Commentary by: CGO Media Research Team
Last reviewed: 16 October 2026
Executive Summary
How does a page become a source in an AI-generated answer? Why might an AI system mention a familiar brand without citing a supporting page? And how can researchers establish whether a correct answer is supported by the evidence collected?
Two papers submitted on 10 October 2026 examine these questions from different perspectives. The first investigates citation patterns, retrieval and model knowledge across commercial search questions. The second evaluates a deep-research system that records and verifies evidence before using it in subsequent reasoning.
Both are preprints. Their findings provide useful research directions, but they should be interpreted within the models, datasets and experimental conditions studied.
For businesses, the practical distinction is between being retrieved, being cited, being mentioned and being recommended. For CGO Media, these papers offer methods for investigating those outcomes and checking whether generated conclusions have adequate evidential support.
1. What Drives AI Citations?
Paper: What Drives AI Citations? Content, Retrieval and Model Knowledge in Generative Engine Optimization (GEO)
Authors: Mahe Chen and Nick Koudas, University of Toronto
Submitted: 10 October 2026
Status: Preprint
The researchers examine two 90-day collections containing 158,099 answers and 468,037 citations. Their study combines observations of commercial search answers with API replays, model-knowledge probes and controlled experiments.
A useful feature of the research is its comparison of pages within the same website. Comparing unrelated websites can make a page feature appear effective when the apparent advantage actually reflects differences between the websites themselves.
Content Relevance Survives the Within-Website Comparison
Within the studied websites, pages whose titles and paragraphs matched the meaning of the questions were cited more often. Common recommendations such as adding FAQ blocks, structured data and summary boxes did not retain the same apparent advantage when the website was held constant.
This does not establish that those features are universally ineffective. It means the study did not find the same independent citation association for them under its within-website comparisons. Their value for accessibility, navigation, conventional search and user understanding remains a separate question.
Retrieval Explains More Than the Measured Page Properties
In the ChatGPT API replay analysis, retrieval explained approximately 32% of the variation examined, compared with 2.8% for the measured page properties. The authors report that every page cited in those ChatGPT replays had been retrieved by the system’s search.
These percentages describe explanatory results within the study. They are not ranking weights, probabilities of earning a citation or estimates of how much a particular website improvement will increase visibility.
The research also examines how engines rewrite questions into searches. This matters because the query used to retrieve a page may differ from the question originally entered by the user.
Brand Knowledge Can Affect Discovery and Selection
The authors report that existing model knowledge influenced which brands appeared in searches and answers. Their controlled experiments also found that changing brand names could alter product rankings while product facts were held constant.
This distinguishes page-level citation from brand-level inclusion. An organisation may be mentioned through the model’s existing knowledge, while a specific page must first be encountered through the retrieval process observed in the study.
Why This Matters for Businesses
Businesses should examine the questions their information answers and the evidence that distinguishes their products or services. They should also measure whether an AI system cites a page, mentions the business or recommends it, because those outcomes can arise through different processes.
The findings do not provide a method for directly controlling model memory. They instead offer a reason to investigate how existing brand familiarity interacts with current retrieved evidence.
CGO Media’s Interpretation
This paper connects our research into AI source selection, entity authority and AI recommendation authority.
Our interpretation is that visibility studies should distinguish pages discovered through retrieval from organisations introduced through existing model knowledge. A citation count alone cannot explain both processes.
Worth reading in full: Sections 4.1–4.2 on comparisons within websites; Sections 6.2–6.6 on retrieval and model knowledge; and the limitations in Section 8.
2. Can an Answer Be Reconstructed from Its Evidence?
Paper: PANINI-DR: Controlled Deep Research through Generative Semantic Workspaces
Authors: M. Yigit Turali, Shreyas Rajesh, M. Efe Lorasdagi, Chenda Duan, Kartik Sharma and Vwani Roychowdhury
Submitted: 10 October 2026
Status: Preprint; under review
PANINI-DR examines how a deep-research system can maintain research state and verify findings across a longer investigation.
