Research commentary: CGO Media Research Team
Executive Summary
Recent research offers useful insights into how AI systems select sources, construct answers and recommend products. The central issue for businesses is that relevant content, credible sources and persuasive explanations do not always lead to the same outcome.
This briefing examines four papers covering source preferences, Google AI Overview citations, evidence-based recommendation and product catalogue search. Three are preprints. One has been accepted to the ACM Internet Measurement Conference 2026. Their findings apply to the systems, datasets and experimental conditions studied; they should not be interpreted as universal rules for commercial AI search platforms.
For CGO Media, the practical value lies in improving how we distinguish and measure retrieval, source selection, evidence use, citation accuracy and recommendation inclusion.
1. Can Source Identity Outweigh Suitability?
Paper: Source Preference in the Wild: How LLM Agents Favor Items by Source, and How to Reduce It
Authors: Jonghyun Song, Haewon Park, Jeonghoon Shim, Woojung Song and Yohan Jo
Submitted: 2 October 2026
Status: Preprint
The researchers examined 12 language-model agents across shopping, accommodation and scholarly search. Their approach compared items from different sources while controlling for display position and the requirements each item satisfied.
Within these experiments, agents showed preferences for particular sources even when the competing items met the same requirements. Source identity could also outweigh differences in suitability.
Among comparisons where exactly one item was selected, a less suitable item from a preferred source beat a more suitable item from a disfavoured source a median 68% of the time. When the source preferences were reversed, the corresponding median was 2%.
The study also tested source identity directly. Hiding identifying information weakened preferences, while changing the source label attached to otherwise unchanged content affected selection.
Why This Matters for Businesses
A business can provide a suitable product or service and still face a disadvantage if an AI agent favours the source presenting a competing option. This makes it useful to investigate both the quality of the offer and the context in which the system encounters it.
The findings do not establish that businesses should pursue a fixed list of preferred domains. Preferences varied across the tested models and settings. They instead provide a reason to test whether source identity affects selection independently of the information supplied.
CGO Media’s Interpretation
For our work on entity authority and recommendation authority, the study supports separating offer suitability from source preference. A recognised source may function as a shortcut for expected quality, but that preference is not itself proof that the selected item is better.
A useful research question is whether a business remains competitive when the same information is presented without its identifying source cues.
Worth reading in full: Sections 4.1–4.2 on observed preferences and selection inversions, followed by Sections 5.1–5.2 on removing and changing source information.
2. Credible Citations Do Not Guarantee Supported Claims
Paper: Measuring Google AI Overviews: Activation, Source Quality, Claim Fidelity, and Publisher Impact
Authors: Haofei Xu, Umar Iqbal and Jacob M. Montgomery
First submitted: 13 May 2026; revised 1 October 2026
Status: Accepted to ACM Internet Measurement Conference 2026
This study analysed 55,393 trending Google queries across 19 topical categories during a 40-day collection period from March to April 2026. It examined when AI Overviews appeared, which sources they cited and whether the generated claims were supported by the captured source material.
AI Overviews appeared for 13.7% of the tested searches, rising to 64.7% for question-form queries. These figures describe the study’s query sample and collection period, rather than all Google searches.
Nearly 30% of cited domains did not appear in the corresponding first-page search results. This demonstrates that the sources cited in AI Overviews were not limited to the domains visible on that first page; it does not, by itself, reveal Google’s underlying selection algorithm.
The researchers also examined 98,020 verifiable claims. Around 11% lacked support in, or contradicted, the captured cited material. Missing page content and changes between collection times could inflate that estimate, so it requires careful interpretation.
Why This Matters for Businesses
Organic rankings and AI citations should be monitored separately. A page’s position in conventional results does not fully describe its presence in an AI Overview.
Businesses should also check how their information is represented. A citation may be visible while an accompanying claim omits a qualification, uses outdated information or exceeds what the source supports.
