GEO Statistics (2026): 83 Research-Backed Stats on AI Search Citations

A data roundup on generative engine optimization (GEO): what changes a page’s visibility inside AI answers, which sources ChatGPT, Claude, Gemini, Perplexity and Google’s AI features cite, how often those citations are right, and what AI answers do to clicks. Every figure comes from peer-reviewed or preprint research, a named research center, a published study by the data owner, or Alphabet’s own earnings remarks. No vendor survey, traffic panel or market-report estimate is used.

The short answer: GEO works, but not the way old SEO did. In the original KDD 2024 study, adding citations, quotations and statistics to a page lifted its visibility in generative engine answers by up to 40%, while keyword stuffing did little or nothing (Aggarwal et al.). A 2025 University of Toronto study then found that AI search shows “a systematic and overwhelming bias” toward earned, third-party media over brand-owned pages (Chen et al.). Put together, the research says two things: write pages that are easy to quote and back with numbers, and get covered by the independent sites AI engines trust. If you only read one section, read what to do with this data.

Key takeaways

• Up to 40% more visibility in generative engine answers from GEO methods, in the paper that coined the term (Aggarwal et al.)
• 115.1% visibility lift from citing sources for sites ranked fifth on Google, while the top-ranked site lost 30.3% on average (Aggarwal et al.)
• 93.5% of ChatGPT’s sources for well-known brands were earned media, with 6.5% from brand sites and none from social (Chen et al.)
• 76.10% of pages cited in Google AI Overviews rank in Google’s top 10 (Ahrefs)
• 58% lower click-through rate for the top-ranking page when an AI Overview appears, in December 2025 data (Ahrefs)
• 8% of Google visits with an AI summary led to a click on a traditional result, against 15% without one (Pew Research Center)
• 0.737 correlation between YouTube mentions and AI visibility, the strongest factor in a 75,000-brand study (Ahrefs)
• 1 billion monthly active users for Google’s AI Mode in Q2 2026 (Alphabet)

What is generative engine optimization?

Generative engine optimization (GEO) is the practice of shaping your content and your presence across the web so that AI systems that answer questions, such as ChatGPT, Perplexity, Claude, Gemini, Google AI Overviews and Google AI Mode, mention your brand and cite your pages. The term comes from a 2023 paper led by researchers at Princeton, which was accepted at KDD 2024, one of the top data mining conferences (arXiv).

GEO differs from search engine optimization in what you can measure. A search engine shows a ranked list of links, so you track position. A generative engine writes an answer and attaches a handful of sources, so you track whether you are mentioned, how prominently, how you are described and which pages get cited. That is why the research below talks about “visibility”, “share of citations” and “earned media” rather than rankings.

Definitions used on this page

• Generative engine: a search or chat product that answers a query by writing text grounded in retrieved web sources, and usually cites them.
• Position-Adjusted Word Count: the GEO paper’s main visibility metric. It counts the words in an answer that come from your source, weighted so that text near the top counts more.
• Subjective Impression: the GEO paper’s second metric, an AI-judged score for how relevant, influential and prominent a source looks in the answer.
• Earned media: third-party, independent sources such as reviews, publishers and comparison sites. Brand media is the company’s own site. Social is user-generated content such as Reddit, YouTube or forums. These are the categories the University of Toronto study uses.
• Citation recall: the share of sentences in an AI answer that are fully supported by the sources it cites. Citation precision is the share of citations that actually support the sentence they are attached to.
• AI Overview / AI summary: the AI-written block at the top of some Google results. Pew calls it an “AI summary”; Google calls it an AI Overview.

Do GEO tactics actually work?

The only controlled, peer-reviewed test of specific GEO tactics is still the original GEO paper. The researchers built GEO-bench, a set of real user queries, rewrote source pages with nine different methods, and measured how much more of each page showed up in generated answers.

