AI Citation Statistics 2026: Which Sources AI Actually Cites

Ranking used to be the whole game. Now there is a second question sitting on top of it: when an AI assistant answers on your behalf, whose page does it point at? Over the last eighteen months the AI visibility platforms have built citation datasets running into the billions, and the picture is stranger than most SEO teams expected. The sources AI systems cite are not simply the pages that rank. A large share of them rank nowhere at all. This piece pulls together what the 2025 and 2026 studies found: how many sources each platform cites, which domains dominate, how little the platforms agree with each other, and how quickly a citation you won in March disappears by June.

Key AI citation statistics for 2026

• ChatGPT averages 15 sources per response and Gemini averages 3, according to Semrush's 2026 AI Visibility Index, built on 126 million US prompts [14]
• Just 37.9% of Google AI Overview citations come from pages in the top 10, down from 76% in July 2025, per Ahrefs [5]
• 31.0% of AI Overview citations point at pages that rank nowhere in Google's top 100 (Ahrefs) [5]
• Wikipedia accounts for 13.15% of ChatGPT citations and Reddit 11.97%, on Similarweb's sample of roughly 600,000 citation events [4]
• Fandom, not Wikipedia, is the single most-cited domain in Google AI Mode at 7.16% (Similarweb) [4]
• Only 3.8% of cited sources appear across all four major engines, and 72-73% are exclusive to a single engine, per Writesonic's study of 161,286 prompts [12]
• Between 40% and 60% of domains cited in AI answers are completely different one month later, on Profound's measure [10]
• Brand-owned sites make up 57% of citations globally across eight assistants, rising to roughly 68% on Google AI Overviews (Profound, 11.84 billion citations) [9]
• 75% of pages cited by LLMs were updated within the past year, but only 42% were originally published in that window, per Seer Interactive [13]
• Pages cited in AI Overviews average 1,282 words, barely above the 1,188-word average of organic results, and length correlates with citation at a Spearman coefficient of 0.04 (Ahrefs) [6]
• Google's own domains climbed 8.4x in AI Mode citation share between mid-April and end of June 2026, per Profound data reported by Search Engine Journal [19]

How many sources does each AI platform cite per answer?

It varies by an order of magnitude. Semrush's 2026 index, covering 126 million US prompts collected between January and April 2026, puts ChatGPT at around 15 sources per response and Gemini at around 3 [14].

The gap matters more than it first appears. A three-source answer is close to winner-take-all, so on Gemini you are competing for a podium finish rather than a place in the field.

Fifteen sources sounds generous by comparison, but many sit behind a collapsed source list that most users never open, so a raw citation count overstates the attention any single link receives.

Semrush also found that on Gemini, the overlap between a brand being mentioned and a brand being cited can be as low as 30%, meaning the assistant talks about you without linking to you roughly two thirds of the time [14]. That distinction between being mentioned and being linked runs right through the AI brand visibility statistics too.

Which domains do AI systems cite most?

User-generated content and encyclopaedic reference sites dominate, but the exact ordering shifts by platform. Similarweb's analysis of nearly 600,000 citation events in the US during January and February 2026 gives the cleanest percentages available [4].

Source: Similarweb, January to February 2026 [4]

Peec AI's larger sample of 30 million cited sources, published in March 2026, ranks Reddit first overall, followed by YouTube, LinkedIn, Wikipedia, Forbes, G2, Yelp, Facebook, Medium and TechRadar [2]. Search Engine Land's write-up of the same study noted that Perplexity leans on Reddit, LinkedIn and G2 for B2B queries whilst Google leans on Facebook and Yelp [3].

Profound, working from 680 million citations gathered between August 2024 and June 2025, put Wikipedia at 7.8% of ChatGPT citations and Reddit at 6.6% of Perplexity's, with .com domains taking over 80% of all citations and .org sites 11.29% [11].

The Fandom result is the one that should stop people short. A fan-wiki network is the most-cited domain in Google AI Mode, which tells you something uncomfortable about what a retrieval system optimises for: dense, entity-heavy pages that answer narrow factual questions, regardless of whether a marketer would call the site authoritative.

These rankings are also less stable than they look. A separate Semrush study of 230,000 prompts and more than 100 million citations, run over 13 weeks from July to October 2025, recorded Reddit's presence in ChatGPT responses collapsing from around 60% to around 10% after a mid-September change [1].

Note also that openai.com sits third on ChatGPT and google.com fifth on AI Mode. Both platforms cite themselves heavily, and that share is growing rather than shrinking.

How much of AI citation traffic comes from pages that already rank?

