Citation vs Mention vs Link

1. Three ways an AI answer can credit a source

A generative answer can credit a source in three distinct ways, or it can provide no credit at all:

• A citation connects content that was lifted or paraphrased to an attributable reference.
• A mention names a brand, product, or author in the answer without linking to it.
• A link provides a clickable source that may not correspond to any text the answer used.

GEO Wiki working definition: In a generative answer, attribution is separate from grounding. An engine may use a source without crediting it, and any credit it does provide can take one of three non-equivalent forms: citation, mention, or link.

All three forms occur during step 4 of the answer loop, synthesis and attribution. This is the stage after grounding when the engine either provides a reference or withholds one.

2. Definitions of citations, mentions, and links

The visible result, value, and appropriate metric differ for each form of credit:

A single synthesized answer can contain all three forms at once:

"Generative engines decouple grounding from attribution.[1]   ← citation (source attached to a reused claim)
 According to Otterly.AI, mentions and citations are tracked   ← mention (brand named, no link)
 as separate KPIs. For more on answer mechanics, see the
 sources below.

 Sources:  [1] geo.wiki/citation-vs-mention
           ▸ example.com/unrelated-page                        ← link (shown in tray, no matching sentence)"

Three common observations illustrate these outcomes:

• “It used my facts but didn’t link to or name me” means the source was used without credit; grounding does not guarantee attribution (see §3).
• “It named us, but we got no traffic” describes a mention, not a citation. It is a different outcome.
• “It linked us but didn’t quote us” describes a link, not a citation. Of the three outcomes, it provides the weakest credit.

GEO Metrics defines the formulas for citation share, share of voice, and referral attribution. Brand Mentions explains how to earn mentions.

3. Why grounding does not guarantee credit

Grounding does not guarantee attribution. An engine can use your content to ground an answer and still provide no credit. It can also name you without a link or link to you without quoting your content. Source use and source credit are separate by design, not because the system has malfunctioned.

  grounded subset
        │
        ▼
  ┌──────────────────────────────┐
  │  SYNTHESIS & ATTRIBUTION     │
  │  emit credit?                │
  └──────────────────────────────┘
        │
        ├──► citation   (used + credited + reference)
        ├──► mention    (named, no link)
        ├──► link        (URL surfaced, maybe not even used)
        └──► nothing     (used, never credited)

The same grounded passage can produce any of these four outcomes. Most cases in which an engine uses content but provides no credit occur at this stage, after grounding. Content that is suitable for grounding is not guaranteed attribution.

The order matters. Attribution depends on groundability, which determines whether a passage can be selected at all (Citability). Groundability, in turn, depends on retrievability, which determines whether the page can become a candidate source (AI Crawlers). If either earlier stage fails, the source cannot receive credit. Diagnose problems in that order; Answer Loop §4 maps the possible failures at each step.

This separation is also visible in platform APIs. Gemini returns groundingChunks, which identify the sources used, separately from groundingSupports, which map answer passages back to those sources. Source use and source credit therefore appear in different response fields (Grounding with Google Search). Anthropic’s web search tool likewise returns a separate url and cited_text for each result (Web search tool).

4. What research shows about attribution

Liu, Zhang, and Liang evaluated Bing Chat, NeevaAI, Perplexity.ai, and YouChat in Evaluating Verifiability in Generative Search Engines (Findings of EMNLP 2023). On average, only 51.5% of generated sentences were fully supported by their citations, a measure the authors call citation recall. Only 74.5% of citations supported the sentence associated with them, which they call citation precision. The authors describe these results as “concerningly low for systems that may serve as a primary tool for information-seeking users.” A citation does not prove that the engine used a source, source use does not guarantee a citation, and even a displayed citation may be incorrect.

Aggarwal et al. provides a useful contrast: its impression metric measures visibility, or whether content was used, rather than whether the source received credit. Liu et al. examine the gap between those events. For a critique of Aggarwal’s headline result of “up to 40%” improvement, see the paper entry (arXiv:2311.09735 · ACM DL).

5. Why GEO teams should track all three

Each outcome has a different metric, a different way to influence it, and a different business value. Combining them into one measure distorts the performance of a GEO program.

Industry tools already measure these outcomes separately. Otterly.AI’s KPI taxonomy defines Brand Mentions, Domain Citations, and Share of Voice as distinct metrics with separate formulas (see Brand Report KPI Definition). In practice, whether an answer names a brand is measured separately from whether it cites the brand’s domain.

An unlinked mention is not a failed citation. It is a different result whose value often accumulates more slowly. Tracking only citations that contain links understates the value of strengthening the entity prior. Citable content may be necessary to earn credit, but it is not sufficient on its own, and each form of credit produces a different kind of value.

6. How attribution differs across platforms

Citation, mention, and link remain distinct across platforms. What changes is how often each form appears and how the platform usually presents it.

Attribution density can also vary by language, which matters for multilingual GEO.

7. Common reporting mistakes

These common reporting mistakes all treat one form of credit as if it were another:

You cannot optimize for credit directly. Improve the two conditions that precede it: groundability and the off-site prior created by mentions. Attribution follows both.

8. What to do next

The answer loop can produce three forms of credit. Credit does not necessarily follow source use, and it is not always reliable. Measure citation, mention, and link separately, then choose the next resource according to the result you want to improve.

References

Academic:

• Liu, N. F., Zhang, T. & Liang, P. (2023). Evaluating Verifiability in Generative Search Engines. Findings of EMNLP 2023. arXiv:2304.09848
• Aggarwal, P., Murahari, V., Rajpurohit, T., Kalyan, A., Narasimhan, K. & Deshpande, A. (2024). GEO: Generative Engine Optimization. KDD ‘24. arXiv:2311.09735 · ACM DL · paper summary

Industry / tooling:

• Otterly.AI: Definition of Brand Report KPIs

Official platform documentation (as of 2026-05):

• Google: Grounding with Google Search (Gemini API) · Google Search Central: AI features and your website
• Anthropic: Web search tool
• OpenAI: ChatGPT search
• Perplexity: What is an answer engine, and how does Perplexity work as one?