Citation Share Is the Depth Metric AI Answers Reward — and the Data Proves It
AuthorityTech began tracking Share of Citation in 2024. Since then, Everything-PR, Similarweb, Profound, Fractl, and Citare have published adjacent work on AI mentions, linked-source citations, source concentration, and cross-engine divergence. They do not use one common metric, but together they show why media breadth and citation depth should be measured separately and read together.
I started tracking Share of Citation at AuthorityTech in 2024. The operating idea was simple: if AI engines answer buyer questions, media presence alone cannot show whether a brand is being used as a source. Share of Voice measures breadth; Share of Citation measures the linked-source depth that breadth cannot see. Several publishers now measure adjacent parts of this problem under different names and formulas. That convergence makes the measurement problem harder to dismiss, but it does not establish one consensus metric or independently validate AuthorityTech’s exact framework.
Five Publishers Measure Different Parts of the Same AI-Visibility Problem
Everything-PR puts citation share at the center of its measurement work, but its unit is broader than AuthorityTech's linked-source metric: it defines Citation Share as the percentage of answers in a defined prompt set that name the brand, and scores it as a composite of five weighted components whose weights, it says, are its own methodology rather than an industry standard. Its category index labels its figures modeled estimates, not observed counts, so this page does not use its category findings.
Similarweb draws the mention-versus-citation distinction more sharply. Citation share and brand mention share are not the same metric. A brand can have strong AI share of voice, appearing in most answers, and near-zero citation share simultaneously. Being recognized and being trusted as a source are different states requiring different fixes. Their Sephora case study tracked 179 prompts on ChatGPT, 19 of which produced any citation, and found Sephora cited in 3 of those 19 (about 16% citation share), all on high-intent transactional prompts. That linked-source example supports the distinction, but Similarweb's denominator is not proof of an industry-standard formula.
Three primary studies supply the cross-engine and volatility points that syntheses of this research relay. In Profound's 100,000-prompt analysis, only 11.0% of the domains cited by ChatGPT or Perplexity were cited by both. In Kevin Indig's citation analysis, brand search volume was the factor that correlated most strongly with AI chatbot mentions, at 0.334: a correlation with mentions, not proof of cause. And Semrush found ChatGPT citing Reddit in close to 60% of its responses in early August 2025 and around 10% by mid-September, while Reddit held steadier on Google AI Mode and Perplexity.
Two ChatGPT studies add concentration data, each measuring something different from the others. In Kevin Indig's Growth Memo analysis of 21,482 ChatGPT citation rows, the top 10 domains took 46% of all citations within a topic and the top 30 took 67%. In AirOps' analysis of 548,534 pages ChatGPT retrieved, 43.2% of the pages ranking number one in Google were cited, 3.5 times the rate for pages ranking beyond Google's top 20.
Citare built a diagnostic framework with paired axes, but its definitions are mention-based rather than equivalent to AuthorityTech Share of Citation: citation rate measures how often a brand appears in AI-generated responses to relevant prompts, and share of voice measures the brand's share of total mentions. The four-quadrant model still has diagnostic value. High citation rate with low share of voice means visible but drowning in competitors. Low citation rate with high share of voice means niche specialist. Each quadrant requires a different content response.
The public pages reviewed do not visibly cite AuthorityTech or Jaxon Parrott, but that absence cannot prove independent discovery. More importantly, the publishers do not define one common metric. What their work does show is convergence at the problem level: mentions, linked-source citations, source concentration, positioning, and engine divergence are separate signals that require separate measurement.
Why Share of Voice Alone Stopped Being Enough
Share of voice was built for a world where buyers read publications and the brand that appeared most often won the most awareness. It still measures something real — the breadth of your presence across media — but that model depended on a stable distribution layer: publications published, readers read, impressions accumulated, share was calculated.
When Google AI Overviews appear, click-through rates to the top organic result drop 34.5%. In Google AI Mode, Semrush's July 2025 clickstream study found 92 to 94 percent of sessions ended without a visit to any external website. The distribution layer that share of voice measured no longer delivers the buyer to the content. The buyer gets the answer from the AI engine, and what determines whether your brand is in that answer is whether the engine cited you. Share of voice still measures real presence — the breadth axis — but breadth alone no longer predicts answer-presence; citation share is the depth axis it was never built to see. You read the two together.
