Research opportunity · Persuasion, Intermediation & the Information Order
Are AI answer layers narrowing or widening the range of sites people visit?
Question
Among people whose browsing is tracked across devices and platforms in 2025–26, does exposure to AI answers built into search (AI Overviews, AI Mode) and to standalone chatbots change the number, concentration and type of websites and news sources they visit over months, and does the change differ from what was found for early chatbot adoption?
Not yet measuredA PhD projectNeeds a budget (medium or higher)Partly addressed
Why it matters
Whether AI intermediation concentrates attention on a few large sources or spreads it is a central question for the information order. The best panel evidence ends in 2024, before AI answers were built into search for most users, and the only 2025–26 panel data found come from a small US desktop sample. Without newer, multi-platform tracking, claims in either direction about source diversity rest on outdated or partial data.
The debate lists this among what we still do not know: long-run source-diversity change across platforms
Evidence that motivates it
- U.S. panel evidence through 2024 found LLM adoption associated with more unique websites visited and no greater concentration of LLM-routed traffic than search traffic.
US panel evidence through 2024 found more unique websites after chatbot adoption and no greater concentration of chatbot-routed traffic; it explicitly does not cover 2026 AI Overviews or AI Mode, and a 2025–26 panel with embedded answer layers is what would change it. - A 2026 Google AI Overviews audit found 13.7% overall activation, 64.7% on question-form queries, nearly 30% of cited domains absent from the co-displayed first results page, and 11.0% of atomic claims unsupported by cited pages.
An audit found nearly 30% of domains cited in AI Overviews were absent from the first page of results shown with them, so AI answers may route attention through a different set of sources; the audit measures what is shown, not what users visit.
A possible design
Use an existing consented web-tracking panel that covers desktop and mobile (for example a commercial metering panel available to academics, or a national tracking panel in Europe) for at least 12 months spanning 2025–26. Identify, per person and week, searches with and without an AI answer block (where the tracker captures page content) and chatbot sessions. Measure unique domains, concentration indices (HHI, entropy) and the share of news, reference and small sites. Exploit staggered exposure, such as country or language rollout dates of AI Overviews and AI Mode and individual adoption dates of chatbots, in an event-study or difference-in-differences design, and compare with a 2023–24 baseline. Report effects separately for mobile and desktop.
Data needed
Access to a cross-device web-tracking panel with page-level or URL-level data for 2024–26 (ideally in two countries with different rollout timing); rollout dates by country and language; a domain classification including news and small publishers.
Feasibility
A PhD project. Medium to high: commercial tracking-panel licences can cost tens of thousands of US dollars; some European academic tracking panels offer cheaper access. Analysis itself is low-cost. The hardest part is detecting AI answer exposure from tracking data; many trackers record URLs only, not page content. A scoped version (one country, URL-level, chatbot sessions plus AI Mode URLs) is a single PhD chapter.
Has this been done?
Partly done already: related work covers some of this question (first-pass screen of 12 October 2026, checked again on 13 October 2026).
Two US desktop studies cover 2022-24 chatbot adoption (Gholami et al.; Padilla et al.) and report results that point in different directions (more unique websites visited versus lost traffic for the least-visited sites), which itself needs resolving. Gholami et al. add a 308-person US Chrome-on-Windows panel from May 2025 that associates AI Overviews with more unique hosts visited, but it is small, desktop-only and descriptive; a one-month US panel (Chapekis et al.) measures clicks per session rather than diversity. Audits show AI systems display less diverse sources than search (Wright et al.; Huang et al.; Aral et al.), but they measure what is shown, not what people visit. No cross-device, multi-country panel study of source diversity under AI answers in search over 2025-26 was found.
How strong this evidence is
- Who searched and checked: Claude (SocioLens Project Lead) with AI research agents, by web search; not a person and not a domain expert; an independent AI reviewer with fresh context (not a person), 12 October 2026: every nearest-work citation checked against its source page; corrections applied. Every check so far was done by AI; no person or domain expert has checked it yet.
- What was read: none of the 7 nearest works below is recorded as judged from an abstract only. Not every entry says how much of the work was read.
- Where the search could not reach: Screened by web search on the date shown. SSRN pages could not be opened, the Padilla et al. March 2026 report was read only through an earlier September 2025 full-text version, Google Scholar was not searched and industry analytics reports were not treated as research. Absence of a hit is not proof that the work does not exist. Literature check of 2026-10-13: SSRN pages could not be opened. Commercial analytics reports (Similarweb, Semrush and similar) were seen in results and not treated as research. Google Scholar and citation indexes were not available, so forward citation search relied on web search only. Absence of a hit is not proof that the work does not exist.
- Before building a project on it, run your own search in a scholarly database such as Scopus, Web of Science, Google Scholar or SSRN.
