Known but Not Cited: Why AI Share of Voice Is the Wrong GEO Metric

Overview

Most GEO dashboards still report a single number: how often the brand appears in AI answers. That number is AI share of voice. It is a recognition metric. It tells you the model knows the name.

It does not tell you whether the model trusts the page.

Similarweb’s June 2026 analysis, What Is AI Citation Share and Why Does It Matter for GEO? (Maayan Zohar Basteker), separates the two. AI citation share is the percentage of citation events, across a fixed prompt set, in which your domain is used as a linked source. A brand can dominate mentions and still be almost absent from the citation graph. That pattern — recognised, not cited — is currently one of the most common misdiagnoses in generative-engine optimisation.

This matters commercially. McKinsey’s 2025 work (already in this library) estimated that unprepared brands risk losing 20–50% of traditional search traffic as $750 billion in US revenue moves through AI-powered search by 2028. Citation, not mention, is what sits closest to that traffic: a linked source is the residual click. A name-drop is not.

The Study

Basteker reports two empirical claims from Similarweb’s AI Search Intelligence stack.

First, a definitional split, illustrated with a worked Sephora example: 179 ChatGPT prompts, of which 19 produced at least one citation event, of which Sephora was a cited source in 3 — ~16% citation share on the cited-prompt base. All three were high-intent transactional queries (“where can I buy…”). Informational and inspiration queries mentioned beauty retail without citing Sephora.

Second, a source-pool split from approximately 600,000 US citation events: ChatGPT and Google AI Mode do not draw from the same web.

| Engine | Leading cited domains | |---|---| | ChatGPT | Wikipedia and Reddit, each ~12–13% of citation events | | Google AI Mode | Fandom.com first, then Wikipedia, then YouTube |

Fandom outranking Wikipedia on Google AI Mode is not a curiosity. It is a retrieval-index fact: Google AI Mode can cite the full Google index, including dense community wikis; ChatGPT’s citation graph is still dominated by encyclopedic and forum priors.

This article does not replace the scientific GEO synthesis already on LLM Wisdom. It adds the 2026 measurement layer that synthesis was missing: what to count, per engine, once you have accepted that citation is the unit of competition.

What They Found

1. Mention and citation diagnose different failures

AI share of voice = how often the brand name appears, relative to competitors, across a prompt set. Recognition.

AI citation share = your citation events ÷ all citation events on that same prompt set. Trust.

Similarweb’s worked formula: 100 ChatGPT prompts produce 800 cited URLs; your domain accounts for 80 of those events → 10% citation share. The denominator is all citations, not all prompts. Prompts that cite nobody do not enter the share calculation in the Sephora walkthrough; the operational question is “of the answers that bother to cite, whose URL is used?”

A high-SoV / low-citation pattern means the model can talk about you and still refuse to send a user to you. Content, structure and third-party corroboration are the usual causes — not awareness.

2. Citation is concentrated

In competitive categories, a small set of domains take most citation events. Most brands that appear in AI answers at all are mentioned far more than they are cited. That is why 16% in beauty, on ChatGPT, is described as a defensible position rather than a rounding error: Sephora is cited in roughly one in six cited prompts, and those prompts are the ones that convert.

3. Citation share is engine-specific, so GEO cannot be platform-agnostic

A page that is structured for Google AI Mode (indexable, wiki-like topical depth, YouTube adjacency) will not automatically earn ChatGPT citations, where Wikipedia and Reddit dominate. The reverse is also true. Optimising “for AI search” as if it were one index is the 2012 mistake of “optimising for search” without distinguishing Google from Bing — except the source pools now diverge more sharply.

4. Intent splits the citation graph inside a single brand

Sephora’s three cited prompts were all purchase-intent. Informational prompts (“what are some luxury skincare products worth investing in?”) did not cite it. That is a content-type gap, not a brand-equity gap. Treating a blended citation-share number as a single KPI will hide the fact that you own commerce queries and have abandoned consideration queries — or the reverse.

Why This Happens

Generative engines do not rank ten blue links. They retrieve passages, synthesise an answer, and optionally attach evidence. Two selection processes run:

1. Entity selection — which brands belong in the answer. This is what share of voice captures. It is closer to co-occurrence, knowledge-graph membership, and training-data frequency.
2. Evidence selection — which URLs are allowed to stand behind a claim. This is what citation share captures. It is closer to extractability, original data, passage-level answerability, and the engine’s preferred source class.

Those processes can disagree. A famous retailer is an obvious entity on “where do I buy fragrance.” The same retailer may not be an obvious source on “which vitamin C serums are worth it,” because the engine would rather cite a wiki, a review round-up, or Reddit.

The engine split follows from index design. ChatGPT’s browsing and citation stack still overweight a small set of high-prior domains. Google AI Mode can, and does, surface whatever the Google index already considers canonical for a fandom-dense or video-dense query. You cannot win both with one page template.

Concrete Recommendations

1. Split the dashboard

Report four numbers, weekly, on a frozen prompt set, across ChatGPT, Perplexity and Google AI Mode / AI Overviews:

• AI share of voice (mentions)
• AI citation share (linked sources)
• Citation gap (prompts where a competitor is cited and you are not)
• Intent mix (what % of your citations are transactional vs informational)

If SoV is healthy and citation share is not, stop spending on “AI brand awareness” content. Fix extractability and third-party evidence.

2. Structure every key page for passage extraction

Similarweb’s own playbook, consistent with the GEO literature already in this library:

• Open each section with a standalone answer.
• Prefer original statistics with attribution over qualitative claims.
• Use tables, numbered procedures and comparison frameworks — they are retrieved disproportionately vs prose covering the same facts.

If the answer lives in paragraph five, it will be mentioned, not cited.

3. Build engine-specific source strategies

• ChatGPT: earn presence on Wikipedia-class pages and Reddit-class discussions in your category; treat those as citation multipliers, not vanity PR.
• Google AI Mode: treat community wikis, YouTube, and densely interlinked topical sites as first-class surfaces. Fandom leading the citation list is a product of index coverage, not a fad.

Do not average these into one “earned media” programme.

4. Run citation-gap analysis before you write more

The highest-ROI list is not “topics we have not covered.” It is “prompts where the engine already cites someone, and that someone is not us.” That list is finite, ranked, and falsifiable.

5. Do not use a universal benchmark

There is no good citation-share number. 16% is a reference point for a strong consumer brand in a crowded category on ChatGPT — not a target for B2B, local, or a monopoly category. Use the category leader on your prompt set as the ceiling.

How This Extends Prior LLM Wisdom

New Front Door to the Internet established the commercial stakes. Scientific and Data-Driven Foundations for LLM SEO and GEO established that citation, not ranking, is the competition. This piece is the measurement upgrade: share of voice is necessary and insufficient, citation share is the trust metric, and the citation graph is not shared across engines.

If your GEO report still has one column, it is out of date as of mid-2026.

Source

This article summarises and adapts findings from Similarweb (June 2026): What Is AI Citation Share and Why Does It Matter for GEO?, authored by Maayan Zohar Basteker, reviewed by Limor Barenholtz. Citation-pool statistics are from Similarweb’s analysis of approximately 600,000 US citation events.

Commercial stakes referenced from McKinsey (October 2025): New front door to the internet: Winning in the age of AI search.

Recommended Reading

• Similarweb: AI citation share
• Similarweb: ~600k citation-event engine split
• LLM Wisdom: New Front Door to the Internet
• LLM Wisdom: Scientific and Data-Driven Foundations for LLM SEO and GEO