Citation vs mention vs link

What is citation vs mention vs link?

When an AI answer references something, the referenced object can surface in three different ways: as a citation, as a mention, or as a link. The three look similar at a glance but signal different things. GEO Glossary measurement entries (attribution rate, citation match rate, brand mention rate) all depend on this distinction to compute their denominators consistently. The taxonomy is practitioner-coined; no vendor or academic source defines an authoritative three-way split. Other indie editorial references have proposed parallel framings (the sibling project geo.wiki has a Citation vs Mention vs Link entry in its Foundations cluster¹); cluster measurement entries on this site previously relied on cross-reference convention to disambiguate, and this entry codifies the explicit terminology for future cluster integration.

The three dimensions:

• Citation is a structured reference indicating the engine is presenting that source as supporting or grounding part of the answer. Typically appears as a numbered footnote, a chip-style entry in a source panel, an inline link with a small icon, or a "Sources" section at the end of the answer. Citation does not automatically mean the source actually supports the specific claim it is attached to; citation granularity varies by engine (claim-level, paragraph-level, answer-level, or source-list-level) and citation alignment can be loose or inaccurate, so the citation act says "the engine is presenting that source as support for some part of the answer", not "this specific piece came verbatim from there".
• Mention is naming a brand, source, entity, product, or concept inline in the answer prose. The engine may say "GEO Glossary defines cite-ability as ..." without attaching a clickable URL. The mention signals recognition (the model knows the entity exists) but does not by itself signal retrieval (the model may be quoting from training-time memory, not a fresh fetch).
• Link is the URL artifact itself. A link can attach to a citation (the chip-style URL in a source panel), to a mention (the brand name turned into a hyperlink), or to a navigation element separate from either (a "see also" recommendation). The link is a technical rendering choice, not a semantic claim about grounding.

The three are best understood as independent dimensions, not as a strict containment relationship. A mention can be linked or unlinked. A citation typically includes a link but the act of citing is separate from the linkability. A link exists as a property orthogonal to whether the engine is asserting grounding or just naming the entity.

Status in 2026

The terminology is unsettled across vendor documentation and across third-party tracking tools. Google's AI Overview product announcement² uses "sources" for what we call citations (every source-panel entry is linked). ChatGPT search, Perplexity, Claude, and Gemini each use slightly different surface treatments and slightly different terminology in their product copy: "citations", "sources", "references", "linked answers", "with attribution". Third-party tools (Profound, Otterly, AthenaHQ, Peec AI, Ahrefs Brand Radar) each implement counting rules that bundle some combination of the three dimensions, often without surfacing the breakdown in their public dashboards.

The result is a practitioner vocabulary mess. A SaaS vendor saying "your brand has been mentioned 200 times by ChatGPT this week" may be counting linked citations only, may be counting brand-name mentions in body prose regardless of links, or may be counting both. Without the three-dimension split, "mentions" and "citations" become interchangeable in marketing copy, and the underlying metric becomes uncomparable across tools.

GEO Glossary measurement entries explicitly separate the three dimensions and label each anchor entry's denominator decision. This entry exists as the taxonomy hub so future cluster entries can reference the distinction by link rather than restating the disambiguation each time.

How to apply

Three operational moves to record citation / mention / link cleanly across an AI citation measurement program:

• Record two independent axes per reference, not a single binary. During each weekly probe, log each reference along two axes: (1) reference type (citation in a source panel / footnote / chip, vs prose mention in body text) and (2) link state (linked vs unlinked). The combination yields four cells in a 2x2 matrix:

  	Linked	Unlinked
  Citation (source panel, footnote, chip)	Standard well-formed reference	Rare, mostly engine rendering bugs
  Mention (inline in answer prose)	Brand name turned into hyperlink mid-prose	Brand name spoken without attribution

  Tracking the four cells separately preserves the signal that compresses out when you record a single "mentioned vs cited" binary. See brand mentions in AI answers for the parallel framing on the mention side.

