AI Citation and LLM-SEOStatistics 2026

On the FORKOFF index, Perplexity leads at a 48 percent cite rate, followed by Google AI Overviews at 35 percent, ChatGPT at 32 percent, Claude at 29 percent, and Gemini at 26 percent, for a 34 percent average. That ordering is not unique to FORKOFF's data. Qwairy's Q3 2025 study of 118,101 answers measured how many citations each engine emits per answer and found Perplexity averages 21.87 citations per question, about 2.76 times ChatGPT's 7.92. Google AI Overviews sits at 17.93, Gemini at 17.11, and Microsoft Copilot last at 2.47. The two methods measure different things, one a hit rate for a single domain and one a raw volume across all answers, yet they agree on which engines are generous with citations and which are stingy.

The mechanism is source-pool width. Perplexity live-indexes the web and surfaces the widest set of sources per answer, so a well-formed page clears its citation floor more often. Engines that answer from a narrower internal pool cite fewer domains, and a single domain has to fight harder to appear. This is why the same remediation work produced the largest first-party movement on Perplexity, plus 17 points between runs, and the smallest on Gemini, plus 7 points. The engine you are trying to earn a citation from determines both the ceiling and the lever that moves it.

The buyer takeaway is that an averaged AI-visibility number is close to meaningless. A 34 percent average hides a 22-point spread between the best and worst engine, and the spread on the open web is wider still. AuthorityTech's 21,143-citation audit measured a Perplexity brand cite rate near 13.05 percent against a ChatGPT rate near 0.59 percent, a 46 times gap on the same brands. Any credible measurement reports per engine, because the practical work of earning a citation differs by engine.

One caveat keeps the ranking honest. Citation volume is not the same as citation quality: Microsoft Copilot emits only 2.47 citations per answer and draws from just 111 distinct domains, which looks stingy, but a narrow, high-trust pool can be easier to enter for an already-authoritative brand than a wide pool where the same slot is contested by thousands of pages. The two Perplexity readings, a 48 percent first-party hit rate and 21.87 citations per answer, are consistent precisely because a wide pool plus a well-formed page produces frequent appearances. The takeaway for a buyer is to read cite rate and citation volume together: a high cite rate on a narrow-pool engine is a stronger signal than the same rate on a wide-pool engine, because the narrow engine had fewer slots to give away. This is why the index logs both the hit rate and the per-answer source count for every surface rather than collapsing them into a single visibility score.

Semrush's study of 150,000 AI citations found Reddit is the single most-referenced domain at 40.1 percent of all LLM references, ahead of Wikipedia at 26.3 percent and YouTube at 23.5 percent. The same finding was reported independently by Search Engine Land. Engines lean on user-generated content because forum threads and video transcripts capture first-hand experience that reads as authentic, and because that content is dense with the natural-language phrasing a model matches against a query. A polished brand blog, by contrast, reads as marketing and gets discounted.

The mix is not uniform across engines. Profound's analysis of 680 million citations found Wikipedia is 47.9 percent of ChatGPT's top-10 sources while Reddit is 46.7 percent of Perplexity's top-10 sources, so the two leading engines lean on almost opposite anchors. Overall, Wikipedia is 7.8 percent of all ChatGPT citations and Reddit is 6.6 percent of all Perplexity citations. Qwairy adds the sharpest per-engine contrast: ChatGPT is the only major engine that cites Wikipedia meaningfully, at 4.8 percent of answers, versus about 0 percent elsewhere. Profound also found 80.41 percent of ChatGPT's cited URLs end in .com, so commercial sources are not shut out, they are simply outweighed by the community and reference giants.

For an operator the implication is concrete. Being present where the engines already read, a credible Reddit footprint, a Wikipedia entity, a YouTube presence, and third-party mentions, matters more than publishing another owned post. The FORKOFF Discovery Gap finding sharpens this: a brand cites at 100 percent on its own name and near 0 percent on the category, because the category answer is assembled from those aggregators and not from the brand's site. Winning the category means being cited inside the sources the engine already trusts.

