What is not cited is not seen - and AI search mainly cites third parties.
What it is about
When AI answers the search, people click less, and what gets cited is mainly third-party sources rather than company websites. Those who stay visible appear in these third-party sources and prepare their content for AI systems.
The shift is fast: according to the Reuters Institute, 38% use generative AI weekly to look for information - up from 11% in 2024 (online survey via YouGov, August to September 2026, about 2,000 respondents in each of six countries, Germany not included).
What the studies show
- Fewer clicks: With an AI summary, US users clicked a traditional result in 8% of Google searches, without one in 15% (Pew Research Center, observed browsing data of 900 US adults with 68,879 searches in March 2025). In a field experiment, click-through rose by 8.8 percentage points when AI Overviews were removed (Wang et al., preregistered field experiment with 1,100 US users in March 2026, preprint). Search traffic to English Wikipedia articles fell by about 5% compared with the same articles in German and French (Khosravi, Yoganarasimhan, 2026, analysis of monthly traffic, preprint).
- Different sources: For software queries, AI search cited third-party coverage in 74.2% of cases and company websites in only 25.8%; on Google, company websites led with 53.8% (Chen et al., 2025, 1,000 consumer queries in the US and Canada, preprint).
- Measurable levers: Content optimised for generative engines was up to 40% more visible in AI answers (Aggarwal et al., KDD 2024, experiment on the GEO-bench benchmark) - in the best case and on partly simulated 2023/24 systems.
All studies come from universities and independent institutes, none from Google or an AI vendor. Almost all data comes from consumer searches in the US. There is no comparable independent measurement yet for B2B queries in Germany.
What we make of it
Visibility will arise in two places: in structured, machine-readable data on the company website and in third-party sources that AI systems cite. How we build the first part is described in Websites for agents: MCP is the new SEO.
The studies in detail
Selection: only studies with a disclosed method - experiments, official statistics, legal texts, independent institutes. Vendor surveys about their own product and consultancy forecasts are not included here.
- 38%
Reuters Institute for the Study of Journalism · 2026 · Surveyindependent
The share of people using generative AI weekly to get information rose from 11% (2024) to 24% (2025) and 38% (2026) (six countries, about 2,000 respondents each).
Base: About 2,000 per country; trend across 6 countries (Argentina, Denmark, France, Japan, UK, US); YouGov online, Aug-Sep 2026
Limits: Germany not included; online sample; self-reported
Generative AI and News Report 2026 ↗We use it in: Websites for agents: MCP is the new SEO
- 8% vs. 15%
Pew Research Center · 2025 · Studyindependent
When a Google search showed an AI summary, US users clicked a traditional result in only 8% of visits, versus 15% without one (browsing data of 900 US adults, March 2025).
Base: 900 U.S. adults (KnowledgePanel Digital), browsing data March 1-31, 2025; 68,879 unique searches
Limits: US only; one month of observed behavior; only Google searches
Google users are less likely to click on links when an AI summary appears in the results ↗We use it in: Websites for agents: MCP is the new SEO
- +8,8 pp
arXiv (Wang et al., Northeastern / UPenn) · 2026 · Experimentindependent
In a preregistered field experiment, external click-through rose by 8.8 percentage points when Google’s AI Overviews were removed; AI-Mode-only search reduced it by 18.8 points.
Base: 1,100 US users of Chrome and Google Search, March 2026, 7 days per condition
Limits: Preprint; US consumers, no B2B searches; short-term; wide confidence interval
AI in Search Reduces Publisher Referrals Without Improving User Experience: Experimental Evidence ↗We use it in: Websites for agents: MCP is the new SEO
- -5,45%
arXiv (Khosravi, Yoganarasimhan, University of Washington) · 2026 · Studyindependent
The rollout of Google AI Overviews reduced search traffic to English Wikipedia articles by about 5%, compared with the same articles in German and French.
Base: Monthly search traffic to English vs German/French Wikipedia articles, difference-in-differences with the staggered rollout
Limits: Preprint, estimate revised between versions; single informational publisher; not directly transferable to B2B sites
Impact of AI Search Summaries on Website Traffic: Evidence from Google AI Overviews and Wikipedia ↗We use it in: Websites for agents: MCP is the new SEO
- 74,2%
arXiv (Chen et al., University of Toronto) · 2025 · Studyindependent
AI search engines cite third-party coverage far more often than company websites: for software queries 74.2% of sources were earned media and only 25.8% brand-owned - Google favored brand sources (53.8%).
Base: 1,000 consumer ranking queries (electronics, automotive, software), US and Canada, Google vs web-enabled GPT
Limits: Preprint; consumer queries, not B2B procurement; automatic source classification; 2025 snapshot
Generative Engine Optimization: How to Dominate AI Search ↗We use it in: Websites for agents: MCP is the new SEO
- bis zu 40%
KDD 2024 (Aggarwal et al., Princeton) · 2024 · Experimentindependent
Optimizing content for generative engines (Generative Engine Optimization) can increase a source’s visibility in AI-generated answers by up to 40%.
Base: GEO-bench: diverse queries across domains, tested on generative search engines (2023/24)
Limits: Best case, not an average; 2023/24 engines, partly simulated; visibility in answer text, not clicks
GEO: Generative Engine Optimization ↗We use it in: Websites for agents: MCP is the new SEO