Less than it used to, and the trend has moved fast. Ahrefs analysed 863,000 keyword SERPs and 4 million AI Overview URLs as of January 2026, finding 37.9% of citations came from top 10 pages, 31.2% from positions 11 to 100, and 31.0% from pages outside the top 100 entirely [5].

The July 2025 version of the same study put the top 10 figure at 76% [5][20]. Halving in six months is not a normal metric movement.

Source: Ahrefs [5][20]

Ahrefs attributes part of the shift to query fan-out, where the assistant decomposes one question into many sub-queries and retrieves against each, and part to its own improved parsing since July 2025 [20]. That second caveat is rarely repeated: some of the drop is measurement, not behaviour.

BrightEdge reaches an almost opposite conclusion using a different method, reporting that 54.5% of AI Overview citations ranked organically as of September 2025, up from 32.3% in May 2024 [8]. Its sector splits are the more useful part of that dataset. Healthcare showed 75.3% overlap between citations and organic rankings, education 72.6% and B2B tech 71.0%, whilst restaurants sat at 19.2%, e-commerce at 22.9% and travel at 23.6% [8].

If you sell to businesses or operate in a regulated space, classic ranking work still carries most of the load. If you are in retail, hospitality or travel, it plainly does not.

How often does a cited page rank nowhere in Google?

Roughly a third of the time, on Ahrefs' January 2026 data. When the analysis is restricted to organic results only, 36.7% of cited pages sat outside Google's top 100 [5].

YouTube explains a chunk of that. Of the citations coming from non-ranking pages, 18.2% were YouTube URLs, and YouTube accounted for 5.6% of all AI Overview citations, having grown 34% in six months [5].

This is the finding with the sharpest practical edge. A page can be invisible in Google's index for a query and still be the source an AI assistant quotes to a buyer, which means rank tracking alone will not tell you whether you are winning or losing.

The reverse risk is real too. Sites that have spent a decade converting rankings into revenue may find their citation share drifting to competitors they have never seen in a SERP, a shift the wider AI search statistics make plain.

Do different AI platforms cite the same sources?

Almost never. Writesonic tested 161,286 prompts across ChatGPT, Gemini, Perplexity and Google AI Overviews in May and June 2026, with 70,879 returning citations from all four engines [12].

Only 3.8% of cited sources appeared on all four platforms, and 72-73% appeared on just one [12]. Pairwise Jaccard similarity ranged from 0.119 for ChatGPT and Gemini up to 0.237 for Perplexity and AI Overviews, meaning no pair of engines shared even a quarter of their sources [12].

Source: Writesonic, May to June 2026 [12]

The strategic reading is blunt. There is no single "AI visibility" number, and any tool or agency reporting one is averaging across systems that barely agree.

The academic literature has landed in the same place. A July 2026 arXiv survey of 45 GEO studies published since November 2023 concluded that commercial audits show "low source overlap, substantial run-to-run variability, and persistent fidelity gaps", and that no reviewed technique demonstrated a stable, cross-platform causal effect on organic discoverability [18].

How volatile are AI citations between identical prompts?

Very. Profound ran roughly 80,000 prompts per platform in June 2025 and repeated them a month later, measuring how many cited domains had changed. Google AI Overviews showed 59.3% drift, ChatGPT 54.1%, Microsoft Copilot 53.4% and Perplexity 40.5%, rising to between 70% and 90% over six months [10].

Similarweb records swings that are wilder still at brand level, including one case where a brand's share of perfume-related citations fell from 86% to 14% between measurement periods [21].

This is the number that should change how agencies report. A single prompt check taken on a Tuesday is closer to a weather reading than a ranking, and month-on-month movement of 50% is the noise floor, not a signal.

It also punctures a lot of case-study marketing. If half your cited domains turn over regardless, attributing a citation gain to last month's content refresh needs far more evidence than most people gather.

Is AI citing user-generated content or brand-owned pages?

Both, and the two most-quoted studies disagree so sharply that the disagreement is the finding.

Profound, working from 11.84 billion citations across eight assistants between April and July 2026, found brand-owned sites account for 57% of citations globally. On ChatGPT the split was 47% brand sites and 30% earned media, whilst Google AI Overviews reached roughly 68% brand sites with 15.3% social [9].

Yext went further. Its analysis of 6.8 million citations from 1.6 million queries per model, run across July and August 2025, found 86% came from brand-managed sources: 44% first-party websites, 42% listings and 8% reviews or social, with third-party forums such as Reddit at just 2% [16].

Set that against Similarweb's finding that Reddit alone is 11.97% of ChatGPT citations [4], and the contradiction looks fatal until you read the methods. Yext tested location and brand-level queries across retail, financial services, healthcare and food service, where the correct answer genuinely is the brand's own listing. Similarweb and Peec sampled broader informational prompts, where forums win.

