How AI Engines Cite the Web: The Six Studies That Define the 2026 Evidence Base

The Six Studies

Study 1 — The Toronto Comparative Audit

Citation: Chen, M., Wang, X., Chen, K., and Koudas, N. Navigating the Shift: A Comparative Analysis of Web Search and Generative AI Response Generation. Proceedings of the Workshops of the EDBT/ICDT 2026 Joint Conference, Tampere, Finland. arxiv.org/pdf/2601.16858

Authors and affiliation: Mahe Chen, Xiaoxuan Wang, Kaiwen Chen, Nick Koudas. Department of Computer Science, University of Toronto. Corresponding author: Koudas (koudas@cs.toronto.edu).

Scope: 1,516 queries total. 1,000 ranking-style queries across ten consumer categories; 216 entity-comparison queries (108 popular brands, 108 niche brands); 300 consumer-electronics queries across three intent categories; 100 freshness queries each in consumer electronics and automotive. Five systems compared in parallel — Google Search, GPT-4o, Claude 4.5 Sonnet, Perplexity Sonar Pro, Gemini 2.5 Flash. Bootstrap resampling with 10,000 iterations for statistical validation.

Headline findings: GPT-4o cites domains overlapping with Google's top-10 results at a mean of 4.0 percent (median 0.0 percent). All four AI engines show under 16 percent mean overlap. Claude cites earned media in 65 percent of references and social content in 1 percent. Median article age in cited automotive sources: Google 492.9 days, Claude 148.0 days. In a controlled perturbation experiment, 16 percent of entities in GPT-4o's SUV rankings appeared with no supporting evidence in retrieved snippets — Cadillac and Infiniti were supplied from prior knowledge 58 percent and 73 percent of the time respectively.

Why this study matters: First published audit running all five major answer surfaces in parallel against an identical query set in the same time window. First controlled perturbation experiment isolating pre-training versus retrieval contribution to LLM answer generation. EPR's full breakdown is here.

Study 2 — Stanford's SourceCheckup

Citation: Wu, K., Wu, E., Wei, K., Zhang, A., Casasola, A., Nguyen, T., Riantawan, S., Shi, P., Ho, D. E., and Zou, J. Y. An automated framework for assessing how well LLMs cite relevant medical references. Nature Communications 16, 3615 (April 16, 2025). doi.org/10.1038/s41467-025-58551-6 · nature.com

Authors and affiliation: Kevin Wu and Eric Wu (co-first authors), with Kevin Wei, Angela Zhang, Allison Casasola, Teresa Nguyen, Sith Riantawan, Patricia Shi, Daniel E. Ho, James Y. Zou. Stanford University, Department of Biomedical Data Science. Senior author: Zou (jamesz@stanford.edu).

Scope: Seven popular LLMs tested — GPT-4o (API and RAG), Claude v2.1, Mistral Medium, Gemini (API and RAG), and others. 800 medical questions; 58,000 statement-source pairs. Validated against three US-licensed medical experts. Published peer-reviewed in Nature Communications.

Headline findings: Between 50 percent and 90 percent of LLM responses are not fully supported by the sources they cite. For GPT-4o with web search enabled, approximately 30 percent of individual statements are unsupported and nearly half of full responses are not fully supported. The SourceCheckup framework achieved 88.7 percent agreement with medical expert consensus — higher than the 86.1 percent average inter-doctor agreement rate. Models without web access produced valid URLs only 40 to 70 percent of the time.

Why this study matters: Peer-reviewed in a top-tier journal. Demonstrates that the citation accuracy problem in LLMs is not edge-case — it is the median case. Establishes the validated methodology (the SourceCheckup pipeline) that subsequent studies have built on.

Study 3 — Salesforce's Answer Engine Evaluation

Citation: Venkit, P. N., Laban, P., Zhou, Y., Mao, Y., and Wu, C.-S. Search Engines in the AI Era: A Qualitative Understanding to the False Promise of Factual and Verifiable Source-Cited Responses in LLM-based Search. Proceedings of the 2025 ACM Conference on Fairness, Accountability, and Transparency (FAccT '25), pages 1325–1340. ACM · arXiv:2410.22349 · GitHub (AEE benchmark)

Authors and affiliation: Pranav Narayanan Venkit (lead), Philippe Laban, Yilun Zhou, Yixin Mao, Chien-Sheng Wu. Salesforce AI Research.

