39 terms you need before buying or running GEO

GEO is the practice of getting a brand named, recommended and cited inside the answers that generative AI assistants produce. Unlike search optimization, which competes for a position in a list of links, GEO competes for a sentence inside the answer itself.

AEO is optimization aimed at systems that return a direct answer instead of a list of links. In practice its scope overlaps GEO almost entirely; the emphasis is on being the source the answer is built from.

LLMO is a broader label for work aimed at how large language models understand and reference a brand, including both live-retrieval answers and what the model has absorbed from training data.

A generative engine is a system that composes an answer in natural language rather than returning a ranked list of documents. DeepSeek, Doubao, ERNIE Bot, Tongyi Qianwen, Tencent Yuanbao and Kimi are the ones people reach for most.

An answer engine is any system whose primary output is a direct answer — generative assistants, but also search results pages that lead with a synthesised summary above the links.

Black-hat GEO covers tactics that manipulate the distribution of information without creating any real information: fabricated reviews, bulk-generated template content across low-quality sites, and attempts to manipulate model inputs directly.

AI visibility is the degree to which a brand is mentioned, recommended and cited in AI-generated answers. It is measured against a fixed set of questions, not as a single global score.

Mention rate is the proportion of question-and-platform combinations in which the brand appears in the answer. If 20 questions are run across 6 platforms, the denominator is 120.

Share of voice is a brand's mentions divided by the total mentions of all brands across the same question set. It answers "of the names that get said, how many are mine".

Citation share is the proportion of cited sources in an answer that belong to your own domains. It is counted only from sources actually marked in the answer body.

Mention context is how a brand is characterised in the sentence that names it, recorded as positive, neutral or negative. It is a separate field from whether the brand was named at all.

A negative mention is an appearance in an answer where the surrounding context is unfavourable — named as the cheap option, named alongside a complaint, named as the one to avoid.

A baseline is the first full measurement: the frozen question set run across every platform, with mention rate, mention context, competitor names and cited source domains recorded.

Re-measurement is re-running the frozen question set under the same conditions on a schedule, and reporting the difference from the previous round.

Zero-click describes a user getting what they needed from the answer itself and never visiting a site. The influence happened; the visit did not.

A semantic keyword is the question a customer actually asks, not the keyword phrase they would type into a search box. "How much does bookkeeping cost for a small company in Shenzhen" is a semantic keyword; "Shenzhen bookkeeping" is a search keyword.

A semantic keyword group is a set of questions that require the same underlying material to answer. "Which provider is reliable" and "how do I choose a provider" belong in one group; "how much does it cost" belongs in another.

A long-tail semantic keyword is a narrow, specific question — usually combining a category with a location, an industry or a situation. Individually low in volume, collectively large.

A brand query names the company directly: "is XX reliable", "what does XX do". The user already knows the brand and is checking it.

A category query asks about the category rather than a company: "how to choose a GEO provider", "which bookkeeping firm is good". The user does not yet have a shortlist.

A question set is the fixed list of customer questions used to measure AI visibility. It is written down once, frozen for the duration of the engagement, and re-run unchanged on every round.

A source is any page or document a model draws on when composing an answer. Your own site is one source among many; media coverage, encyclopedia entries, B2B platforms and community discussion are others.

An authoritative source is one a model treats as carrying weight: outlets with a real editorial process, encyclopedia entries, official registries and standards bodies. Authority here is behavioural — it is whatever models actually cite.

A source matrix is the deliberate spread of accurate, consistent material about a brand across several source types, so that a model encounters the same facts from more than one direction.

Structured data is machine-readable markup that states explicitly what a page's content means — that this string is a company name, that one a phone number, this block a question and answer.

Schema.org is the shared vocabulary used for structured data on the web — Organization, Article, FAQPage, Product, BreadcrumbList and several hundred other types.

FAQPage is the structured data type for question-and-answer content. It tells a model explicitly which text is a question and which text answers it.

llms.txt is a plain-text file at the site root that indexes the site for large language models — what the organisation is, which pages matter, and how they relate. It is an emerging convention rather than a standard.

robots.txt is the file that tells crawlers which parts of a site they may fetch. For GEO the relevant part is whether AI crawlers are allowed.

An AI crawler is a bot operated by a model vendor to fetch web content — for training, for live retrieval, or for both. Each identifies itself with its own user agent.

Crawlability is whether a crawler can actually retrieve a page — covering robots rules, firewall and CDN behaviour, rate limiting, redirects and status codes.

Render readability is whether a page's content exists in the HTML that a crawler receives, or only appears after JavaScript runs in a browser.

A canonical URL declares which address is the authoritative one for a piece of content when the same content is reachable at more than one address.

An entity is a specific, identifiable thing — a company, a person, a product — as distinct from the words used to refer to it. Models work with entities, which is why two companies with similar names get confused.

A knowledge graph is a structured store of entities and the relationships between them. Search and AI systems use one to decide what a name refers to and what is known about it.

A hallucination is a confident statement by a model that is not supported by any source — an invented service, an invented year, an invented case study.

Live web search is a model fetching current web pages at the moment it answers, rather than relying only on what it absorbed during training.

Training data is the corpus a model absorbed before release, as distinct from the material it retrieves while answering a specific question.

The retrieval candidate pool is the set of documents a model pulls in before composing an answer. Most of them do not end up being used, and most are never shown to the user.