The system records subquestions, dependencies, findings and supporting evidence in a shared workspace. Research agents investigate individual questions, while a separate verification process checks proposed findings before they guide further work.
Verification Before Reuse
An incorrect intermediate finding can redirect an entire investigation. For example, identifying the wrong organisation may lead a system to retrieve accurate information about an unrelated business and then build a plausible but incorrect conclusion.
PANINI-DR addresses this problem by checking findings against both their evidence and the subquestions they are intended to resolve. Verification feedback can trigger further investigation before a finding is reused.
Answer Accuracy and Evidence Quality Are Different Tests
The researchers compare the system with seven other deep-research approaches across three benchmarks, using a shared language-model backbone and frozen document collections.
They report the strongest average performance among the evaluated systems. More relevant to citation research, they also assess whether correct answers can be recovered from the evidence saved during the investigation.
This is a useful distinction because a model can produce a correct answer from its existing knowledge even when the evidence it collected is insufficient. Answer correctness alone does not demonstrate that the research process adequately supported the conclusion.
Why This Matters for Businesses
When AI systems support supplier comparisons, market research or purchasing decisions, the evidence record matters. A business needs to understand whether a conclusion follows from the collected material and whether the evidence concerns the correct organisation, product, location and period.
The study evaluates a particular research architecture under benchmark conditions. It does not establish how all commercial AI assistants verify evidence or demonstrate improved business outcomes.
CGO Media’s Interpretation
The evidence-recovery approach is relevant to our work on AI answer construction and our Research Methodology.
A useful future CGO Media experiment would ask an independent evaluator to reconstruct a conclusion using only its recorded evidence. This could help distinguish supported conclusions from correct answers whose provenance remains unclear.
Worth reading in full: Section 2.3 on verifying findings; Section 4.2 on evidence quality; and Appendix D.2.2 on verification before reuse.
What These Papers Suggest for AI Visibility Measurement
The papers address different research problems, but together they support asking more precise questions about how information enters an AI answer.
- Retrieval: was the page returned by the system’s search?
- Citation: was the page visibly referenced?
- Brand inclusion: was the organisation named, and was supporting evidence available?
- Recommendation: was the organisation presented as a suitable option?
- Evidence support: can the conclusion be justified using the material retained?
These are separate analytical questions rather than a fixed sequence that every AI system follows. A brand mention can occur without a citation, and a correct answer can occur without a sufficient evidence record.
For commercial measurement, these visibility outcomes should also be distinguished from enquiries, sales and revenue. Neither paper establishes that an increase in citations automatically produces a corresponding increase in business performance.
Conclusion
This week’s research places attention on retrieval, existing model knowledge and evidence verification. It suggests that businesses need to understand both how their information is discovered and how it is represented once it enters an AI-supported decision.
CGO Media’s interpretation is that better visibility research requires separate measurements for page retrieval, citations, brand inclusion, recommendations and evidential support. Controlled comparisons and clear limitations can help prevent those measurements from being mistaken for universal optimisation rules.
Visit our Research Library for related work, or contact CGO Media to discuss your organisation’s visibility across AI search, GEO and traditional search.
References and Further Reading
- Chen, M. and Koudas, N. (2026).
What Drives AI Citations? Content, Retrieval and Model Knowledge in Generative Engine Optimization (GEO).
Preprint, submitted 10 October 2026.
Paper and publication record. - Turali, M. Y., Rajesh, S., Lorasdagi, M. E., Duan, C., Sharma, K. and Roychowdhury, V. (2026).
PANINI-DR: Controlled Deep Research through Generative Semantic Workspaces.
Preprint, submitted 10 October 2026.
Paper and publication record.
Editorial note: This briefing summarises external research and provides CGO Media’s interpretation. The reported experimental findings belong to the original authors. Proposed applications to CGO Media’s research programme are identified as interpretations or future research directions.
About the author: The CGO Media Research Team publishes research and commentary on SEO, AI search, generative engine optimisation, digital authority and business discovery. Read more about our research team.
