The vocabulary of AI visibility.

AI visibility is what an AI assistant says about you when somebody asks it a buying question and does not mention you by name. It has three parts: whether you are named at all, where you sit when the answer is ordered, and how the answer describes you.

It is not a ranking. There is no results page to be first on. An assistant returns one answer, and you are either in it or you are not.

See also: GEO (Generative Engine Optimization), AEO (Answer Engine Optimization), Recommendation position

Generative Engine Optimization is the discipline of shaping the public evidence about a business so that generative systems include it in their answers. Where SEO optimizes for a position in a list of links, GEO optimizes for inclusion in a written answer.

The levers overlap with SEO but are weighted differently: entity consistency, structured data, third-party mentions and depth of public evidence matter more, and keyword placement matters much less.

See also: AEO (Answer Engine Optimization), AI visibility, Structured data (schema markup)

Answer Engine Optimization targets surfaces that return one answer instead of ten links: featured snippets, voice assistants, AI Overviews. The unit of success is being the source the answer rests on.

GEO and AEO are often used interchangeably. The useful distinction: AEO is about being the answer to a question, GEO is about being present in generated prose that may not be a question at all.

See also: GEO (Generative Engine Optimization), Google AI Overviews, Citation

A large language model produces text by predicting what comes next, given everything it has read during training and everything it is handed at question time. ChatGPT, Gemini, Perplexity and Claude are products built on such models.

Two consequences matter for visibility. What the model absorbed during training is fixed until the next training run, and what it retrieves at question time can change hour to hour. Your visibility rides on both.

See also: Training data, Grounding, Retrieval

Training data is the corpus a model learned from. If your business is described consistently across many independent sources in that corpus, the model has a stable internal picture of you. If it is described three different ways, or barely at all, it does not.

You cannot edit training data. You can change what the next crawl of the open web says about you, which is the only lever anyone actually has.

See also: Large language model (LLM), Brand mention, NAP consistency

AI Overviews are the block of generated text Google places above the blue links for many queries, with a small set of linked sources. For a large share of informational and local queries it is now the first thing on the page.

Being cited in an Overview is a different outcome from ranking on the page below it. A page can rank fourth and still be one of the cited sources, and a page can rank first and be absent.

See also: Citation, Zero-click, Google AI Mode

AI Mode is the fully conversational version of Google search: a dialogue with follow-up turns rather than a single query and a page of results.

It changes the shape of the problem. Visibility has to survive several turns of a conversation, not one query, and a business can be introduced in turn one and dropped by turn three.

See also: Google AI Overviews, Prompt

An answer engine reads the sources for you and returns a conclusion. Perplexity is the clearest example; AI Overviews, voice assistants and chat assistants all behave this way for many queries.

The economics differ from search: there is one winner per answer, not ten positions, and the click is optional.

See also: AEO (Answer Engine Optimization), Zero-click, Google AI Overviews

A zero-click result is one where the searcher gets what they needed without opening any page. Analytics shows nothing at all: no session, no referrer, no event.

This is why analytics alone cannot tell you whether your AI visibility is improving. The measurement has to happen on the answer side, by asking the question and reading what comes back.

See also: Google AI Overviews, Prompt-level tracking

In AI search, a citation is a source an answer links to or attributes a claim to. In local SEO, a citation is a listing of your name, address and phone number on a directory. Both matter, and confusing them costs people months.

Directory citations feed the entity picture that models build. Answer citations are an outcome of that picture, plus the depth and readability of the page being cited.

See also: NAP consistency, Google AI Overviews, Brand mention

A grounded answer is one the system has attached to documents it actually retrieved, rather than one produced from memory alone. Grounding is why the same assistant can name a business it has never been trained on.

For a business, grounding is the fast lane: it does not require the next training run. It requires that a retrievable page states the thing clearly enough to be quoted.

See also: Retrieval, RAG (retrieval-augmented generation), Hallucination

Retrieval is the step where a system searches for relevant material after the question arrives, instead of relying only on what it memorized. Most assistants now do this for anything local, recent or specific.

It is the reason AI visibility can move within days rather than waiting for a model refresh, and the reason it can also move for reasons that have nothing to do with you.

See also: RAG (retrieval-augmented generation), Grounding, Large language model (LLM)

Retrieval-augmented generation is the standard architecture behind assistants that cite sources: search, select passages, then generate an answer constrained by those passages.

The practical implication is unglamorous. Pages that state facts plainly, in text, near a heading that matches the question, get selected. Pages that bury the same fact in an image or a script do not.

See also: Retrieval, Grounding, Structured data (schema markup)

A hallucination is a fabricated detail delivered in the same tone as a true one: wrong hours, a service you do not offer, an address you moved out of years ago.

The defense is redundancy. When the same facts appear consistently across your site, your listing and independent sources, there is less room for a model to fill a gap with invention.