1. Up to 40% more visibility in generative engine responses from GEO methods (Aggarwal et al.).
2. 10K queries make up GEO-bench, split into 8K training, 1K validation and 1K test queries (Aggarwal et al.).
3. 80% of GEO-bench queries are informational, with 10% transactional and 10% navigational (Aggarwal et al.).
4. 30-40% relative improvement in Position-Adjusted Word Count, and 15-30% in Subjective Impression, from the three best methods: Cite Sources, Quotation Addition and Statistics Addition (Aggarwal et al.).
5. 41% and 28% improvements over baseline from the best methods, on Position-Adjusted Word Count and Subjective Impression respectively (Aggarwal et al.).
6. 15-30% visibility boost from stylistic changes alone, meaning Fluency Optimization and Easy-to-Understand rewrites (Aggarwal et al.).
7. Little to no improvement from keyword stuffing, the classic SEO tactic (Aggarwal et al.).
8. 31.4% average gain when Cite Sources is combined with other methods (Aggarwal et al.).
9. More than 5.5% better results from the best combination, Fluency Optimization plus Statistics Addition, than from any single method (Aggarwal et al.).
10. 22% improvement in Position-Adjusted Word Count from Quotation Addition when tested on Perplexity.ai, a live generative engine (Aggarwal et al.).
11. Up to 9% and 37% improvements on Perplexity.ai from Cite Sources and Statistics Addition on the two metrics (Aggarwal et al.).

The most practical finding is that the three winning methods all add evidence. Citing credible sources, quoting a named expert and adding numbers each make a passage more useful to an engine trying to write a grounded answer, and each one gave a bigger lift than rewriting for persuasion or authority tone. Keyword stuffing, the tactic most people think of as “optimization”, did close to nothing.

Does GEO help smaller sites more than big ones?

Yes, according to the same paper. When the researchers split results by where a site ranked on Google, the lower-ranked sites gained the most and the top site often lost share.

12. 115.1% visibility increase from Cite Sources for websites ranked fifth on the search results page, while the top-ranked website lost 30.3% on average (Aggarwal et al.).
13. 99.7 relative gain for fifth-ranked sites from Quotation Addition, against a 22.9 drop for the top-ranked site (Aggarwal et al.).
14. 97.9 relative gain for fifth-ranked sites from Statistics Addition, against a 20.6 drop for the top-ranked site (Aggarwal et al.).

Source: Aggarwal et al., GEO: Generative Engine Optimization, Table 2. Relative change in visibility by the site’s Google rank. Fetched October 28, 2026.

For a challenger brand, this is the most encouraging result in the GEO literature. In a ranked list you cannot jump from fifth to first by adding three statistics, but in a generated answer you can take a much larger share of the words.

Which sources do AI search engines cite?

The largest controlled comparison of AI search and Google sourcing so far comes from a 2025 University of Toronto study. Its researchers ran ranking-style queries (“best laptops”, “best CRM software”) through ChatGPT, Claude, Gemini, Perplexity and Google, then labeled every source as earned, brand or social.

15. Overwhelming bias toward earned media: the authors conclude that AI search exhibits “a systematic and overwhelming bias towards Earned media (third-party, authoritative sources) over Brand-owned and Social content” (Chen et al.).
16. 1,000 consumer ranking prompts, 100 in each of 10 categories, were used for the Google vs AI search comparison (Chen et al.).
17. 93.5% of ChatGPT’s sources for well-known brands were earned media, 6.5% brand and none social (Chen et al.).
18. 87.3% of Claude’s sources for well-known brands were earned, 6.8% brand and 5.9% social (Chen et al.).
19. 23.8% of Perplexity’s sources for well-known brands were social content, with 67.4% earned and 8.8% brand (Chen et al.).
20. 25.1% of Gemini’s sources for well-known brands were brand-owned, with 63.4% earned and 11.5% social (Chen et al.).
21. 95.1% of ChatGPT’s sources for niche brands were earned media, with only 4.9% brand (Chen et al.).
22. 76–81% agreement between AI systems on well-known brands, falling to 71–76% for niche brands (Chen et al.).
23. 56.3% of ChatGPT’s brand mentions went to major brands (274 of 487); for Perplexity the figure was 67.9% (339 of 499) (Chen et al.).
24. 19.7% of ChatGPT’s citations came from Wikipedia in the brand-bias experiment (Chen et al.).

Source: Chen, Wang, Chen and Koudas, Generative Engine Optimization: How to Dominate AI Search, well-known brand queries. Fetched October 28, 2026.