The obvious over-reading of these numbers deserves a caution. That Reddit is 40.1 percent of references does not mean a brand should spam Reddit; low-effort promotional threads are exactly what moderators remove and what engines learn to discount. The signal an engine rewards is a genuine, upvoted, answer-shaped discussion that a retrieval system can lift as a self-contained response. The same logic applies to the .com share: Profound's 80.41 percent commercial-URL figure for ChatGPT shows brand sites are read, so the lesson is not to abandon owned content but to make it liftable, structured, sourced, and quotable, so it can sit alongside the community and reference giants rather than be discounted beneath them. A page that reads like a neutral, well-cited reference clears the bar; a page that reads like a brochure does not, regardless of which domain hosts it.

Start with disagreement. AuthorityTech found only 11 percent of cited domains overlap between ChatGPT and Perplexity, even though both crawl the same open web. Ahrefs found only 12 percent of AI-cited URLs rank in Google's top 10 for the original prompt. So the set of pages an engine cites is largely its own, and it is largely decoupled from classic search rankings. A domain optimized for one engine can be invisible on another, which is the whole argument for measuring each surface separately.

Now concentration. Within any single engine the citations pile onto a handful of domains. Profound measured Wikipedia at 47.9 percent of ChatGPT's top-10 sources and Reddit at 46.7 percent of Perplexity's top-10. Ahrefs found the top 50 brands take 28.9 percent of all AI Overview citations. Qwairy's unique-domain counts show the same skew from the other side: ChatGPT drew from 42,592 distinct domains while Microsoft Copilot drew from just 111, a near-total concentration on that engine. The FORKOFF first-party read matches the shape: Perplexity cites 8 to 12 distinct domains per answer and FORKOFF appears in about half of them, while Gemini's median is 3 distinct domains, the narrowest pool of the five.

The synthesis is that AI citation is a concentrated, per-engine game of a few slots. Ahrefs' finding that only 38 percent of AI Overview citations now come from a Google top-10 page, down from 76 percent, confirms the slots are opening up to sources beyond the classic winners. The opportunity is real but narrow: on each engine there are only a few cited positions per answer, they are not the same positions across engines, and winning one does not win the others.

There is a strategic reading of the decoupling that matters more than the headline. When only 12 percent of AI-cited URLs also rank in Google's top 10, and the top-10 share of AI Overview citations has fallen from 76 percent to 38 percent, a page that cannot win the classic SERP still has a path to the answer box. The concentration cuts both ways: the top 50 brands hold 28.9 percent of AI Overview citations, which is a moat for incumbents, but the other 71 percent is spread across a long tail that a focused, well-sourced page can enter without out-ranking a giant on the blue links. That is the opening a challenger plays for. The discipline is to pick the two or three engines that matter for a category, measure each on its own scorecard, and win the specific slots that engine gives, rather than chase an averaged visibility number that no single answer box ever produces.

The Discovery Gap is the single most counter-intuitive finding in the first-party data, and it reframes the whole category. When a query names the brand, the kind a customer types after they have already heard of you, the engine cites the brand at 100 percent across ChatGPT, Claude, Gemini, and Perplexity, because it trained on the brand's own site. When a query is a category head term, the kind a stranger types to find a vendor, the same engines cite the brand at near 0 percent, because they answer the category with aggregators, review platforms, and competitors.

This is why a brand can feel visible in AI and still be undiscovered. A founder tests the engine with their own name, sees a glowing, accurate answer, and concludes AI search is handled. The test measured retention, not discovery. The gap shape repeats on every brand FORKOFF audits, at varying scales, and it is invisible unless the branded and head-term queries are measured as a matched pair. Of the 50 buyer-intent prompts in the index cluster, 14 returned zero citations on any engine, and those blanks cluster on exactly the head-term and tooling queries where discovery, not retention, is at stake.

The fix is not more owned content about yourself, because that moves 100 to 100 for zero discovery lift. The fix is earning presence inside the sources the engine consults for the category: third-party mentions, comparison and listicle placements, community threads, and the entity and schema work that lets the engine connect the brand to the category in the first place. The full study, including the three-lever fix, is the companion dataset linked below.

The Discovery Gap also explains why so much AI-visibility advice disappoints. Advice that optimizes the owned site, better meta, cleaner headings, more FAQ blocks, tends to lift the branded number that was already near 100 percent, so it feels productive and changes nothing on the query that actually acquires customers. The 14 of 50 index prompts that returned zero citations on any engine are the measurable shape of that trap: they cluster on the category and tooling questions a buyer types before they know the brand exists, the exact queries owned-content work cannot reach. Reading the gap correctly reorders the whole budget. It moves spend away from publishing more about yourself and toward earning citations inside the aggregators, comparisons, and communities the engine assembles a category answer from, which is where a stranger first meets the brand.