Scope: 21-participant qualitative user study comparing answer engines (You.com, Perplexity.ai, BingChat) to traditional search. 16 identified answer-engine limitations. Eight measurement metrics. Released as the open-source Answer Engine Evaluation (AEE) benchmark.

Headline findings: Users systematically misperceived the reliability of cited sources in answer engines. The qualitative analysis identified frequent hallucinated citation, misattribution between cited URL and supporting claim, and variable answer confidence across engines for identical queries. The automated AEE benchmark quantifies common failure modes across the three tested engines.

Why this study matters: The qualitative dimension the Toronto paper does not have — real users describing what fails. ACM FAccT is a top peer-reviewed venue for AI accountability research. The companion paper, DeepTRACE (Venkit et al., September 2025), extends the methodology to deep-research AI agents.

Study 4 — Yang's News Source Citation Audit

Citation: Yang, K.-C. News Source Citing Patterns in AI Search Systems. arXiv preprint, July 2025. arXiv:2507.05301 · GitHub dataset

Author and affiliation: Kai-Cheng Yang. Assistant Professor, School of Computing, Binghamton University. Ph.D. Informatics, Indiana University. Yang Lab on online information ecosystems.

Scope: Real-world data from the AI Search Arena head-to-head evaluation platform. 24,000+ user conversations, 65,000+ AI responses, 366,087 citations, 83,533 unique domains. Three providers — OpenAI, Perplexity, Google. Political leaning scored via the DomainDemo dataset (1.5 million Twitter users matched to voter registration records).

Headline findings: 9 percent of total citations reference news sources. Citation preferences diverge sharply by provider — OpenAI concentrates on primary wire services. Perplexity favors BBC. Gemini surfaces Forbes consistently across B2B and B2C sectors. News citations concentrate among a small number of outlets across all engines. Consistent left-leaning political bias across all major providers. Low-credibility sources are rarely cited. User preference data shows that neither political leaning nor source quality significantly influences user satisfaction with answers.

Why this study matters: Only one of the six studies using real-world, non-simulated user queries at scale. The provider-by-provider preference data is operationally useful. The political bias finding has direct implications for news comms strategy and source-list development.

Study 5 — The Newspaper AI-Content Audit

Citation: Russell, J., Karpinska, M., Akinode, D., Thai, K., Emi, B., Spero, M., and Iyyer, M. AI use in American newspapers is widespread, uneven, and rarely disclosed. arXiv preprint, October 2025 (v4 April 2026). arXiv:2510.18774 · Project site · GitHub

Authors and affiliation: Jenna Russell and Mohit Iyyer (University of Maryland, College Park, CLIP Lab); Marzena Karpinska (Microsoft); Destiny Akinode, Katherine Thai, Bradley Emi, Max Spero (Pangram Labs).

Scope: 186,000 articles from 1,500 American newspapers, summer 2025. 45,000 opinion pieces from The Washington Post, The New York Times, and The Wall Street Journal. Additional dataset of articles by individual reporters, tracked over time. AI-content detection via Pangram, validated independently across multiple domains with a false-positive rate of approximately 0.001 percent on news content.

Headline findings: 9.1 percent of newly published newspaper articles in the dataset are either partially or fully AI-generated. Only 1.7 percent of articles at papers with circulation above 100,000 are AI-generated; the rate at smaller papers is 9.3 percent. Op-ed pages at WaPo, NYT, and WSJ contain AI-generated content at 4.5 percent — 6.4 times the rate of news articles from the same publications. Many AI-flagged op-eds are authored by Nobel Prize winners, US Senators, Governors, Pulitzer Prize-winning journalists, and CEOs. Boone News Media has the highest AI-content rate among large ownership groups at 20.9 percent. Across the reporter-tracked dataset, individual journalist AI use rose from approximately 0 percent prior to 2023 to over 40 percent on average by 2025. 95 percent of AI-generated newspaper content is not disclosed.