The engines are not interchangeable. ChatGPT and Claude lean almost entirely on independent publishers, Perplexity pulls in far more Reddit and YouTube, and Gemini gives brand sites the most room. A brand that only optimizes its own site will look strongest in Gemini and weakest in ChatGPT, which is one reason to track each engine separately. Our Perplexity statistics, ChatGPT statistics and Gemini statistics pages cover each engine’s usage.

How different are AI search sources from Google’s?

The same study compared AI answers with Google’s organic results for identical queries in the United States and Canada. For software, the two systems pointed to almost opposite kinds of sources.

25. 72.7% earned, 26.7% brand and negligible social sources in AI search answers for U.S. software queries (Chen et al.).
26. 43.7% brand, 10.9% social and 45.4% earned sources in Google’s results for the same U.S. software queries (Chen et al.).
27. 53.8% of Google’s software sources in Canada were brand-owned, against 74.2% earned and no social in AI search (Chen et al.).
28. 92.1% earned content in AI search answers for U.S. consumer electronics queries (Chen et al.).
29. 81.9% earned and 18.1% brand sources in AI search answers for U.S. automotive queries (Chen et al.).
30. 20.6% domain overlap between AI and Google for local Home Cleaning searches, the highest of any local category; IT Support overlapped just 0.1% (Chen et al.).
31. 60.8% of GPT’s cited automotive domains were unique to GPT, against 50.3% for Claude and 56.5% for Perplexity (Chen et al.).
32. 67.6% of GPT’s consumer electronics domains, and 67.2% of Perplexity’s, were cited by no other engine (Chen et al.).

Source: Chen et al., University of Toronto study, software vertical, United States. Fetched October 28, 2026.

For software companies this is the headline number. On Google, your own pricing and feature pages make up a large share of what people see for “best X software” queries. In AI search, roughly three in four sources are someone else’s page about you. That makes review sites, comparison pages, analyst coverage and listicles the main lever, and it makes checking how those pages describe your pricing and features part of the job. If you sell software, our guide to the best AI visibility tools for SaaS covers the tools built for that check.

Does Google ranking still matter for AI citations?

For Google’s own AI features, yes. Ahrefs matched 1.9 million AI Overview citations against its ranking data and found that most cited pages already rank on page one.

33. 1.9M citations from 1M AI Overviews were analyzed (Ahrefs).
34. 76.10% of AI Overview-cited pages rank in Google’s top 10 (Ahrefs).
35. 9.50% of cited pages rank between positions 11 and 100 (Ahrefs).
36. 14.40% of pages cited in AI Overviews do not rank in the top 100 at all (Ahrefs).

Read this next to the Toronto study. Google’s AI Overviews borrow heavily from Google’s own ranking, so classic SEO still feeds them. ChatGPT, Claude and Perplexity rely far more on earned media, and their source lists overlap little with Google’s. A GEO plan needs both: rank for the query, and get mentioned on the independent pages that chat-style engines retrieve. Our Google AI Overviews statistics page has more on how often Overviews appear.

What predicts whether AI mentions a brand?

Ahrefs also ran the largest correlation study of brand visibility in AI answers, covering 75,000 brands across ChatGPT, Google AI Mode and AI Overviews. Correlation is not causation, but the pattern is consistent: off-site mentions matter more than on-site size.

37. 75,000 brands were analyzed for mentions in ChatGPT, AI Mode and AI Overviews (Ahrefs).
38. 0.737 correlation between YouTube mentions and AI visibility, the strongest factor measured (Ahrefs).
39. 0.66–0.71 correlation between branded web mentions and AI visibility (Ahrefs).
40. 0.194 correlation for the number of pages on a brand’s site, one of the weakest factors (Ahrefs).
41. 0.352 and 0.266 correlations between ChatGPT visibility and branded search volume and Domain Rating, respectively (Ahrefs).
42. 0.628 correlation between AI Mode visibility and branded anchors (Ahrefs).
43. 0.779 output overlap correlation, meaning the three AI surfaces largely mention the same brands (Ahrefs).

Source: Ahrefs, AI brand visibility correlations, 75,000 brands, published December 12, 2025. Fetched October 28, 2026.

