AEO Glossary: 27 Source-Backed AI Visibility Terms
Clear definitions, examples, measurement boundaries, and reviewed sources for 27 terms used in answer engine optimization and AI visibility work.
Mar 12, 2026
An AEO glossary gives teams a shared, evidence-aware language for discussing what a website can control and what external answer systems actually showed. This page defines 27 terms without treating site checks as AI visibility or turning unavailable measurements into negative results.
Last reviewed: August 18, 2026. Sources are dated in each expanded term record so readers can see when they were checked.
Explore 27 source-backed AEO terms
Search a term or browse by its first letter. Each short definition expands to show an example, measurement boundaries, reviewed sources, useful pages, and related terminology.
agent.json is an experimental SkillAEO format for organizing selected public facts about a site or product in a machine-readable file. It is not an industry standard, authorization protocol, or crawler directive. Publishing it does not guarantee discovery, indexing, citations, recommendations, rankings, or use by any external AI system.
Review evidence and related terms
Example
A team generates a draft agent.json, checks every fact against its public pages, removes unsupported claims, and treats the file as optional documentation.
Related measurements
File availability and factual review status only
Excluded from observed visibility and site-readiness scores
An AI mention is an observed response that names the tracked brand or domain. A mention is not automatically a citation, recommendation, endorsement, or first-choice result. It should be stored with the exact prompt, provider, model or surface when available, timestamp, raw response, and the rule used to classify the mention.
Review evidence and related terms
Example
A saved response says “Acme provides payroll software” but links only to a third-party review. Acme is mentioned, but it does not have an owned citation.
AI visibility describes what a defined set of AI answers actually showed about a brand, page, or source at a recorded time. It is an observation scope, not a permanent property or universal ranking. A useful visibility statement names the prompts, systems, valid-run denominator, time, locale, and evidence behind each mention or citation.
Review evidence and related terms
Example
A report states that a brand appeared in four of five valid Perplexity answers on August 18, rather than claiming the brand is broadly visible everywhere.
Related measurements
Mention rate, citation rate, and first-choice rate
Answer engine optimization is the practice of making public information easier for answer systems and their users to find, understand, verify, and cite. It combines useful content, clear site structure, crawlable pages, evidence, and measurement. AEO cannot force an external system to mention, cite, rank, or recommend a brand.
Review evidence and related terms
Example
A software company publishes a clear pricing explanation, comparison page, methodology, and support documentation, then measures the same approved prompts over time.
Related measurements
Observed mention, citation, and first-choice rates for a fixed prompt set
Site-readiness checks reported separately from external observations
Answer readiness is a site-side assessment of whether important public information is accessible, specific, internally consistent, and supported by relevant pages. It can include crawlability, factual clarity, architecture, structured data, and support coverage. Readiness is not observed AI visibility and does not predict or guarantee an external mention, citation, or recommendation.
Review evidence and related terms
Example
A pricing page names plan limits and billing terms, links to supporting policies, and matches its structured data, so reviewers can verify the claims without inference.
Related measurements
Separate crawlability, factual clarity, support coverage, and structure checks
Unavailable checks excluded rather than scored as zero
An answer surface is the specific product, mode, or interface where a generated response is observed, such as ChatGPT Search, Perplexity, Google AI Overviews, or Google AI Mode. Surfaces can use different retrieval systems, models, locations, and presentation rules, so results from one surface should not be generalized to every other surface.
Review evidence and related terms
Example
A team records Google AI Mode and Google AI Overviews as separate surfaces because their responses and displayed links can differ for the same query.
Related measurements
Valid and unavailable runs grouped by exact surface
Surface, mode, locale, and model recorded with each run when exposed
A baseline is the first saved, reviewable measurement used for a later comparison. It must preserve the approved prompt set, answer surfaces, provider context, timestamps, raw responses, citations, and unavailable states. A later rerun is comparable only when material method changes are disclosed; sequence alone does not prove that site changes caused answer changes.
Review evidence and related terms
Example
Before revising its comparison page, a team saves ten prompt results and their citations, then reruns the same approved set after the revision.
Related measurements
Baseline date and immutable prompt-set identifier
Method changes flagged before comparing later runs
A citation is a source reference or link attached to an observed answer. It identifies material the answer surface displayed as supporting or related evidence, but it does not by itself prove endorsement, accuracy, ranking, or causation. Citation records should preserve the exact destination, source domain, answer, prompt, surface, and run time.
Review evidence and related terms
Example
An answer names Acme and links to an industry directory. The directory is a citation; Acme is mentioned but the citation is not owned by Acme.