The lesson is that UGC dominance is query-dependent, not universal, and any statistic quoted without its query mix is close to meaningless.

Semrush's separate work with Quora found the platform was the fourth most-cited domain in Google AI Mode at 7.25% of responses, roughly one in fourteen answers, and that threads marked relevant by Quora's own machine learning were nine times more likely to be cited [15].

How old and how long are the pages AI systems cite?

Fresher than organic results, but not by as much as the freshness advice suggests, and length barely matters at all.

Ahrefs analysed 174,048 pages drawn from 1,677,876 cited URLs across 560,346 AI Overviews in December 2025. Cited pages averaged 1,282 words against 1,188 for organic results, 53.4% were under 1,000 words, and the correlation between length and citation was 0.04 [6].

On age, Ahrefs looked at 16.975 million cited URLs in July 2025 and found AI assistants cite content averaging 1,064 days old, about 2.9 years, against 1,432 days for Google organic results [7].

Seer Interactive's 2026 study reads very differently at first glance. Across 7,683 dated pages and 47,097 citations from March to June 2026, 75% of cited pages had been updated within the past year and 88% within two [13].

The reconciliation sits in Seer's own data. Measured by original publication date, only 42% of those pages were less than a year old, and more than a quarter of "fresh" content was first published over two years earlier [13].

Seer also found that pages cited in all four months of its study had a median age of six months and only 68% updated within a year, whilst one-month citation spikes skewed to two-month-old pages [13]. Durable citation looks less like novelty and more like maintained substance.

Which sectors see the most concentrated citations?

Concentration varies enormously. Semrush found the top three brands hold 82.9% of visibility in news and media and 76.9% in consumer electronics, against 41.4% in finance and 42.2% in industrial [14].

Goodie's analysis of 58.6 million citations across 31 sectors between October 2025 and March 2026 shows how thin the leading shares are outside the giants. Wikipedia took 3.4% of all citations, YouTube 2.1% and Reddit 1.4% [17].

Within verticals, specialists win. TechRadar took 8.9% of CRM and sales software citations, NerdWallet 3.9% in retail banking and fintech, and Wikipedia 10.4% in hotels and resorts [17]. Seer's concentration figures agree: the top 10% of sites held 78% of citations in retail energy but only 49% in commercial banking [13].

In a concentrated sector, displacing an incumbent citation is the entire job. In a fragmented one, breadth of coverage does more work than any single page.

How to read these numbers

Almost every figure here comes from a commercial AI visibility platform running its own prompt panel. That is not the same as observing real user queries, and prompt sets are chosen partly to suit the vendor's customer base.

Sample sizes are large but not comparable. Profound's 11.84 billion citations, Ahrefs' 4 million URLs and Seer's 47,097 citations measure different things at different grains, and a bigger number is not automatically a better estimate.

Definitions move too. "Citation" can mean a visible inline link, an entry in a collapsed source list, or any URL retrieved during generation, and vendors rarely state which. Ahrefs' own top 10 overlap figure halved in six months partly because of a parsing change on their side [20], so treat any citation statistic older than about nine months as historical.

Where credible sources disagree, check the query mix first. The Yext 86% brand-managed figure and the Similarweb 11.97% Reddit figure are not really in conflict, they are answers to different questions [16][4].

Finally, Google's self-citation growth of 8.4x between April and June 2026 [19] is a reminder that the denominator is unstable. Share of citations can fall whilst absolute citations rise, simply because the platform started citing itself more.

What this means going into 2027

Rank tracking is now a partial measure. With around a third of AI Overview citations coming from pages outside Google's top 100 [5], any reporting stack that only watches positions will miss a growing share of where recommendations come from.

Platform-specific strategy stops being optional. At 3.8% universal source overlap [12], optimising for "AI" as a single channel is a category error. Pick the one or two engines your buyers actually use.

Measurement cadence has to rise. With 40-60% monthly citation drift [10], single-point checks cannot support a decision, and fortnightly sampling across a stable prompt set is the minimum for anything you plan to act on.

Maintenance beats publication volume. Seer's split between 75% recently updated and 42% recently published [13] points at refreshing durable pages rather than adding new ones, and Ahrefs' 0.04 length correlation [6] says padding word counts is wasted effort.

Expect the platforms to keep taking share. Google properties and openai.com already sit inside the top five most-cited domains on their own systems [4][19], which caps how much of the citation pool is available to anyone else. If you want a clear read on which sources AI actually cites in your category, and a plan to earn more of them, you can book a call with me and we will look at your citation data together.