How accurate are AI search citations?

Being cited is only half the story. Several independent studies show AI engines often cite the wrong page, attach a citation that does not support the sentence, or invent a link. That is why brands need to check what engines actually say, not just whether they appear.

44. 51.5% of sentences from generative search engines were fully supported by their citations, and only 74.5% of citations supported their sentence, in Stanford’s 2023 audit (Liu, Zhang and Liang).
45. 68.7 average citation recall for perplexity.ai, the highest tested, against 11.1 for YouChat (Liu, Zhang and Liang).
46. 89.5 average citation precision for Bing Chat, the highest tested (Liu, Zhang and Liang).
47. More than 60% of 1,600 news-citation queries were answered incorrectly across eight AI search tools (Tow Center, CJR).
48. 1,600 queries were run: 20 publishers, times 10 articles, times 8 chatbots (Tow Center, CJR).
49. 37% of Perplexity’s answers were incorrect, against 94% for Grok 3 (Tow Center, CJR).
50. 134 articles were misidentified by ChatGPT, which signaled low confidence just 15 times in 200 responses (Tow Center, CJR).
51. 115 out of 200 sources were misattributed by DeepSeek (Tow Center, CJR).
52. 154 of Grok 3’s 200 citations led to error pages (Tow Center, CJR).

These studies tested news articles and early engines, not product pages, so treat them as a warning rather than a measure of how engines describe your company today. The practical point stands: an AI answer can name you and still get your price, plan or feature set wrong, and nothing in a rank tracker will tell you.

Which news and reference sites get cited most?

Two 2025 papers analyzed the public Search Arena dataset, where people compared answers from OpenAI, Google and Perplexity search models side by side. They show how concentrated citations are.

53. 366,087 citations pointed to 83,533 unique domains across the dataset (News citing patterns study).
54. 9.0% of all citations went to news outlets, which were just 2.1% of unique domains (News citing patterns study).
55. 10% of citations pointed to social media platforms (News citing patterns study).
56. 67.3% of OpenAI models’ news citations went to their top 20 news sources; for Google the figure was 31.9% and for Perplexity 28.5% (News citing patterns study).
57. 1,430 unique news sources were cited by Perplexity models, against 881 for Google and 707 for OpenAI (News citing patterns study).

Source: News Source Citing Patterns in AI Search Systems, Search Arena data. Fetched October 28, 2026.

OpenAI’s models concentrate their citations in a small set of outlets, while Perplexity spreads them across many more. For a smaller brand, that suggests Perplexity is the easier engine to break into through niche publications, while ChatGPT rewards coverage in the few outlets it already trusts.

What do people ask AI search, and which answers do they prefer?

Search Arena is also the best public look at what people actually ask search-enabled AI, and how citations affect which answer they prefer.

58. 24,069 conversations and 12,652 human preference votes were released in the Search Arena dataset (Miroyan et al.).
59. 19.3% of prompts were factual lookups, so most use is something other than simple fact-finding (Miroyan et al.).
60. 22.0% of prompts were technology comparisons, the most common topic, ahead of market analysis at 12.3% (Miroyan et al.).
61. 22.4% of conversations were multi-turn (Miroyan et al.).
62. 6.9 citations on average in answers to recommendation prompts, against 6.8 for info synthesis and 5.7 for factual lookups (Miroyan et al.).
63. More references preferred: users favored responses with more citations (Miroyan et al.).
64. Negative correlation between citing Wikipedia and user preference (Miroyan et al.).

Technology comparisons being the single largest topic matters for software brands. People are using AI search to compare products, and the answers to those prompts carry the most citations, which means more slots for third-party pages that mention you.

How do AI answers change clicks from Google?

The best independent measure of click behavior comes from Pew Research Center, which tracked the real Google searches of 900 U.S. adults in March 2025. Ahrefs has since measured the effect at the level of click-through rates.