• Map each measurement metric to which cells it counts. In this glossary's measurement system: attribution rate counts citation events (linked or unlinked, though unlinked is rare). Citation match rate counts the linked-citation cell only (the strictest definition). Brand mention rate counts the mention cells (linked mention + unlinked mention). Citation share is normalized against total citation instances across all sources for the prompt set, inheriting the citation-vs-mention distinction at the denominator. Knowing which cells each metric counts is what makes per-engine and per-tool numbers comparable.

• Treat per-engine variation as structural. Google AI Overview is the simplest case (every source-panel entry is linked, so citation match rate equals attribution rate). ChatGPT, Claude, and Gemini link state depends on session settings (web search enabled, grounding mode, model variant). Practitioners commonly observe Perplexity as citation-heavy relative to other major consumer AI engines. Per-engine measurement-coverage caveats apply regardless of which dimension you track.

What to skip:

• Collapsing the three into a single "AI references" number. The composite hides the operational signal.
• Assuming a vendor's "mention count" matches your definition. Cross-check the tool's methodology page before treating their number as comparable to your own measurements.
• Treating an unlinked brand mention as a failure of attribution. Unlinked mentions are a distinct downstream channel (delayed direct discovery, branded search lift) and have their own value that linked citation rate does not measure.

What remains contested or unverified

• Whether unlinked mentions correlate with eventual linked citations (the leading-indicator hypothesis). Practitioners commonly assume that accumulating unlinked brand mentions on a topic eventually converts to linked citations as the engine's confidence in the brand grows. The causal chain is not directly demonstrated; the correlation could equally reflect that both metrics respond to the same underlying authority signals over time.
• The exact link-vs-no-link logic per engine. No vendor publishes a per-response policy for when a source gets linked in a citation chip versus when a brand gets mentioned without attribution. Practitioner inference says it correlates with grounding mode and source-trust scoring, but the per-engine rules are not documented.
• Whether structured data markup (DefinedTerm, FAQ, Article schema) shifts citation vs mention behavior. Practitioner speculation in some GEO marketing copy suggests that schema markup encourages engines to treat a source as a citation rather than a mention, but no controlled study has separated this from confounders (schema correlates with editorial quality and primary-source citation discipline, which are independent citation-friendliness drivers). See the DefinedTerm schema cluster entries for the parallel hedging.
• How AI dev tool citations (AI dev tool citations) fit the three-dimension split. Cursor, GitHub Copilot, Codex, and similar coding-context AI tools render references differently from chat AI engines: code-comment cite (rendered as a comment inline in generated code), IDE-panel cite (rendered in a separate side panel), chat-only cite (rendered only in the AI chat output, not in the code), or Cmd+K result attribution (small-popup completion mode). A separate AI-dev-tool citation taxonomy entry may eventually be warranted; the three-dimension taxonomy here is a starting point but does not yet cover the code-context surface cleanly.
• Whether convergent terminology will settle across practitioner references. Both this entry and geo.wiki/citation-vs-mention codify a three-dimension split, but vendor and SaaS-tool usage have not converged on a shared vocabulary. The three-dimension taxonomy described here is one of several viable framings.

How it relates to other concepts

• Attribution rate counts the broad "was the source referenced" question, primarily measuring citations (linked or unlinked).
• Citation match rate counts the linked subset of attribution: the strictest definition.
• Brand mentions in AI answers counts the mention column: linked + unlinked brand-name references regardless of citation chip status.
• Citation share normalizes citations against competitors for the same prompt set, inheriting the citation-vs-mention distinction at the denominator.
• Cite-ability is the content-side property that drives whether content gets cited at all. A passage with low cite-ability may still be mentioned (the brand or topic gets named) without being cited (no source-panel chip), separating the two outcomes.
• AI citation metrics pillar synthesizes the six measurement anchors; the pillar's decision matrix assumes the citation-vs-mention-vs-link split is already known to the reader.
• AI Overview citation is the simplest per-engine case: every source-panel entry is a linked citation, so the three-dimension split collapses to two (linked-citation in the panel, vs mention in body prose).