

Start with the definition. What is GEO explains how these fit together and what we measured across the six sites AI assistants recommend in this category.