65. 8% of visits with an AI summary led to a click on a traditional result, against 15% of visits without one (Pew Research Center).
66. 1% of visits to pages with an AI summary included a click on a link inside the summary (Pew Research Center).
67. 26% of pages with an AI summary ended the browsing session, against 16% of pages with only traditional results (Pew Research Center).
68. 18% of Google searches in the study produced an AI summary (Pew Research Center).
69. 68,879 unique Google searches were analyzed, and 12,593 produced an AI summary (Pew Research Center).
70. 88% of AI summaries cited three or more sources; only 1% cited a single source (Pew Research Center).
71. 53% of searches with 10 or more words produced an AI summary, against 8% of one- or two-word searches (Pew Research Center).
72. 60% of searches starting with question words such as who, what, when or why produced an AI summary (Pew Research Center).
73. 15% of the sources in AI summaries were Wikipedia, YouTube or Reddit (Pew Research Center).
74. 6% of AI summary sources were .gov sites, against 2% of standard results (Pew Research Center).
75. 58% lower average click-through rate for the top-ranking page when an AI Overview is present, in December 2025 data (Ahrefs).
76. 34.5% was the click reduction Ahrefs measured in April 2025, so the effect grew in eight months (Ahrefs).

Source: Pew Research Center, 68,879 searches by 900 U.S. adults, March 2025. Fetched October 28, 2026.

Long, question-shaped searches are exactly the ones that trigger AI answers, and those are the searches where clicks fall the most. For a brand, being named in the answer is becoming the outcome that matters, because the click often never comes.

How big are Google’s AI search surfaces now?

Alphabet’s own numbers show how fast AI answers have become the default search experience.

77. 1 billion monthly active users for AI Mode, after Google expanded it globally the October before (Alphabet).
78. Billions of clicks go to websites every week through AI features in Search, according to Google (Alphabet).
79. 17% revenue growth in Search and Other in Q2 2026 (Alphabet).
80. 950 million monthly active users for the Gemini app (Alphabet).

Google says AI features still send “billions of clicks”, while Pew and Ahrefs show each individual result gets fewer. Both can be true: the total grows with search volume while the share per result falls. For more on AI Mode, see our Google AI Mode statistics.

How fresh are the pages AI engines cite?

81. 117 days was the mean age of Claude’s consumer electronics links, with a median of 62 days (Chen et al.).
82. 331 days was the mean age of Claude’s automotive links, with a median of 148 days, showing a heavy reliance on older reviews and ranking pages (Chen et al.).
83. KDD 2024 accepted the original GEO paper, the first peer-reviewed study of the field (arXiv).

Freshness varies a lot by category. In fast-moving categories engines lean on recent coverage, while in slower ones a review from a year ago can keep being cited. An outdated review that misstates your pricing can therefore shape AI answers for months.

Stats ledger: what each source measured

Source: tools8020 stats ledger for this page. Every row links to the primary page we fetched on October 28, 2026.

What to do with this data

1. Add evidence to the pages you want cited. The GEO paper’s three best methods were citing sources, adding quotations and adding statistics. Start with your pricing, comparison and “what is” pages.
2. Stop keyword stuffing. It produced little to no gain in generative answers.
3. Earn third-party coverage. AI search draws roughly three quarters of software sources from earned media, so reviews, listicles, analyst pages and comparison posts carry more weight than your own blog.
4. Keep ranking on Google. About three in four AI Overview citations rank in Google’s top 10, so classic SEO still feeds Google’s AI surfaces.
5. Track each engine separately. ChatGPT, Claude, Perplexity and Gemini cite very different mixes of sources and overlap little.
6. Check accuracy, not just presence. Independent audits show AI engines misattribute and misstate often. Monitor what they say about your pricing and features.
7. Watch older third-party pages. Engines can keep citing a review for months, so outdated coverage needs correcting at the source.

A tool such as Mentionova automates steps 5 and 6. It tracks your prompts across ChatGPT, Perplexity, Claude, Gemini and Google AI, shows which sources each engine cites, and checks whether the engines describe your pricing and features accurately. For a broader shortlist, see the best AI visibility tools, the best GEO tools and the best AI citation tracking tools. For usage numbers across engines, see our AI search statistics.

Frequently asked questions

What is GEO in marketing?

GEO, or generative engine optimization, is the work of getting your brand mentioned and your pages cited in AI-written answers from tools such as ChatGPT, Perplexity, Gemini, Claude and Google’s AI Overviews and AI Mode. The term comes from a 2023 research paper accepted at KDD 2024.

