---
title: "AI Citations vs. Brand Mentions: What 182,789 Answers Show"
canonical: https://snezzi.com/blog/ai-citations-vs-brand-mentions/
source: https://snezzi.com/blog/ai-citations-vs-brand-mentions/
published: 2026-10-09
modified: 2026-10-09
author: "Nikunj Thakkar"
category: "Industry Insights & Research"
---

> Canonical page: https://snezzi.com/blog/ai-citations-vs-brand-mentions/

# AI Citations vs. Brand Mentions: What 182,789 Answers Show

*Original research on why a source link and a place in the answer are different AI visibility metrics.*

Imagine asking an AI assistant for the best tools in a category. It links to a company’s guide, then names three other companies in the answer. The guide helped support the response. The company itself never made the list.

That is the gap we wanted to measure. A **citation** is a link to a brand’s own website in an AI answer. A **mention** is the brand’s name in the answer text. A mention is *not necessarily a recommendation*: it may be neutral, critical, or part of a comparison. But if a company is absent from the answer altogether, it cannot be recommended there.

We examined 185,912 valid answers recorded by Snezzi from November 2025 through September 2026 across five AI search surfaces. The main analysis removes prompts that already contained the brand name, leaving **182,789 answers and 9,974 brand–prompt–engine combinations**. These were real prompts selected for tracking, rather than a representative sample of every AI search.

The result: **46% of combinations in which a brand’s site was cited never produced an answer naming that brand.** Looking the other way, **74% of combinations in which a brand was named never cited its own site.** Citation and answer presence overlap, but neither stands in for the other.

![A four-cell map of brand site citation and brand mention across 9,974 unbranded combinations. 298 had both, 254 citation only, 839 mention only, and 8,583 neither.](https://images.snezzi.com/2026/3713x1mb3450ys9/914obh1t7f9949v/ref_mfvhB.webp)

*Figure 1. Outcomes across 9,974 brand–prompt–engine combinations. “Ever” means at least once during the study window, not necessarily in the same answer.*

## What the study counted

The tracker ran a fixed set of customer-facing prompts repeatedly across ChatGPT, Google AI Overviews, Google AI Mode, Perplexity, and Claude. Each **combination** is one tracked brand, one prompt, and one engine, observed across multiple answers. The prompts were selected for tracking, so this is a picture of that prompt set rather than a random sample of everything people ask AI.

We counted an **own-site citation** when a brand's registered domain, subdomain, or known alias appeared in the engine's recorded source list at least once. We counted a **mention** when its name appeared in answer text at least once. We also looked at answer-level mention rates, the best source-list position recorded for a combination, and the cumulative number of own-site citation events.

| Measure | What it tells us | What it does not establish |
| --- | --- | --- |
| Ever cited | The brand's own site appeared in a recorded source list during tracking. | That the answer named or endorsed the brand. |
| Ever mentioned | At least one answer named the brand. | That the name-drop was favorable or a recommendation. |
| Best citation position | The highest source slot the brand's site reached during tracking. | That this position coincided with a mention in the same answer. |

The wording matters. In the tracker, **61.7% of mentions were labeled neutral, 29.3% positive, 0.9% negative, and 8.1% unclassified**. A brand cannot be recommended in an answer that never names it, but a name-drop alone is not a confirmed recommendation. This is a study of the gap between *citations* and *answer presence*.

Citation records are also stored cumulatively per combination. They do not tie one source link to one individual answer. That limitation shapes every interpretation below.

## Do AI citations lead to brand mentions?

Among the 552 combinations where a brand’s own domain appeared in a source list, 298 eventually had at least one answer naming the brand. That is **54%**. Among the 9,422 combinations without a recorded own-domain citation, **8.9%** had at least one mention.

| Outcome across repeated answers | Combinations | Share of the full discovery set |
| --- | ---: | ---: |
| Own site cited and brand named | 298 | 3.0% |
| Own site cited, brand never named | 254 | 2.5% |
| Brand named, own site not recorded as cited | 839 | 8.4% |
| Neither outcome recorded | 8,583 | 86.1% |

Among combinations with a recorded own-site citation, **254 of 552** never named the brand: **46.0%**. Among combinations that named it, **839 of 1,137** had no recorded own-site citation: **73.8%**. This second statistic does not prove those mentions came from third-party sources; some answers may have shown no source list at all.