Related measurements
Cited responses divided by valid responses
Owned, competitor-owned, and unknown source relationships reported separately
Crawler control is the use of documented site and platform mechanisms to manage automated access or presentation. Different crawler tokens can serve search, training, advertising, or user-triggered purposes, and their rules are not interchangeable. A crawler policy can permit or restrict requests, but it cannot guarantee indexing, exclusion, citations, or product behavior.
Review evidence and related terms
Example
A site allows OAI-SearchBot for ChatGPT search while disallowing GPTBot for training, following OpenAI’s documented independent controls and published tokens.
Related measurements
Allowed, blocked, or not-declared state per documented crawler token
HTTP availability of robots.txt recorded separately from policy interpretation
An entity is a distinct thing that can be named and described, such as an organization, product, person, place, or concept. On a website, consistent names, factual descriptions, relationships, and matching visible structured data can reduce ambiguity. Entity clarity helps interpretation, but it does not guarantee search inclusion or AI citations.
Review evidence and related terms
Example
A company consistently uses one product name, identifies the operating organization, and keeps pricing, support, and structured information aligned across public pages.
Related measurements
Name and fact consistency across reviewed public pages
Structured data compared with visible page content
First-choice rate is the share of valid prompt runs in which the tracked brand is clearly presented as the primary option under a documented classification rule. It is stricter than a mention and different from a citation. The denominator must exclude unavailable runs, and ambiguous lists should not be silently treated as first-choice recommendations.
Review evidence and related terms
Example
If a brand is clearly selected first in two of eight valid answers, its first-choice rate is 2/8; a failed ninth run remains unavailable.
Related measurements
First-choice runs divided by valid runs
Ambiguous or unranked lists kept outside the positive numerator
Generative engine optimization is a research and practice area concerned with how content appears in responses produced by generative retrieval systems. The original GEO work proposed visibility measures and a black-box evaluation framework. Results depend on the domain, query set, system, and method, so a reported experiment should not become a universal ranking promise.
Review evidence and related terms
Example
A research team tests controlled content variants against a fixed benchmark and reports visibility differences by domain instead of claiming one tactic works everywhere.
Related measurements
Defined visibility measure for a fixed benchmark and system
Domain, query set, experimental method, and uncertainty disclosed
Internal linking is the use of crawlable links between pages on the same site. Clear anchors help people and crawlers find related definitions, evidence, policies, comparisons, and product information. Internal links can clarify site structure and context, but adding many repetitive links or pages does not guarantee indexing, rankings, AI use, or citations.
Review evidence and related terms
Example
A glossary definition links to the methodology and a related term, while both related term records link back to the original definition on the same page.
Related measurements
Important pages reachable through crawlable internal links
Reciprocal related-term links resolve to visible stable anchors
llms.txt is an optional community proposal for publishing a concise Markdown overview and selected links for possible use by language-model tools. It is not a W3C or IETF standard, crawler permission file, sitemap replacement, or Google AI requirement. Publishing one does not guarantee reading, discovery, indexing, citations, recommendations, or rankings.
Review evidence and related terms
Example
A documentation site publishes a reviewed llms.txt as an extra navigation aid while keeping normal HTML, internal links, sitemaps, and robots controls authoritative.
Related measurements
File availability, freshness, and factual review status only
Excluded from observed visibility and site-readiness scores
Mention rate is the number of valid prompt runs that mention the tracked brand divided by all valid runs in the approved scope. Provider failures, parse failures, and other unavailable observations stay outside the denominator. The result is specific to the recorded prompt set, surfaces, models, locale, and time; it is not universal visibility.
Review evidence and related terms
Example
Four mentions across five valid responses produce a 4/5 mention rate. A sixth provider failure is reported separately and does not become a negative response.
Related measurements
Mentioned valid runs divided by all valid runs
Numerator, denominator, and unavailable count shown together
An owned citation is a cited URL whose normalized domain belongs to the tracked organization under the report’s documented ownership rule. It is narrower than any citation and separate from a brand mention. Redirects, subdomains, regional domains, and third-party profiles require explicit treatment so ownership is not inferred inconsistently after the run.
Review evidence and related terms
Example
A response cites docs.acme.com, which the audit records as an approved Acme subdomain. A citation to an independent review site is not owned.
Related measurements
Valid responses with at least one owned citation
Source relationship stored as owned, competitor-owned, or unknown
A parse failure occurs when a provider response was received but the required response fields could not be parsed or validated under the recorded contract. It differs from a provider failure, where no usable provider response returned, and from a negative brand result. A parse failure remains unavailable and must not be converted to zero.