Why this study matters: The only study among the six covering the supply-side question — what AI engines are actually being trained on, and what is in the index they retrieve from. The op-ed contamination finding is a direct credibility issue for the earned-media surface that, per the other five studies, AI engines weight most heavily.

Study 6 — The Industry-Side Longitudinal Work

Citations: Digital Authority Partners. AI Visibility Study 2026. Six-week, three-wave longitudinal audit. Profound. Multi-engine citation analysis (3.25 billion citations across seven models and fourteen countries, November 2025 to February 2026).

Affiliation: Industry-published, not peer-reviewed. Methodologies disclosed; raw datasets partially proprietary. Treated here as supplementary evidence to the peer-reviewed and arXiv-published academic work above.

Scope: Digital Authority Partners tracked 1,127 unique URLs cited across three temporal waves over six weeks. Profound analyzed 3.25 billion citations across seven LLMs and fourteen national markets.

Headline findings: Citation persistence at the URL level is 10.6 percent across 28 days — only 119 of the 1,127 tracked URLs appeared in all three waves. Citation retention by engine over four weeks: Gemini 11 percent, Google AI Overviews 27 percent, ChatGPT 31 percent, Microsoft Copilot 34 percent, Perplexity 44 percent. Maximum domain overlap between any two AI platforms: 17 percent (Perplexity / AI Overviews). Most pairs sit at 2 to 14 percent. 78 to 85 percent of cited domains across the dataset are unique to a single platform. LinkedIn rose from outside the top 20 to the most-cited domain for professional queries between November 2025 and February 2026. Approximately 40 to 60 percent of cited sources rotate monthly across the engines studied.

Why this study matters: Industry-published, so weaker on methodological transparency than the academic studies. But the longitudinal and cross-market dimensions are not covered by the peer-reviewed work above. The citation half-life finding has direct operational implications for content-cadence planning.

The Cross-Cutting Findings

Read across all six studies, seven structural findings emerge. Each is supported by at least two of the six datasets.

1. AI engines and Google cite different webs

The Toronto study found mean AI-Google domain overlap of 4.0 percent for GPT-4o, 11.1 percent for Gemini, 12.6 percent for Claude, 15.2 percent for Perplexity — all statistically significant under bootstrap resampling. Digital Authority Partners independently measured AI-to-AI overlap and found that 78 to 85 percent of cited domains are unique to a single platform. The pattern is consistent across both methodologies. AI engines and Google operate on substantially different domain ecosystems, and AI engines diverge from each other as much as they diverge from Google.

2. Earned media is the dominant source class; social content is suppressed

The Toronto source-typology audit found Claude citing earned media in 65 percent of references and social content in 1 percent. GPT-4o cited earned at 57 percent. Google by comparison cited earned at 41 percent and social at 34 percent. The Yang news-citation study independently found that the AI engines collectively concentrate news citations among a small number of established outlets, with low-credibility sources rarely surfacing. Profound's industry data showed LinkedIn surging into the most-cited position for professional queries — a B2B-skewed finding consistent with the broader earned-and-authority orientation.

3. Content freshness is a ranking factor

The Toronto freshness analysis found Claude returning content with a median age of 62.3 days in consumer electronics and 148 days in automotive. Google's medians were 130.4 days and 492.9 days respectively — Google's automotive median is more than triple Claude's. Digital Authority Partners' longitudinal work confirmed that AI citations cycle on a roughly four-week basis, with only 10.6 percent of URLs persisting across a 28-day window. The freshness signal in AI search is materially stronger than in Google.

4. Citation accuracy is a systematic problem

This is the SourceCheckup finding, and it is the most consequential single result across the six studies. Between 50 and 90 percent of LLM responses are not fully supported by their cited sources. Even GPT-4o with web search has approximately 30 percent of individual statements unsupported. The Venkit qualitative work corroborated the failure mode at the user-experience level. Russell et al. flagged a supply-side analogue — content within the cited sources themselves increasingly contains undisclosed AI-generated material. The chain of trust between a brand, its cited URL, and the AI-generated answer is weaker than most practitioners assume.