The relationship is strong. It is also easy to overread. Those “ever cited” and “ever mentioned” labels summarize a history of repeated answers. The citation and mention need not have happened together. Longer histories also provide more opportunities for both events. We cannot conclude that the citation *caused* the mention.

Looking at individual answers gives a more grounded sense of scale. Answers from combinations that had ever cited the brand named it **13.1% of the time**, compared with **2.8%** for combinations with no recorded own-domain citation. That is an association between two groups of prompts, not a test of what happens after a citation in a particular answer.

The source draft reports 95% Wilson intervals of **49.8%–58.1%** for the 54.0% combination rate and **8.3%–9.5%** for the 8.9% rate. Those intervals describe sampling uncertainty within the tracked combinations; they are not a margin of error for all brands or AI searches. Repeated answers and prompts from the same brands are related observations, so a broader population claim would need a more careful clustered analysis.

![Among combinations with a brand site citation, 54.0% ever mentioned the brand; among combinations without one, 8.9% did. At the answer level, the corresponding mention rates were 13.1% and 2.8%.](https://images.snezzi.com/2026/3713x1mb3450ys9/914obh1t7f9949v/ref_vEKSa.webp)

*Figure 2. Combination-level “ever mentioned” and answer-level mention rates. The answer groups are defined by whether their combination ever had an own-domain citation.*

## More AI citations did not reliably mean more mentions

The first recorded citation marked a large difference: **51.0%** of combinations with exactly one own-domain citation ever mentioned the brand, versus **8.9%** with none. But combinations with two citations were at **54.4%**, and those with three to five were at **51.5%**. The six-or-more group reached **65.0%**, based on just 80 combinations.

| Own-site citation events during the window | Combinations | Ever named the brand |
| --- | ---: | ---: |
| 0 | 9,422 | 8.9% |
| 1 | 200 | 51.0% |
| 2 | 136 | 54.4% |
| 3–5 | 136 | 51.5% |
| 6 or more | 80 | 65.0% |

![Share of discovery combinations that ever named the brand by cumulative own-site citations: 8.9% with none, 51.0% with one, 54.4% with two, 51.5% with three to five, and 65.0% with six or more.](https://images.snezzi.com/2026/3713x1mb3450ys9/914obh1t7f9949v/ref_hAOXL.webp)

*Figure 3. Citation events are cumulative across runs. The six-or-more group is small; its 65.0% rate should not be read as a smooth dose-response curve.*

This pattern does not show that getting one citation changes an AI system’s mind. Brands that are better known or more relevant to a prompt may simply be more likely to earn both citations and mentions. What it *does* show is that a rising citation count for the same prompt and engine is a poor substitute for measuring whether the brand enters the answer.

The original analysis also restricted cited combinations to those with at least eight runs. The rates in the one, two, three-to-five, and six-or-more citation groups remained broadly similar at **58.3%, 51.0%, 51.9%, and 66.7%**. That check reduces the simplest concern that long-lived combinations merely had more chances to accumulate everything. It still cannot isolate the effect of citations from brand recognition, query fit, or page quality.

## Where your site appears in AI citations matters

When a brand’s site reached the first position in an engine’s source list, **67.8%** of those combinations ever mentioned the brand. When its best position was ninth or lower, **37.2%** did. At the individual-answer level, mention rates for those two groups were **22.4%** and **5.6%**.

| Best own-site source-list position | Cited combinations | Ever named | Answer-level mention rate for the group |
| --- | ---: | ---: | ---: |
| 1st | 146 | 67.8% | 22.4% |
| 2nd–3rd | 134 | 51.5% | 13.2% |
| 4th–8th | 178 | 53.4% | 11.3% |
| 9th or lower | 94 | 37.2% | 5.6% |

That is a large difference, but the source position is the *best recorded across a combination’s history*. We cannot say a first-place citation in one answer produced a mention in that same answer. We can say that a top source position and answer presence tend to travel together in this dataset.