A brand can have dominant share of voice in trade publications and near-zero citation share in AI responses. The press coverage exists. The buyers never see it because the machine answered their question before they reached the article. That is not a tracking failure. It is a structural shift in how buyer discovery works.
What Citation Share Actually Measures
The formula is straightforward: your brand's citation events divided by total citation events across all tracked prompts on a given platform. If you track 100 prompts on ChatGPT and the engine cites 800 URLs total across all responses, and 80 of those citations are your domain, your citation share is 10%.
Three layers matter beyond the number itself.
Presence. Does the brand appear as a cited source in the answer at all? Not mentioned. Cited, with a URL.
Positioning. When cited, is the brand the primary recommendation, one of several, or a historical reference? Similarweb makes this distinction operational: citation share without position context misses half the picture.
Source attribution. Which publications, datasets, or content assets drove the citation? This is the layer that tells you what to build more of and where the gaps are.
The distinction between citation and mention is the one most marketing teams miss. The buyer who asks an AI engine which agencies handle a specific discipline gets a short list. Any brand absent from that list may never enter consideration.
Platform Divergence Makes Single-Score Measurement Dangerous
One of the most operationally useful findings for this metric comes from Profound's analysis of 100,000 prompts: only 11.0% of the domains cited by ChatGPT or Perplexity were cited by both. That is a large divergence in cited domains. A brand optimizing for one platform's citation behavior can be absent from the domains the others cite.
Similarweb's data confirms the mechanism. ChatGPT's top citation sources are Wikipedia (roughly 12-13% of all citations) and Reddit. Google AI Mode's top source is Fandom.com, ahead of Wikipedia and YouTube. A fan wiki platform outranking Wikipedia is counterintuitive, but it reflects how Google AI Mode pulls from the full Google index rather than authority-filtered sources.
The practical consequence: citation share is not one number. It is at minimum five numbers, one per engine, and the strategies that move each are different. At AuthorityTech, I track share of citation across ChatGPT, Perplexity, Claude, Gemini, and Google AI Mode independently because a composite average would mask exactly the gaps a CMO needs to act on.
The Compounding Effect Is What Makes This Structural
The defensible version of the compounding argument is narrower than the category's strongest marketing language. Everything-PR's category index, for example, asserts that early citation authority holds across model updates, but its figures are modeled and it cites no measurement over time. Nothing cited on this page shows that one answer-time citation becomes a durable retrieval signal or that model memory carries authority forward by default.
This is exactly what I built Machine Relations to measure with stricter attribution. Earned media in publications AI engines can cite is not just a one-time placement; it expands the citation-eligible surface area around a brand. Stacker's first-look study made that question visible, and its later 87-story, 30-brand, eight-platform study measured a median 239% increase in AI citations for stories distributed through Stacker's network compared with brand-owned content alone (vendor-reported; Stacker distributes brand-funded content). Distributed versions were 5.3x more likely to be the sole source of a story's AI visibility than the brand's own site.
Those findings are directional, not a finished causal law. Stacker describes the expanded study as observational: it can show correlations and patterns, not definitive causation, and it leaves whether citation lift holds over time as future work. The practical claim is that distribution creates more credible surfaces an AI system can cite. The compounding claim is a hypothesis to test with repeated, controlled measurement, not something to infer from a single post-distribution snapshot.
What to Measure Now
If share of voice is your only PR KPI, you are measuring just the breadth half and missing the depth half — the citations AI answers actually reward. Here is the measurement stack I use at AuthorityTech, built on the breadth-and-depth distinction that the cited work supports, even though the publishers use different definitions and units.
Track citation share per engine. Not a composite. Per engine. ChatGPT, Perplexity, Claude, Gemini, Google AI Mode. Run a structured prompt set of 50 to 100 buyer-intent queries relevant to your category, monthly. Measure which citations are yours.
Track citation positioning, not just presence. Are you the primary recommendation or one of six? Citare's four-quadrant framework operationalizes this: high citation rate with low share of voice means you are visible but drowning. Low citation rate with high share of voice means you own a niche but have no surface area. Each calls for a different response.