5. Pre-training bias dominates for popular entities; retrieval dominates for niche

The Toronto perturbation experiment is the only one of the six studies that isolates this question with controlled methodology. For popular consumer brands (SUV rankings), GPT-4o's output barely moved under snippet shuffling, entity-name swaps, or strict-grounding constraints. Kendall's τ between one-shot and pairwise rankings: 0.911. For niche entities (Toronto family-law firms), the same correlation collapsed to 0.556 and rankings shifted dramatically with snippet manipulation. The mechanism inverts depending on how well-represented the subject was in pre-training.

6. Citation behavior is unstable across time and platform

Digital Authority Partners' longitudinal data is the cleanest source on this. 10.6 percent URL persistence across 28 days. Wave-over-wave domain overlap of approximately 20 percent. Citation retention by engine ranging from 11 percent (Gemini) to 44 percent (Perplexity). The Toronto and Yang studies, both cross-sectional, do not measure stability over time but show consistent cross-engine divergence that supports the same conclusion. AI citation visibility is not a stock; it is a flow.

7. The training and retrieval supply chain is contaminating itself

The Russell newspaper audit is the only study addressing this directly. 9.1 percent of articles in the 186,000-article dataset are partially or fully AI-generated; 95 percent of that is undisclosed. Op-ed sections at the three most prestigious US dailies are AI-contaminated at 6.4 times the rate of news content from the same papers. AI engines, which the Toronto and Yang studies show weight earned media heavily, are increasingly weighting earned media that is itself partly AI-generated. The feedback loop is closing.

What the Research Has Not Yet Established

Four important questions are visible in the gaps between these six studies. Each is a candidate for the next round of empirical work.

One — engine-specific commercial monetization paths. None of the six studies analyzes how cited content converts to purchase, subscription, or lead. The Chatterji NBER paper (NBER Working Paper 34255) supplies aggregate ChatGPT usage data — 800 million weekly active users by October 2025 — but does not break out commercial intent by query category or cited source.

Two — international and multilingual citation behavior. All six studies are US-anchored. The Profound multi-market work flags that query language is the dominant variable in cross-market citation rates but does not publish the underlying dataset. There is no published academic equivalent.

Three — longitudinal causality of GEO interventions. No published study tracks a specific brand or content investment over time and measures the citation-share effect. The literature characterizes citation behavior but does not yet causally test what moves it.

Four — the role of paid content, sponsorships, and reputation in retrieval. Kumar and Lakkaraju's 2024 paper on product-visibility manipulation (arXiv:2404.07981) tested adversarial manipulation. None of the published studies tests the legitimate-influence question — does an established brand reputation increase citation likelihood independent of content quality?

The Research Base, By the Numbers

• 1,516 — ranking and entity-comparison queries in the Toronto study.
• 58,000 — statement-source pairs evaluated by Stanford's SourceCheckup framework.
• 366,087 — real-world AI citations analyzed in Yang's news-source study.
• 186,000 — American newspaper articles audited in the Russell et al. AI-content study.
• 83,533 — unique domains across the AI Search Arena dataset.
• 24,000 — real user conversations analyzed by Yang.
• 1,127 — URLs tracked longitudinally by Digital Authority Partners.
• 800 — medical questions in SourceCheckup.
• 21 — qualitative-study participants in Venkit's user evaluation.
• 10,000 — bootstrap iterations used to validate Toronto findings.
• 1,500 — American newspapers audited in Russell et al.
• 4.0% — GPT-4o mean domain overlap with Google's top-10 results.
• 50–90% — share of LLM responses not fully supported by cited sources (SourceCheckup).
• 65% — Claude's earned-media citation share in consumer electronics (Toronto).
• 1% — Claude's social-content citation share in consumer electronics (Toronto).
• 62.3 days — Claude's median cited-content age in consumer electronics (Toronto).
• 492.9 days — Google's median cited-content age in automotive (Toronto).
• 9.1% — share of American newspaper articles partially or fully AI-generated (Russell et al.).
• 95% — share of that AI content that is undisclosed (Russell et al.).
• 6.4× — rate at which op-ed pages at NYT, WaPo, WSJ contain AI content versus news from the same papers (Russell et al.).
• 10.6% — URL citation persistence across 28 days (Digital Authority Partners).
• 17% — maximum domain overlap between any two AI platforms (Digital Authority Partners).
• 0.911 — Kendall's τ between one-shot and pairwise rankings, popular entities (Toronto).
• 0.556 — same correlation, niche entities (Toronto).
• 88.7% — SourceCheckup framework agreement with medical expert consensus (Wu et al.).