CGO Media’s Interpretation
This paper is relevant to our AI Search Statistics UK research and our approach to citation measurement. Source credibility, citation presence and claim fidelity represent different dimensions of performance.
Its statistics should retain their collection dates and sampling context wherever they are cited. The study examined trending queries collected in March and April 2026; the October revision is not a new October measurement of Google’s performance.
Worth reading in full: Sections 4.1–4.3 on activation, source selection and claim fidelity, alongside the collection and verification limitations.
3. Does an AI Recommendation’s Explanation Reflect Its Evidence?
Paper: Reasoning with Evidence, Not Merely Rationales: Verifiable Preference Proofs for LLM-Based Recommendation
Authors: Yu Hou, Nathaniel Kang, Pengkai Wang and Hua Li
Submitted: 2 October 2026
Status: Preprint
This paper investigates a gap between the evidence available to a recommender, the explanation it produces and the factors that affect its final ranking.
The proposed PROVE-Rec framework first constructs preference claims linked to selected evidence from a user’s history. A second stage recommends an item using those claims and their supporting evidence.
The researchers test two relationships: whether the evidence supports the preference claim, and whether removing that claim changes the recommendation’s ranking margin. This provides a way to evaluate explanations beyond whether they sound plausible.
Why This Matters for Businesses
When an AI system explains why it recommends a product, that explanation may not reveal what actually drove its choice. Businesses analysing recommendation visibility should therefore be cautious about treating generated reasons as a complete account of the decision.
CGO Media’s Interpretation
The method offers useful ideas for our recommendation-authority research: assess evidence grounding and decision influence separately. Applying similar tests to open-web recommendations would require additional research, because this paper studies recommendation from user histories and reviews.
Worth reading in full: Sections 4.4–4.5 on evidence and claim interventions, followed by Sections 5.4 and 5.7 on grounding and influence.
4. Product Information Must Be Usable Within the AI’s Context
Paper: RPTune: Learned Context Curation for LLM Catalog Search
Authors: Chuxuan Hu, Hejie Cui, Norman Huang, Shubham Kumar Bharti, Wang-Chiew Tan and Sercan Ö. Arık
Submitted: 1 October 2026; revised 2 October 2026
Status: Preprint
RPTune examines product search for smaller merchants whose catalogues fit within a language model’s context window. Its starting point is that supplying the complete catalogue does not guarantee the model will use every product effectively.
The framework learns to prune and order product information before selection, then adapts the model using catalogue-grounded supervision. Across seven real merchants, the authors report context-curation gains of up to 31.4 percentage points in search accuracy.
Why This Matters for Businesses
Making product information available is only part of the problem. The system must also retain and use the details that distinguish the right product from close alternatives.
These results concern a controlled catalogue-search system. They do not demonstrate that rearranging a public product page will deliver equivalent gains in Google, ChatGPT or another commercial search platform.
CGO Media’s Interpretation
The paper is relevant to our Ecommerce Product Discovery and Retailer Selection Model. It reinforces the need to distinguish information availability from effective use during product comparison and selection.
Worth reading in full: Section 3 on the method, Sections 4.2 and 4.4 on results and ablations, and Appendix A.7 on failure cases.
Conclusion: Measure Each Outcome Separately
These papers provide different forms of evidence about AI search and recommendation. Together, they suggest a useful direction for CGO Media’s research: measure whether information is available, retrieved, used, accurately cited and incorporated into a recommendation as separate outcomes.
For businesses, that means asking more precise questions. Is the organisation being found? Is its evidence being used correctly? Is it included in relevant comparisons? Is it recommended for the right reasons?
CGO Media’s interpretation is that this approach can produce more informative audits than citation counts alone. Establishing how these outcomes relate to enquiries, sales and other commercial results requires further measurement.
Visit our Research Library for related work, or read our Research Methodology for how evidence, observations and analytical frameworks are distinguished.
Need to understand your organisation’s visibility in AI search? Contact CGO Media to discuss your search visibility and research requirements.