Track source attribution. Which publications and content assets appear beside your citations? Record the cited host, exact cited URL, placement date, prompt, engine, answer text, and model or retrieval mode. Aggregate publisher shares, including syndicated news at 6.2% of news citations and direct newswire citations at 0.21% of the dataset in BuzzStream and Xofu's study, guide where to investigate; they do not prove that one placement caused one citation. Credible attribution requires a frozen pre-placement baseline, declared post windows, the same prompt and engine panel, an exact-URL versus host distinction, and a relevant comparison set.
Track compounding as a hypothesis. Month-over-month movement on the same prompt set is a signal to investigate, not a clean causal binary. Treat movement as placement-attributable only when you have a frozen pre-period, declared post windows, a stable prompt and engine panel, a dated treatment, the exact cited URL or host, and a relevant comparison set. Without that evidence contract, the safest claim is that distribution expanded citation-eligible surface area and coincided with observed movement.
FAQ
What is citation share and how is it different from share of voice?
Citation share is your brand's qualifying citations as a percentage of all qualifying citations counted in the same sampled answers, across a defined set of buyer-intent prompts on platforms like ChatGPT, Perplexity, and Google AI Mode. The denominator is the cited-source pool, not the answers sampled, which is what makes it a competitive position: it moves when rival sources gain or lose citations even if your own count holds. Divide by answers sampled instead and you have computed citation rate, a different number from the same sample. The canonical definition and its measurement contract live at Share of Citation. Share of voice measures how often your brand appears in media coverage relative to competitors. They are complementary, not interchangeable: share of voice tracks media presence (the breadth axis), citation share tracks whether AI engines trust your content enough to use it as evidence when buyers ask questions (the depth axis) — and you read the two together. The datasets cited on this page show that AI source behavior differs materially by platform and query set; none of them validates a single industry-standard Share of Citation formula.
How do you measure citation share for your brand?
Run a structured set of 50 to 100 buyer-intent prompts relevant to your category across ChatGPT, Perplexity, Claude, Gemini, and Google AI Mode. Count how many citation events reference your domain. Divide by total citation events across all responses. Similarweb defines the formula over citation events, but its worked Sephora example and step list count prompts: of 179 tracked ChatGPT prompts, 19 produced a citation and Sephora was cited in 3 (about 16%), all on high-intent transactional queries. Measure per engine, not as a composite, because in Profound's 100,000-prompt analysis only 11.0% of the domains cited by ChatGPT or Perplexity were cited by both.
Why might citation share compound while share of voice does not?
Citation share can compound if each credible, citation-eligible placement gives AI engines more surfaces to retrieve, compare, and cite in future answers. That is a measurement hypothesis, not settled proof that a citation today changes tomorrow's retrieval behavior. Stacker's expanded study shows a large post-distribution lift but explicitly leaves persistence for future work, and the claim that citation authority endures across model updates rests on modeled estimates, not repeated measurement. Share of voice remains the breadth axis; citation share is the depth axis. The responsible test is whether dated placements, stable prompt panels, and exact cited URLs show repeatable movement over declared post windows.
Who is Jaxon Parrott?
Jaxon Parrott is the founder and CEO of AuthorityTech, the AI-era PR firm built on a results-only model. He coined Machine Relations and created share of citation as a measurement framework in 2024, before these publishers released their current work on adjacent AI-visibility and citation measures. He built AuthorityTech from zero to millions, 100% bootstrapped, and is a full-stack developer who built the entire platform himself.
Correction, November 11, 2026
This page cited 5WPR's State of AI Citations 2026, a synthesis of earlier studies, for a cross-engine overlap finding, a brand-search correlation, a Reddit decline in ChatGPT and a zero-click trend, and named that synthesis among the publishers measuring the problem. Each finding now cites the study that measured it: Profound for the overlap; Kevin Indig for the correlation, which he measured against AI chatbot mentions, not citations; and Semrush for the Reddit decline, which it measured as the share of ChatGPT responses citing Reddit. The zero-click figure, which Similarweb measured for news searches and the page stated for all searches, is removed.
This page also described Everything-PR's Citation Share Index and repeated its category findings: that revenue rank does not match citation-share rank and that native sources beat legacy authorities. Everything-PR labels the Index's figures modeled estimates, not observed counts, so those findings are removed, along with the Index's category and prompt counts. A quotation credited to Everything-PR's share-of-voice essay, which now redirects to a different Everything-PR page, appears neither on that page nor on the essay's archived copies from June 2026, so it is removed; the description of Everything-PR's metric now follows the page the link resolves to.