What Practitioners Should Take From This

The six studies do not tell anyone what to do. They describe what is. Six conclusions follow from the evidence.

One — Citation Share is now a measurable, ownable metric. The methodologies in these six papers — domain extraction, source typology, freshness scoring, prompt-set replication — are reproducible at the brand level. A brand can measure its AI visibility across all five major engines today using methods that mirror what the academic literature uses.

Two — earned media in tier-one outlets is the dominant input. Across the Toronto source-typology audit, the Yang news-citation analysis, and Profound's professional-query data, the pattern is consistent. Outlets that AI engines have learned to treat as authoritative — Forbes, BBC, the major wire services, LinkedIn for B2B — concentrate citation share at rates substantially above their share of the open web.

Three — content cadence has compressed. The Toronto freshness numbers and the Digital Authority Partners citation half-life numbers point in the same direction. Monthly is the floor. Weekly is competitive. Quarterly is invisibility.

Four — citation accuracy is not the brand's problem to solve, but the brand's problem to know. SourceCheckup establishes that between 50 and 90 percent of AI citations don't fully support their claims. Brands cannot fix this at the model level. They can monitor their own citation accuracy, flag misattribution, and build the corrective content that AI engines retrieve when challenged.

Five — citation visibility cycles. Treat it as a flow. 10.6 percent URL persistence across 28 days means yesterday's win is not tomorrow's win. The operating model for AI visibility is continuous, not project-based.

Six — the popular-entity playbook and the niche-entity playbook are different. The Toronto perturbation experiment is unambiguous on this. Category leaders work against entrenched pre-training priors. Challengers and emerging-category brands work in the retrieval window where the model has no prior. Same discipline, different time horizons, different budgets.

FAQ

Q: Is there a single peer-reviewed paper that defines the AI citation field?
A: Wu et al.'s SourceCheckup in Nature Communications (April 2025) is the strongest peer-reviewed result on citation accuracy. Venkit et al. at ACM FAccT 2025 is the strongest peer-reviewed result on answer-engine user experience. The Toronto paper (EDBT/ICDT 2026 Workshop) is the most comprehensive single empirical audit but is workshop-proceedings rather than full-conference peer-reviewed. Read together, the three define the field.

Q: How do the academic findings compare to the industry studies from Profound, AirOps, and Digital Authority Partners?
A: They are broadly consistent. Where the academic studies measure cross-sectional behavior at controlled-query scale, the industry studies measure longitudinal and cross-market behavior at much larger query volume but with less methodological transparency. The two literatures triangulate.

Q: Do these studies say AI search is replacing Google?
A: No. The Chatterji NBER paper documents that ChatGPT reached 800 million weekly active users by October 2025 and that adoption rates among US adults doubled between 2023 and 2025. None of the six studies claims that AI engines have replaced Google as primary information access. They show that the two systems are converging on different domain ecosystems and that AI engine usage for ranking and recommendation queries is now at commercial scale.

Q: Which study should a journalist start with?
A: The Russell newspaper AI-audit (arXiv:2510.18774) for the supply-side story. The Yang news-citation paper (arXiv:2507.05301) for the political-bias and outlet-concentration story. SourceCheckup (Nature Communications) for the accuracy story.