![Brand mention presence was 67.8% when the brand site's best source position was first, 51.5% for positions two to three, 53.4% for positions four to eight, and 37.2% for position nine or lower.](https://images.snezzi.com/2026/3713x1mb3450ys9/914obh1t7f9949v/ref_mnnM4.webp)

*Figure 4. Share of cited combinations that ever mentioned the brand, grouped by the brand site’s best recorded source-list position. Group sizes: 146, 134, 178, and 94.*

Position is the **best** slot achieved across the whole tracking history. One first-place citation in the source analysis still coexisted with no mentions across many runs. Prominence and answer presence tend to travel together, but a top citation is no guarantee.

## AI citation rates differ by search surface

Perplexity had the highest share of combinations with an own-domain citation in this dataset (**8.1%**). Yet only **47.1%** of those cited combinations ever named the brand. ChatGPT had the lowest own-domain citation share (**2.4%**), while **77.6%** of its cited combinations ever named the brand.

| AI search surface | Discovery answers | Mention rate per answer | Combinations with own-site citation | Cited combinations ever named |
| --- | ---: | ---: | ---: | ---: |
| ChatGPT | 48,857 | 3.3% | 2.4% | 77.6% |
| Google AI Overviews | 41,990 | 3.9% | 6.6% | 52.0% |
| Google AI Mode | 48,958 | 3.6% | 5.7% | 57.0% |
| Perplexity | 38,926 | 4.3% | 8.1% | 47.1% |
| Claude | 4,058 | 0.3% | 2.8% | 42.9% |

This is a reason to keep engine results separate. A higher citation rate on one platform does not automatically mean stronger brand presence in its answers. Claude’s sample started later and included only 14 cited combinations, so its estimate is especially uncertain.

The four larger surfaces had answer-level mention rates in a narrower **3.3%–4.3%** band despite their different own-site citation rates. Engines can assemble and display sources differently, and the prompt mix tracked on each surface may differ too. The table describes the observed pattern, not an engine ranking or a causal conversion rate.

![Own-domain citation rate and the share of cited combinations that ever mentioned the brand vary by engine. ChatGPT 2.4% and 77.6%; Google AI Overviews 6.6% and 52.0%; Google AI Mode 5.7% and 57.0%; Perplexity 8.1% and 47.1%; Claude 2.8% and 42.9%.](https://images.snezzi.com/2026/3713x1mb3450ys9/914obh1t7f9949v/ref_NbJ0b.webp)

*Figure 5. Grey shows the share of combinations with an own-domain citation. Green shows the share of those cited combinations that ever mentioned the brand. These are different denominators. Claude is shown for completeness, with a small-sample caution.*

## The question itself changes the odds

For “best of” prompts, **19.3%** of combinations ever named the tracked brand. For informational prompts, it was **8.8%**. A shopping or shortlist question creates more room to name brands than a general explanation does.

The original analysis assigned prompts to five groups using their wording. Those labels are useful approximations of intent, not a perfect account of why a person asked the question.

| Prompt type | Combinations | Ever named, overall | Ever named if own site cited | Ever named with no recorded own-site citation |
| --- | ---: | ---: | ---: | ---: |
| Informational | 6,847 | 8.8% | 49.1% | 6.4% |
| Best-of or list | 2,362 | 19.3% | 71.2% | 16.4% |
| Comparison | 383 | 9.9% | 22.2% | 9.3% |
| Alternatives | 212 | 13.2% | 75.0% | 9.5% |
| Price-focused | 170 | 7.6% | 66.7% | 5.5% |

The comparison group deserves restraint. Its cited subset was small, and the original analysis did not find a statistically clear citation–mention relationship there. It may be that comparisons draw on names supplied by users or on third-party coverage, but this dataset cannot establish that explanation.

