Citation Share: The New Share of Voice for SEO, GEO and AEO

Classic share of voice combines three signals — impressions, clicks and rankings. All three are eroding at the same time, on the same page, for the same query. The metric hasn't lied to us; the surface has moved. What replaces it is citation share: whether AI engines cite your domain when buyers ask real questions in your category. This article lays out the evidence, defines the metric precisely, and shares what we found running 25 companies through it.

Two numbers to start with

Ahrefs, using 300,000 keywords in Google Search Console, measured position-1 click-through rate at 58% below forecast when an AI Overview appears in the results (December 2025 data, published February 2026). The Japanese cut of the same study moved from −38% in December 2025 to −62.7% in June 2026. The trajectory is getting worse.

Pew Research, watching real US search sessions in July 2025, found that users who saw an AI summary clicked a search result in 8% of visits, against 15% without one. 26% of those sessions ended on the results page, against 16% without the summary.

Ranking is decoupling from being chosen. The answer itself is the surface most users touch, and the link sits underneath it. Any share-of-voice number built on impressions and clicks is counting the box traffic used to flow through, while users open a different box.

The old share of voice is measuring an emptying box

Position one no longer means a visit. The Ahrefs February 2026 update showed CTR falling at every position when an AI Overview appears: 58% down at position one, 50.8% at two, 46.4% at three, 38.8% at four, 32.6% at five. Seer Interactive reports between 49.4% and 65.2%. Kevin Indig reports over 50%. Authoritas 47.5%. Different samples, same direction.

A visit no longer means the content is read: when Google composes an AI Overview above your listing, the user often has the answer before scrolling. And a read no longer means being chosen — if the AI response names three vendors in its own paragraph, the buyer may act on that shortlist without clicking anywhere. The signal we still count as share of voice describes a surface that fewer people are touching.

Every serious vendor is shipping the same primitive

The tooling market has already converged, under different names. Ahrefs Brand Radar defines AI share of voice as the percentage of brand impressions you own versus competitors across ChatGPT, Perplexity, Gemini, Copilot and Google AI Overviews — and splits mentions (the name in the text) from citations (the source link) as separate primitives. Semrush normalizes visibility to a hundred percent within a category. Profound separates share of voice (mentions vs competitors) from citation share (how often your brand is cited). Peec tracks mention rate, citations, position and sentiment per prompt, reporting citation rate and mention rate as different rows.

The wedge is real and worth keeping: a brand mentioned without a source link contributes to awareness; a domain cited without a brand mention contributes to authority. Both matter, and both belong on their own row in the report.

What citation share actually is

For working purposes I use one operational definition: citation share is the percentage of AI responses to a fixed set of buyer questions in your category where your domain is cited as a source, relative to all cited domains in that question class, per engine, tracked over time.

At CitePulse we compute it deterministically. We string-match on the domain inside the answer's source block, and on the brand name inside the answer text. We do not ask the model how it feels about the brand, and we do not infer a ranking. We count. That matters because a number you can reproduce next month is the only kind you can build a decision on.

The empirical spine: off-domain beats on-domain, measured

The claim that off-domain signals matter more than on-domain signals is no longer a hunch. Ahrefs correlated common SEO variables with AI citations across 75,000 brands. Brand web mentions correlated at roughly 0.664. Branded anchor text at 0.527. Brand search volume at 0.392. Domain Rating at 0.25. Backlinks at 0.10 — the weakest of the five predictors. Off-site brand signals beat classic SEO signals, on the same brands, in the same study, at the same time.

In the same body of work, 91% of AI answers cite third-party sources. Only about 9% of brand mentions in AI responses come from the brand's own site. Whatever “AI optimization” means in practice, nine times out of ten the answer is assembled from somebody else's page.

Which somebody else? Similarweb's most-cited-domains study for the US in January–February 2026 puts Wikipedia at 13.15% of ChatGPT citations and Reddit at 11.97%; OpenAI's own properties at 6.21%, Walmart 2.90%, YouTube 2.67%, LinkedIn 2.42%. Google AI Mode has a different leader: Fandom at 7.16%. On review-driven queries, the AmICited dataset shows G2 alone at 22.4% of review-based citations, and the top five review platforms — G2, Capterra, Gartner Peer Insights, TrustRadius and Software Advice — at 88% combined. This is the measured shape of the field, and it is the base rate any strategy has to answer to.

Different engines, different games

Glen Allsopp ran the top 1,000 cited sites per engine through six months of Ahrefs Brand Radar data. His count, reported by Tim Soulo in August 2026, is blunt: last month ChatGPT cited 80 spammy sites, Perplexity 33, Copilot 122 — and Google AI Overviews cited zero. Manual review of those sites turned up fake or missing authors, thousands of AI-generated articles with no coherent topic, and e-commerce stores with a hundred thousand random items.

The story lives in the index layer, not the model. Google AI Overviews sits on top of two decades of anti-spam infrastructure — SpamBrain, manual actions, link-graph analysis. ChatGPT, Perplexity and Copilot draw from different retrieval layers with different filtering histories. The output reads as a model problem because a model wrote the sentence; the upstream cause is which pages the retriever surfaced in the first place. An arXiv paper on 761,495 citation pairs across ten search-augmented models adds numbers: 30.6% of citations distort their sources, 27.1% originate from domain-inappropriate sources, and up to 96% of users encounter at least one structurally misleading citation per session.