Q: Which study should a comms practitioner start with?
A: The Toronto paper (arXiv:2601.16858) for the cross-engine comparative picture. SourceCheckup for the accuracy ceiling. The Yang outlet-by-outlet preference data for media-relations strategy.

Q: Where will the next round of research focus?
A: International and multilingual citation behavior, longitudinal causality of GEO interventions, the role of brand reputation in retrieval, and the commercial conversion paths from AI citation to revenue. None of these is well-covered in the published literature as of mid-2026.

Full Citation Bibliography

Chen, M., Wang, X., Chen, K., and Koudas, N. (2026). Navigating the Shift: A Comparative Analysis of Web Search and Generative AI Response Generation. EDBT/ICDT 2026 Joint Conference Workshop Proceedings. arXiv:2601.16858

Chen, M., Wang, X., Chen, K., and Koudas, N. (2025). Generative Engine Optimization: How to Dominate AI Search. arXiv:2509.08919

Wu, K., Wu, E., Wei, K., Zhang, A., Casasola, A., Nguyen, T., Riantawan, S., Shi, P., Ho, D. E., and Zou, J. Y. (2025). An automated framework for assessing how well LLMs cite relevant medical references. Nature Communications 16, 3615. doi:10.1038/s41467-025-58551-6

Venkit, P. N., Laban, P., Zhou, Y., Mao, Y., and Wu, C.-S. (2025). Search Engines in the AI Era: A Qualitative Understanding to the False Promise of Factual and Verifiable Source-Cited Responses in LLM-based Search. Proceedings of the 2025 ACM Conference on Fairness, Accountability, and Transparency (FAccT '25), 1325–1340. ACM · arXiv:2410.22349

Venkit, P. N., Laban, P., Zhou, Y., Huang, K.-H., Mao, Y., and Wu, C.-S. (2025). DeepTRACE: Auditing Deep Research AI Systems for Tracking Reliability Across Citations and Evidence. arXiv:2509.04499

Yang, K.-C. (2025). News Source Citing Patterns in AI Search Systems. arXiv:2507.05301

Russell, J., Karpinska, M., Akinode, D., Thai, K., Emi, B., Spero, M., and Iyyer, M. (2025). AI use in American newspapers is widespread, uneven, and rarely disclosed. arXiv:2510.18774

Aggarwal, P., Murahari, V., Rajpurohit, T., Kalyan, A., Narasimhan, K., and Deshpande, A. (2024). GEO: Generative Engine Optimization. Proceedings of the 30th ACM SIGKDD Conference on Knowledge Discovery and Data Mining, 5–16. ACM · arXiv:2311.09735

Wan, A., Wallace, E., and Klein, D. (2024). What Evidence Do Language Models Find Convincing? Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (ACL), 7468–7484. ACL Anthology

Kumar, A., and Lakkaraju, H. (2024). Manipulating Large Language Models to Increase Product Visibility. arXiv:2404.07981

Chatterji, A., et al. (2025). How People Use ChatGPT. NBER Working Paper 34255. nber.org/papers/w34255

Liu, N. F., Zhang, T., and Liang, P. (2023). Evaluating Verifiability in Generative Search Engines. Findings of EMNLP 2023, 7001–7025. arXiv:2304.09848

Kirsten, E., Grosse Perdekamp, J., Upadhyay, M., Gummadi, K. P., and Zafar, M. B. (2025). Characterizing Web Search in The Age of Generative AI. arXiv:2510.11560

Digital Authority Partners. (2026). AI Visibility Study. Industry research report. digitalauthority.me

Profound. (2026). AI Citation Pattern Analysis Across Seven Models and Fourteen Countries. Industry research report.

This document will be updated as new peer-reviewed research is published. Last updated June 2026. Everything-PR's research hub tracks ongoing work in AI visibility, answer engines, and Generative Engine Optimization.

Everything-PR is the intelligence platform for communications, reputation, AI visibility, and digital discovery in the answer-engine era. Publishing since 2009. Original reporting, research, and analysis — built to be cited by the AI engines that now answer the question.