Does GEO actually work?

In the only controlled peer-reviewed test, yes. The GEO paper found its methods raised a source’s visibility in generated answers by up to 40%, with the best methods beating the baseline by 41% on word-count visibility.

What is the most effective GEO tactic?

Adding evidence. Citing sources, adding quotations and adding statistics were the three best methods in the GEO paper, each improving visibility by 30-40%. The best combination was fluency edits plus added statistics.

Does keyword stuffing help in AI search?

No. The GEO paper found keyword stuffing offered little to no improvement in generative engine responses.

Do AI search engines prefer brand websites?

No. The University of Toronto study found AI search has an overwhelming bias toward earned, third-party media. For U.S. software queries, 72.7% of AI search sources were earned media and 26.7% brand-owned, while Google’s results were 43.7% brand-owned.

Does ranking on Google help you get cited in AI Overviews?

Yes. Ahrefs found 76.10% of pages cited in AI Overviews rank in Google’s top 10, and 14.40% do not rank in the top 100.

Which engine cites Reddit and social content most?

Perplexity. In the Toronto study, 23.8% of Perplexity’s sources for well-known brands were social content, while ChatGPT cited no social sources for those queries.

How much do AI Overviews reduce clicks?

Pew found people clicked a traditional result in 8% of visits with an AI summary, against 15% without. Ahrefs found an AI Overview correlated with a 58% lower click-through rate for the top-ranking page in December 2025 data.

How accurate are AI search citations?

Not very, in independent tests. Stanford found only 51.5% of generated sentences were fully supported by citations, and the Tow Center found eight AI search tools answered more than 60% of citation queries incorrectly.

What predicts whether ChatGPT mentions my brand?

Off-site mentions. Ahrefs found YouTube mentions (about 0.737) and branded web mentions (0.66–0.71) correlated most with AI visibility, while Domain Rating correlated only 0.266 with ChatGPT visibility.

How many people use Google AI Mode?

Alphabet said AI Mode surpassed 1 billion monthly active users in its Q2 2026 earnings remarks.

How do I measure GEO results?

Track a fixed set of prompts across each AI engine and record whether you are mentioned, how you are described and which sources are cited. AI visibility tools such as Mentionova do this automatically and flag answers that misstate your pricing or features.

How we compiled these statistics

We fetched 11 primary source pages on October 28, 2026 and recorded 83 data points in docs/content/stats-ledgers/geo-statistics.json. Every numbered figure above maps to a ledger row with a verbatim quote from the page we fetched. Sources are peer-reviewed or preprint research on arXiv (the GEO paper, the University of Toronto study, the Stanford verifiability audit, the Search Arena paper and the news-citation study), Pew Research Center, the Tow Center for Digital Journalism at Columbia, Ahrefs studies run on Ahrefs’ own data, and Alphabet’s Q2 2026 earnings remarks. We did not use traffic panels, browser-share trackers, vendor surveys or market-report estimates. Percentages are quoted as each source printed them, and the GEO paper’s rank table values are relative changes as published. Preprints have not all been peer reviewed, and several studies predate today’s model versions, so treat older accuracy figures as directional.

Sources

• Aggarwal et al., GEO: Generative Engine Optimization (arXiv HTML, KDD 2024)
• GEO: Generative Engine Optimization, arXiv abstract
• Chen, Wang, Chen and Koudas, Generative Engine Optimization: How to Dominate AI Search (University of Toronto, 2025)
• Tow Center for Digital Journalism: We compared eight AI search engines (Columbia Journalism Review, March 6, 2025)
• Pew Research Center: Google users are less likely to click on links when an AI summary appears (July 22, 2025)
• Liu, Zhang and Liang: Evaluating Verifiability in Generative Search Engines (Stanford, 2023)
• News Source Citing Patterns in AI Search Systems (2025)
• Miroyan et al.: Search Arena (2025)
• Ahrefs: AI brand visibility correlations, 75K brands (December 12, 2025)
• Ahrefs: Search rankings and AI Overview citations (July 21, 2025)
• Ahrefs: AI Overviews reduce clicks, update (February 4, 2026)
• Alphabet Q2 2026 earnings: CEO remarks