Reference
Thirty working definitions for reading AI visibility research and comparing measurement approaches. Each term has a permanent page you can link to.
Whether and how a business appears in answers from an AI system for a defined set of questions. Visibility depends on the prompts, collection route, location and time tested.
Example: A brand mentioned in one sampled answer has observed visibility for that answer.
The practice of improving how a business can be found, understood and cited in generative AI answers. GEO work can involve first-party content and third-party sources.
Example: An accurate business profile and a clear service page are possible GEO inputs; their effects still need measurement.
The practice of making information useful and accessible to systems that return direct answers. Its scope overlaps with GEO and search engine optimization.
Example: A concise, sourced answer to a common customer question is answer-oriented content.
A search experience that uses generative AI to produce an answer, sometimes with retrieved web sources. Different products and modes can return different answers.
Example: A search-enabled chat answer and a search results page with an AI summary are distinct surfaces.
In Outrigger’s published benchmark context, the percentage of tested buying-intent prompts in which a brand is mentioned. Always specify the prompt set, system and denominator.
Example: Compare rates within the same test design; a change in the prompt menu can change the rate.
A brand’s share of measured mentions relative to a defined comparison set. In AI measurement, the counting rule and competitor set must be stated.
Example: A share among competitor mentions uses a different denominator from the share of prompts mentioning one brand.
An occurrence of a business or brand name in an AI answer. A mention alone does not show endorsement or a source link.
Example: An answer can mention a business while describing a limitation.
A source reference attached to an AI answer, often as a link. A citation and a brand mention are separate observations.
Example: An answer can link to a company page without recommending that company.
An answer that presents a business as a suitable option for the question asked. The rule for classifying a recommendation should be explicit.
Example: A list of suggested dentists differs from a historical reference to a dental practice.
The domain hosting a source linked in an AI answer. Domain-level analysis groups links from the same site.
Example: Several cited pages from one publisher can count as one unique citation domain.
The question or instruction submitted to an AI system, together with any specified context. Wording is part of the measurement design.
Example: “Which plumber serves this area?” and “Is this plumber reliable?” test different questions.
The information need a prompt is designed to represent, such as discovery, comparison or checking a named business.
Example: A “best options” question measures discovery; a named-brand question measures prompted awareness.
An alternative wording intended to express the same underlying information need. A paraphrase family helps test sensitivity to phrasing.
Example: “Who offers this service nearby?” and “Which local companies provide this service?” belong to a possible wording family.
A fresh collection of an answer to the same exact prompt under a stated setup. Repeats help expose run-to-run variation.
Example: Repeating the same wording differs from asking several paraphrases.
The product interface through which a person receives an AI answer. Interface settings and web access can affect what is observed.
Example: A consumer chat application is a different collection route from a developer API.
Obtaining AI responses through a programmatic interface. An API result needs calibration before it is treated as representative of a consumer product.
Example: Record whether web search was enabled when comparing API and consumer answers.
Using external information, such as retrieved web pages, to inform a generated answer. Grounding does not by itself establish correctness.
Example: An answer can consult live web sources and still omit a relevant business.
Finding external material for use in an answer. The material retrieved and the material finally cited need not be identical.
Example: A system may retrieve several pages but cite only some of them.
The real-world business, person or organization a name refers to. Consistent identity matters when matching answers to the intended subject.
Example: Two companies sharing a name need location or domain evidence to distinguish them.
Matching a name or reference in an answer to the correct real-world entity. Ambiguous matches should be retained as unresolved.
Example: A matching name without enough location context may be insufficient to identify the intended business.
Information about a business on sites it does not control, such as directories, review platforms, news sites and forums.
Example: A third-party business listing is an off-page observation.
Machine-readable markup that describes page content and entities. Its presence is an implementation fact, not proof of an AI citation effect.
Example: Organization markup can describe a company’s name and website.
A proposed plain-text convention for pointing language-model tools toward selected website information. Publishing the file does not establish that an AI system will use it.
Example: Treat an llms.txt file as a discoverability aid whose effect requires testing.
A website file expressing crawling directives for user agents. Interpretation depends on the crawler and the specific directive.
Example: A rule for one crawler does not describe every AI system’s access.
A reference distribution or test design used for comparison. A benchmark describes its measured sample and conditions.
Example: Compare a local business with the matching industry and region bucket.
A position in an ordered distribution. In the published benchmark interface, p50 is the median; p25, p75 and p90 describe other points in the distribution.
Example: The median is a reference point, not a recommended target.
The fraction of valid measured answer opportunities in a defined benchmark that mention a target entity. It depends on the test’s intent and wording mix.
Example: This rate is not a market-demand-weighted probability unless the sampling design supports that interpretation.
No target mention was found in the valid answers collected for a specified test. It does not establish that the system cannot mention or recognize the business.
Example: A timeout is a missing measurement, not a valid answer with no mention.
Differences between answers collected on repeated runs of the same setup. Report it separately from changes caused by a new prompt menu.
Example: An option appearing on one repeat and disappearing on another is observed variability.
Correlation describes an association between measured variables. A causal claim says that changing one variable produces a change in another and needs a design that supports that inference.
Example: Businesses with more directory listings may also differ in brand size; an association alone cannot isolate the directory effect.
These are Outrigger’s working definitions. Metric names can mean different things across products. Check the denominator and collection method before comparing numbers.
Start with what AI says about your business.
See what AI says about youSee pricing