The practical consequence: there is no single “optimize for AI search” play that survives contact with all four engines. The source types cited in Google AI Overviews differ from those in ChatGPT, and neither matches Perplexity or Copilot. Priorities have to be set per engine.

The llms.txt lesson: measure before you invest

Ahrefs studied 137,000 domains in June 2026. 28% had published an llms.txt file. 97% of valid files received zero requests in May, and 96% of the requests that did arrive came from bots, not from language models pulling grounding data. SE Ranking's 300,000-domain study found no statistically significant effect of llms.txt on AI citations. Google is on the record — John Mueller, Gary Illyes, and Google's May 2026 optimization guide all say Search does not use llms.txt, and file it under “myths.” We detect the file, and we refuse to score it; the full reasoning is in llms.txt: what it is, who actually reads it, and whether you need one.

Thousands of teams invested in a signal without checking whether the engine consumed the signal. Citation share is that check. If your citations don't move after a change, the change didn't work — whatever the tactic promised.

What we found: 25 companies

We ran 25 companies through the same methodology — about 45 real buyer questions per company (not vanity brand queries), tested across ChatGPT, Perplexity, Gemini and Google AI Overviews, with deterministic detection of brand name in answer text and domain in cited sources, scored 0–100 across engines.

The median score was 12 out of 100. Eight companies scored zero. One category leader scored 70. The median for SaaS was 37. The median for SEO agencies selling AI visibility as a service was zero. Four stories from the set make it concrete:

• Popsa, a consumer photo-book app, scored 2; Shutterfly was cited on 11 of 45 category buyer questions.
• Archy in dental software scored 7; Dentrix, the legacy incumbent, was cited on 10 of 44 — including “how much can I save switching from Dentrix,” a query a challenger should own.
• An agency at seoagencyinessex.co.uk scored 0; competitor seoessex.io was cited on 6 of 45, including exact-brand queries the domain is named after.
• The leader, SmartMoving, scored 70 — with Movegistics already cited on 12 of 44 and closing.

Two of the 25 categories — dental marketing and Gloucester local SEO — had no vendor owning the answer at all. That is an empty seat: first-mover opportunity, no incumbent to unseat. Twenty-five companies is a small sample; the value sits in the method, not the count. It is cheap, repeatable, and any team can run it on their own domain and get a number that means the same thing next month.

What this means for SEO, GEO and AEO practice

SEO is not going away — its role is shifting from a channel to a grounding layer. Rankings still build the authority that gets a domain surfaced in the first place, and your content still needs to be citable. The change is downstream: ranking is now the input to somebody else's answer, not the endpoint of a user journey.

GEO and AEO are largely off-domain work. 91% of AI answers cite pages that are not yours. If your AI-visibility portfolio is 90% on-page changes and 10% off-domain, the portfolio is inverted. Review platforms carry 88% of review-based citations from five destinations; Wikipedia, Reddit threads that rank in Google, YouTube, LinkedIn long-form and a few publisher relationships do most of the visible lifting.

Per-engine strategy is not optional. Fandom leads Google AI Mode; Wikipedia and Reddit lead ChatGPT; the spam-tolerant engines cite thin sites Google won't touch. Optimizing as if all four engines share one taste produces work that fails halfway on every engine.

Three metrics deserve permanent dashboard space: citation share per engine; the top five cited sources in your category, so you know where to spend; and the distribution of your brand versus top competitor versus everyone else across the answer set. These are exactly the numbers a free CitePulse audit returns, and the reasoning behind the score is in our guide to generative engine optimization.

Honest caveats

Twenty-five companies is a small sample. Citations shift day to day with retrieval and temperature. No engine publishes its source weights. Nobody — not OpenAI, not any agency, not I — can guarantee that a specific brand will be cited by a specific model on a specific query. What is different now is that measurement is possible, and measurement is what turns a hype cycle into a discipline.

FAQ: citation share and AI visibility

What is citation share?

Citation share is the percentage of AI responses to a fixed set of buyer questions in your category where your domain is cited as a source, relative to all cited domains in that question class, per engine, tracked over time. It measures the surface users actually touch — the answer — rather than impressions, clicks and rankings on a page they increasingly skip.

How is it different from classic share of voice?

Classic share of voice combines impressions, clicks and rankings, all three eroding on the same page for the same query: position-1 CTR runs 58% below forecast under an AI Overview (Ahrefs), and real clicks roughly halve when a summary is shown (Pew). Share of voice built on those signals counts a box fewer people open. Citation share counts the answer itself.

Is off-domain or on-domain work more important?

Off-domain, on the measured evidence. Across 75,000 brands, brand web mentions correlate with AI citations at ~0.664 and backlinks at only 0.10; 91% of answers cite third-party sources and ~9% cite the brand's own site. A 90%-on-page budget is inverted relative to where citations come from.

Does llms.txt improve my AI citations?

No measured effect. 97% of published llms.txt files got zero requests (Ahrefs, 137,000 domains), SE Ranking found no statistically significant effect, and Google Search does not use it. Treat it as agent-readiness hygiene, not a citation lever.

Can anyone guarantee my brand gets cited by AI?

No. Citations shift day to day, no engine publishes its weights, and no one can guarantee a specific citation. What you can do is measure citation share on a fixed cadence and let it tell you whether a change worked.

Related reading: what generative engine optimization is and how to track it · how to measure AI visibility across ChatGPT, Perplexity, Gemini and Claude · the llms.txt reality check.