That matters when reviewing a dashboard. If the prompt mix changes, the overall mention rate can change even when a brand’s underlying visibility has not. Report discovery prompts by intent and keep prompts that already name the brand in a separate bucket. In this dataset, branded prompts mentioned the brand in roughly 99% of answers regardless of whether its site was cited. They are useful for studying how AI describes a brand, but tell us little about discovery.

## What a mention without an own-site citation means

The 839 combinations in which the brand was named without a recorded citation to its own site are easy to overinterpret. Some answers may have linked to outside pages. Others may have displayed no sources, and some source records may be missing. The data show that a brand can enter an answer without its own site appearing in the recorded source list; they do not show which outside websites, if any, made that happen.

To understand a specific prompt, read the answer and the displayed sources together. Check whether the brand is mentioned favorably, neutrally, or critically, and whether the cited pages actually discuss it. That review can generate a useful content or outreach hypothesis, but this study did not test which source types cause AI systems to mention a brand.

## How to measure AI citations and brand mentions

Start with three questions for each discovery prompt and engine:

1. **Did the answer name us?** Count the share of individual answers that do, and review whether the mention is favorable, neutral, or negative.
2. **Did it cite our own site?** Track citations and source position, while remembering that a source link is not an endorsement.
3. **Which third-party pages appear when we are named or absent?** Read those answers and sources. They can reveal where buyers are learning about the category and how the brand is framed.

The four outcomes in Figure 1 make a useful review queue:

1. **Cited and named:** inspect how the brand is described. A neutral comparison is different from an explicit recommendation.
2. **Cited but unnamed:** read the cited page and the answer together. Does the page clearly connect the brand to the problem the prompt asks about?
3. **Named but own site uncited:** inspect the displayed outside evidence and check whether it describes the brand accurately. Some answers have no displayed sources, so keep that possibility separate.
4. **Neither:** decide whether the prompt represents a buyer question worth pursuing before investing in content or outreach.

These are diagnostic starting points, not proven fixes. This observational study did not test whether an on-site rewrite, a comparison page, digital PR, or any other intervention increases favorable recommendations.

Snezzi's [AI citation tracking](https://snezzi.com/features/ai-citation-tracking/) and [AI visibility tracking](https://snezzi.com/features/ai-visibility-tracking/) pages explain how to review those two signals separately.

The clearest lesson is simple: **a link to your site and a place in the answer are different outcomes. Measure both.** A dashboard that reports citations alone can make a brand look visible while missing the moment a reader asks, “Which company should I consider?”

---

### How we measured this

**Observation window and unit.** The study includes 185,912 valid tracked answers from November 2025 through September 2026. Excluding prompts that contained the tracked brand's name leaves 9,974 brand–prompt–engine combinations and 182,789 answers. A combination was observed repeatedly, usually on a recurring schedule. Customer identities, counts, prompt text, and brand-level results are withheld.

**Matching.** Own-site citations were matched to registered domains, subdomains, and known aliases, then checked against the tracker's own-source classification. Mentions were detected from answer text. Incomplete aliases or a rebrand can produce small undercounts. Mention tone is a tracker classification; it has not been independently hand-labeled for this publication.

**Timing.** Citation storage is cumulative by combination, including usage count and best source-list position. It does not identify the exact answer in which a specific source was cited. “Ever cited, ever mentioned” can therefore overstate same-answer overlap, especially for combinations with many runs. The answer-level mention rates in Figures 2 and 4 partly address unequal exposure time, but those answers are still grouped by cumulative citation history.

**Missing sources and classification.** 23.5% of combinations had no stored citation rows. Some engines omit source lists, so “no recorded own-site citation” is more precise than “the brand was never cited.” The source records do not support a reliable attribution of mentions to particular kinds of outside websites.

**Scope of inference.** Tracked brands, selected prompts, engines, and repeat runs are not independent random draws. Brand recognition, query relevance, category competition, and content quality could influence both outcomes. These percentages describe the tracked set; they do not establish causality or predict the result of a particular optimization.

*Research by Snezzi. [Google’s documentation on AI features](https://developers.google.com/search/docs/appearance/ai-features) provides background on how links can appear in AI Overviews and AI Mode; it does not establish the causal interpretation of our data.*