Review evidence and related terms
Example
The provider returns text but omits a required structured field, so the run is stored as a parse failure rather than “brand not mentioned.”
Related measurements
Parse-failure count and rate reported separately
Original provider context retained for debugging without inventing a result
A prompt is the exact instruction or question submitted to an answer surface for a recorded run. Small wording changes can produce different responses, so measurement should preserve the complete text, intent classification, locale, and relevant context. A prompt is evidence input, not a keyword ranking position or a promise of a stable answer.
Review evidence and related terms
Example
The audit saves “What payroll tools support contractors?” exactly as run, rather than later shortening it to “payroll tools” in the evidence record.
Related measurements
Exact approved prompt text and stable identifier
Prompt type, locale, surface, and run time stored with the response
A prompt set is the approved collection of exact prompts used together for an audit, baseline, or rerun. Its size, selection method, topic coverage, locale, and version shape the resulting metrics. Reusing an immutable set improves comparability, while adding or rewriting prompts creates a method change that should be disclosed before comparison.
Review evidence and related terms
Example
A team approves five discovery and direct questions, stores the version, and reuses those exact prompts for the next measurement instead of regenerating them.
Related measurements
Prompt count, composition, version, and approval time
Coverage by intent or topic without treating the set as population-complete
A provider failure occurs when the selected external provider did not return a usable answer because of an upstream error, timeout, access limit, or service condition. It is not a negative brand observation and differs from a parse failure after a response. The run remains unavailable, with provider and time context preserved for review.
Review evidence and related terms
Example
An answer request times out before any response arrives. The report records the provider failure and excludes that run from mention and citation denominators.
Related measurements
Provider-failure count and rate by surface or provider
Unavailable run excluded from visibility numerators and denominators
A raw response is the complete answer text and provider payload retained from a recorded run before summary metrics or editorial interpretation. It gives reviewers the evidence needed to inspect mentions, citations, recommendation wording, and parsing decisions. Sensitive or unsupported fields may require controlled storage, but a score without inspectable underlying evidence is weaker.
Review evidence and related terms
Example
A report lets an authorized reviewer expand the complete answer, see its cited URLs, and compare the stored classification with the original wording.
Related measurements
Raw-answer availability for each valid run
Classification trace linked to the exact stored response and citations
Repeat consistency describes how often a documented classification remains the same across comparable valid runs of an approved prompt. It does not mean answers are deterministic or permanently stable. The calculation must identify the compared runs, preserve method changes, and exclude unavailable observations rather than treating provider or parsing failures as changed answers.
Review evidence and related terms
Example
A brand remains mentioned in three comparable valid reruns of one prompt; a fourth provider failure is unavailable and not counted as inconsistency.
Related measurements
Matching classifications divided by comparable valid reruns
Method changes and unavailable runs displayed alongside the rate
robots.txt is a root-level text file that communicates crawl permissions to user agents that follow the Robots Exclusion Protocol. Rules depend on the documented token and purpose. The file manages requests, not authentication, and it should not be used as a guaranteed indexing-removal mechanism, ranking control, citation switch, or substitute for page-level directives.
Review evidence and related terms
Example
A site reviews separate groups for Googlebot, OAI-SearchBot, and GPTBot instead of assuming one wildcard rule has the same product effect everywhere.
Related measurements
HTTP status, final URL, and matched root rule
Allowed, blocked, or not-declared state per documented user agent
Run context is the recorded information needed to interpret and reproduce an observation: exact prompt, surface, provider, exposed model, locale, timestamp, prompt-set version, and relevant method version. Without that context, two answers may look comparable when they were produced under different conditions, and a metric may imply more stability than the evidence supports.
Review evidence and related terms
Example
A saved answer includes its prompt, Perplexity provider, model label, English locale, UTC timestamp, and approved prompt-set identifier beside the raw text.
Related measurements
Completeness of required provenance fields per run
Method or model changes flagged before baseline comparison
Site readiness is a direct review of the audited website’s accessible content and technical presentation, kept separate from observed answers produced by external systems. It can cover crawlability, factual clarity, support coverage, architecture, trust consistency, and structured data. A readiness score describes reviewed site checks; it is not an AI visibility score or prediction.
Review evidence and related terms
Example
A report shows strong crawlability but missing comparison support, while a separate observed-visibility section records whether external answers mentioned or cited the site.