Ahrefs studied 75,000 brands and found that brand web mentions correlate 0.664 with AI visibility while backlinks correlate only about 0.218, and YouTube mentions correlate 0.737 with ChatGPT brand visibility, the single strongest factor measured. Domain Rating, the classic authority metric, correlates a weak 0.326. The direction is unambiguous: for AI visibility, being mentioned across the web out-predicts being linked to. Ahrefs also found brands in the top mention quartile earn 10 times more AI Overview mentions than the next quartile.

Princeton's peer-reviewed GEO study (KDD 2024) ran controlled experiments on what changes to a page move its generative-engine visibility, and ranked the levers by measured impact: citing authoritative sources lifts visibility 115 percent for lower-ranked pages, adding statistics lifts it 41 percent, and adding expert quotations lifts it 28 percent, with GEO methods together lifting visibility up to 40 percent. Ahrefs separately measured that AI Overview content skews 25.7 percent fresher than content cited in traditional organic results, so recency is a lever in its own right.

These are not two unrelated lists. The correlation study says the highest-impact off-site signal is being mentioned and quoted across authoritative sources, and the experiment says the highest-impact on-page change is citing authoritative sources and packing verifiable statistics into the content. Both point at the same underlying behavior: engines reward content that reads as well-sourced and entity-connected, and they reward brands the broader web already talks about. That is the design brief for a page built to be cited, and it is the brief this hub follows.

A correlation is not a guarantee, and the honest reading matters. A 0.664 correlation between brand mentions and AI visibility does not mean mentions alone cause citations; large, well-mentioned brands also tend to have better content, stronger entities, and more of everything, so some of the signal is confounded. What makes the conclusion durable is that the observational study and the controlled Princeton experiment converge from different directions: the experiment isolates causation by changing one page at a time and still finds authoritative sourcing worth up to a 115 percent lift, and the correlation study finds the same lever dominates at scale across 75,000 brands. When a controlled test and a large field study point at the same lever, the recommendation is safe to act on. The weak backlink correlation of about 0.218 is the useful negative result: it does not say links are worthless for classic SEO, it says they are the wrong first place to spend for AI visibility.

Pew Research studied real browsing behavior and found that when an AI summary appears in Google results, users click a traditional result only 8 percent of the time, versus 15 percent without a summary, and they click a link inside the summary just 1 percent of the time. About 18 percent of US Google searches in March 2025 already produced an AI summary. Ahrefs measured a 34.5 percent click reduction when an AI Overview is present. The click that used to reward a top ranking is being absorbed by the answer itself.

This did not start with AI. SparkToro and Datos found that 58.5 percent of US Google searches already ended without a click to the open web in 2024, so the zero-click search was the baseline the AI era is now accelerating. Gartner forecasts that traditional search-engine volume will fall 25 percent by 2026 as AI chatbots and virtual agents become substitute answer engines, and BrightEdge tracked Google AI Overview coverage growing from 26.6 percent to 44.4 percent of queries across nine industries between May 2024 and September 2025. The surface is expanding, not plateauing.

There is a counter-signal worth holding: the traffic that does come through AI search converts. Ahrefs reported that AI-search visitors convert about 23 times better than traditional organic visitors, despite AI search being roughly 0.5 percent of traffic today, because a user who arrives after an AI answer has already been pre-qualified by it. Meanwhile the audience keeps scaling: ChatGPT reached about 900 million weekly active users in February 2026, up from 400 million a year earlier. Fewer clicks, higher intent, a bigger surface. The rational response is to compete for the citation, because the citation is what the user now sees.

The strategic conclusion is not that clicks disappear but that their value concentrates. If an AI summary answers the low-intent questions in the answer box, the clicks that still happen are the ones from users who need more than a summary, and those users convert at the 23-times rate Ahrefs measured. A brand that earns the citation wins twice: it is named inside the answer the majority read without clicking, and it captures the smaller, higher-intent stream that does click. A brand that is absent from the citation loses both, and it loses them on a surface that is still growing, from 26.6 percent to 44.4 percent AI Overview coverage and toward a 900-million-user chat audience. That asymmetry, most of the visibility with none of the click for those who stay in the answer, plus the best of the clicks for those who leave it, is why the citation rate, not the ranking, is the number this index tracks.