See also: NAP consistency, Grounding, Entity

An entity is the concept behind the name. Two businesses with similar names are two entities; one business written five ways is still one entity, provided the machine can tell.

Entity clarity is the single most under-rated AI visibility lever. Everything downstream, from citation to recommendation, depends on the system being sure which business you are.

See also: Knowledge graph, NAP consistency, Structured data (schema markup)

A knowledge graph stores entities and the relationships between them: this business is in this category, in this city, offering these services, with these credentials.

Structured data is how you volunteer your own edges of that graph rather than leaving them to be inferred.

See also: Entity, Structured data (schema markup)

Structured data is JSON-LD embedded in a page that states, unambiguously, what the page is about: LocalBusiness, Service, FAQPage, Review, Organization and so on.

It does not make claims true, and it does not rank a page on its own. What it does is remove ambiguity, which is exactly the failure mode that keeps businesses out of generated answers.

See also: Knowledge graph, Entity, RAG (retrieval-augmented generation)

NAP consistency means your name, address and phone number are written identically across your site, your Google listing and every directory that carries you.

Inconsistency is not a cosmetic problem. It splits one entity into several weaker ones, and a system that is unsure which record is current tends to name a competitor it is sure about.

See also: Entity, Citation, Google Business Profile

A brand mention is any independent reference to your business in public text: press, forums, roundups, partner pages, reviews. It counts whether or not it carries a link.

Mentions are how a model learns that other people, not just you, consider your business worth naming. That is a different signal from anything you can publish about yourself.

See also: Citation, Training data, Entity

A Google Business Profile carries your category, service area, hours, attributes, photos and reviews, and it is one of the most heavily cross-referenced public records a local business has.

It is strong corroborating evidence for AI visibility rather than the thing itself. A complete listing makes it easier for a system to be confident about you; it does not make an assistant recommend you.

See also: NAP consistency, Entity, Local evidence

A prompt is the literal text sent to an assistant. Wording changes results, so a measurement is only comparable over time if the prompt is held stable.

A measurement prompt must not contain the name of the business being measured. Ask an assistant about a named business and it will discuss that business, which measures the question rather than the visibility.

See also: Prompt-level tracking, AI visibility

Prompt-level tracking means asking the same buyer questions across the same assistants at a set cadence, and storing every answer, so that changes in the result can be attributed to changes in the world rather than to a reworded question.

It is the only way to observe a zero-click surface. Nothing about it shows up in your analytics.

See also: Prompt, Zero-click, Visibility score

Mention presence is the first and bluntest reading taken from an answer: your business is in it, or it is not.

Name matching has to be tolerant. Businesses are named in shortened forms, without their legal suffix, or with words reordered, and a strict string comparison reports false absences.

See also: Recommendation position, Answer sentiment, AI visibility

When an answer returns a ranked list, position is the second reading. Being named first and being named eighth are different results and should not collapse into the same number.

Not every answer is ordered. A prose answer that mentions you without ranking anything has presence and sentiment but no position, and inventing one would be fiction.

See also: Mention presence, Answer sentiment

Sentiment is the third reading: whether the wording about your business is positive, neutral or negative. A mention hedged with a caveat is worth less than a confident recommendation.

It is read from the answer text itself, not from your reviews. It measures what the assistant said, which is the thing a potential customer would actually see.

See also: Mention presence, Recommendation position

Share of voice is borrowed from advertising: your mentions divided by all mentions in the category. Applied to AI answers it needs a count of every mention of every business in your category, which no tool observes.

Recometric does not report a share-of-voice figure. What it reports instead is a per-competitor mention count taken from the identical answers, which is a smaller claim and a true one.

See also: Mention presence, Competitor tracking

Competitor tracking takes the same three readings for rival businesses, out of the same answers, rather than running them a separate and friendlier question.

That is what makes the comparison meaningful: same prompt, same assistant, same moment, different name.

See also: Mention presence, Share of voice

Local evidence is the part of the picture that can be checked against public records: the listing, the category, the service area, consistency between them.

It is deliberately kept separate from the AI reading. Mixing verified facts and generated opinion into one undifferentiated number hides which of the two moved.

See also: Google Business Profile, NAP consistency, Visibility score

A visibility score compresses many readings into one 0-100 figure. It is only meaningful if you can see what went into it: which assistants answered, what each contributed, and what the non-AI evidence added.

The Recometric Score is built from 4 assistants plus website and Google Business Profile evidence, with each component's share shown next to the number and every input stored, so the figure can be recomputed rather than trusted.

See also: Mention presence, Local evidence, Coverage

Coverage is the share of the intended measurement that produced usable evidence. Assistants rate-limit, time out and occasionally refuse, so partial coverage is normal rather than exceptional.

A partial measurement should be published with its uncertainty rather than presented as a confident point, and a missing answer must never be recorded as a bad one.

See also: Visibility score, Prompt-level tracking