Related measurements
Valid site checks and unavailable checks shown separately
Category-level readiness based only on directly inspected site evidence
Structured data is machine-readable markup that describes visible page information using a shared vocabulary such as Schema.org. It should match the page users can see and follow the policies for any claimed search feature. Google states that generative search requires no special schema, so markup does not guarantee AI inclusion, rankings, or citations.
Review evidence and related terms
Example
A glossary emits a DefinedTermSet whose names, descriptions, and anchors match the visible records, rather than publishing hidden definitions that users cannot inspect.
Related measurements
Markup values compared with visible content and canonical URLs
Syntax validation without treating valid markup as a visibility result
An unavailable observation is a planned measurement that produced no valid classified result because of a provider failure, parse failure, or another recorded execution problem. It is not converted to zero, “not mentioned,” or “not cited.” Reports should show the unavailable count and exclude those runs from valid-result denominators while preserving the reason.
Review evidence and related terms
Example
Five prompts were planned, four returned valid answers, and one failed upstream. Metrics use four as the denominator and show one unavailable observation.
Related measurements
Unavailable count by explicit failure reason
Planned runs, valid runs, and unavailable runs reconciled
Start with the 40–80 word direct definition, then expand a term when you need its example, related measurements, official and research references, methodology sources, tools, or connected terminology.
The glossary intentionally keeps these definitions on one page, not across many nearly identical thin term pages. Stable anchors let guides and related terms point to the exact record while the surrounding context remains visible.
Three boundaries matter throughout the page:
Observed AI visibility records what defined answer surfaces returned for an approved prompt set at a stated time.
Site readiness records direct checks of the website and stays separate from external observations.
Unavailable observations preserve provider failures and parse failures as explicit states rather than converting them to “not mentioned” or a score of zero.
The Methodology explains the complete measurement rules, while Data Sources explains what evidence is retained and where coverage has limits. You can apply the terminology to a real domain with the free AI visibility audit.
A glossary can help readers understand unfamiliar terms and find the relevant product or documentation page. It is not evidence that an AI system has indexed, cited, or recommended your website. Start with terminology your customers actually need, rather than creating a page for every keyword variation.
For example, a workflow product might define approval routing, show how an invoice moves from a requester to an approver, and link to setup instructions. “Approval routing improves efficiency” alone does not explain the inputs, permissions, or exceptions a buyer needs to understand. This is an illustrative content example, not a measured ranking result.
Before publishing a definition, check that it includes:
A plain-language answer that stands on its own.
A concrete example and an important limitation or common misunderstanding.
A relevant source for factual claims, where available.
A useful next page, such as a setup guide or feature explanation—not a link added only to repeat a keyword.
To evaluate the change, review that page's actual search queries and clicks in Search Console separately from any recorded AI answers. Keep the prompt and run context when comparing AI observations. More glossary entries or a higher site-readiness score alone do not establish better visibility. Google's AI features guidance does not require a special AI file or schema for Search AI features.
If an audit reports a citation gap, first open the raw answer and check whether a source URL was actually returned. An unlinked brand mention is not a verified citation. If it reports a site-readiness issue, inspect the specific page and finding before editing it. If a run is unavailable, investigate the failure instead of rewriting content to explain a result that was never measured.
Open the sample report to see the difference, or audit your own domain and choose one evidence-backed issue to review. After publishing a fix, verify the page itself before comparing a new model response.
llms.txt is an optional community proposal, not a search-engine requirement or a promise of discovery, ranking, inclusion, citation, or recommendation.
agent.json is an experimental SkillAEO format, not an industry standard, crawler directive, or authorization protocol. It does not guarantee that an external AI system will read or use the file.
An AEO glossary is a shared reference for terms used in answer engine optimization, site readiness, and observed AI visibility. This glossary pairs each concise definition with an example, measurement boundary, dated source, useful tool or guide, and related terminology.
The labels overlap, but they are not universal standards. This glossary uses AEO for the practical work of making public information findable, understandable, verifiable, and measurable, while GEO refers to the research and practice of improving content visibility within generative-engine responses.
No. Site readiness describes checks performed on the website. Observed AI visibility describes what defined external answer surfaces returned. The two may be reviewed together, but one cannot be presented as proof or a guarantee of the other.
A provider failure produced no response, while a parse failure produced a response that could not be classified reliably. A valid response that does not contain the brand is a different observed result. Combining them would distort the denominator and understate uncertainty.
The catalog records a visible review date and a checked date for every source. Review dates show when references and wording were checked; they do not imply that an external product, model, or search surface will remain unchanged.