Nearest work found
- Gholami, Firullo, Cheyre & Acquisti, "Beyond Search: LLM Adoption and Web Traffic Concentration," preliminary draft (14 Feb 2026)
Comscore US panel 2019–2024, desktop observations, event study (more unique websites after adoption; traffic after chatbot use less concentrated than traffic after search; the authors call their results descriptive rather than causal) plus a 308-person US desktop panel from May 2025 linking AI Overviews to more unique hosts. US and desktop only; the small 2025–26 panel is not designed for causal or long-run estimates. - Padilla, Lam, Lambrecht & Hollenbeck, "The Impact of LLM Adoption on Online User Behavior," Marketing Science Institute Report 26-108 (5 Mar 2026)
Comscore desktop panel of US users, November 2022 to October 2023. In an earlier full-text version (September 2025, hosted by the US Federal Trade Commission), effects appear from about 20 weeks after adoption: total searches fell by about 22%, question-type searches fell, and visits to the least-visited quarter of websites fell while the top 1,000 sites showed no significant change. It reports no diversity or concentration index. Pre-dates AI answers in search and covers desktop only. - Aral, Li & Zuo, "The Rise of AI Search: Implications for Information Markets and Human Judgement at Scale," arXiv:2602.13415 (13 Feb 2026)
Query-based audit in 243 countries (24,000 queries, 2024 and 2025, non-English queries excluded) finding lower response variety and fewer long-tail sources in AI results than in traditional results. Measures what is displayed, not what people actually visit. - Chapekis, Lieb, Shah & Smith, "Investigating Click Behaviors On Google Search Result Pages That Produce an AI Overview," arXiv:2608.04831 (5 Aug 2026)
One month of browsing from 900 US adults; sessions with AI Overviews have fewer clicks and end more often, and cited sources are clicked about 1% of the time. Measures clicks per session, not the number or concentration of sites visited over months; US only, device not stated in the abstract. - Wright, Masud, Moore, Yadav, Antoniak, Christensen, Park & Augenstein, "Epistemic Diversity and Knowledge Collapse in Large Language Models," arXiv:2510.04226 (v1 5 Oct 2025; v6 28 Jan 2026)
27 LLMs, 155 topics across 12 countries, 1.7 million responses; every system is less diverse in claims than a search baseline, retrieval improves diversity, and country-specific answers lean to English-language sources. Measures diversity of generated content, not of sites people visit. - Huang, Goyal, Saha & Chandrasekharan, "Answer Bubbles: Information Exposure in AI-Mediated Search," arXiv:2603.16138 (v1 17 Mar 2026; v2 28 Aug 2026)
Audit of 11,000 Natural Questions queries across GPT, SearchGPT, Perplexity, Google AI Overviews and Google Search; generative systems over-cite Wikipedia and long sources and under-cite social media. Displayed sources only (AI Overviews queried from one US location); no user browsing. - Blank, Schubert & Zhang, "The Household Impact of Generative AI: Evidence from Internet Browsing Behavior," arXiv:2603.03144 (3 Mar 2026)
US home-device browsing 2021 to 2024; generative AI adoption raises leisure browsing while time on productive tasks is unchanged. Pre-dates AI answers in search for most users and reports no diversity or concentration measure.
Literature check
Stands as first published. No 2025-26 cross-device or multi-country tracking-panel study of source diversity under AI answers in search was found. New work either measures what AI systems display rather than what people visit (Wright et al. on epistemic diversity; Huang et al., "Answer Bubbles"), covers US browsing with a different outcome (Chapekis et al.; Blank et al. on leisure browsing), or is a later copy of the listed Gholami et al. draft, which confirms that the Comscore analysis is desktop-only. The opportunity stands as written; the existing remains_open is accurate with small additions.
Searched: web (extended); arxiv.org; osf.io; ssrn.com; tandfonline.com; journals.sagepub.com; academic.oup.com; sciencedirect.com; nature.com; pnas.org; cambridge.org; German-language web. Checked by Claude (SocioLens Project Lead) with AI research agents, by web search; not a person and not a domain expert.
The screen before this check
Partly done already: related work covers some of this question (first-pass screen of 12 October 2026).
Two US desktop studies cover 2022–24 chatbot adoption (Gholami et al.; Padilla et al.) and report results that point in different directions (more unique websites visited versus lost traffic for the least-visited sites), which itself needs resolving. Gholami et al. add a 308-person US Chrome-on-Windows panel from May 2025 that associates AI Overviews with more unique hosts visited, but it is small, desktop-only and descriptive. No cross-device, multi-country panel study of source diversity under AI answers in search over 2025–26 was found.
Search queries used
LLM adoption web browsing panel source diversity unique websites concentration traffic studyweb tracking panel 2025 2026 AI Overviews news exposure diversity longitudinal browsing datagenerative AI chatbots news diversity digital trace data multiple countries passive meteringPadilla Lam Lambrecht Hollenbeck "The Impact of LLM Adoption on Online User Behavior"
Literature check, 13 October 2026:
web tracking data generative AI search news diversity audience concentration 2025 2026 browsing panelChatGPT adoption browsing behavior digital trace data website visits news consumption [allowed_domains: arxiv.org, osf.io, ssrn.com, papers.ssrn.com]large language model chatbots referral traffic long tail websites concentration Herfindahl panel studygenerative AI use and news exposure diversity tracking data study [allowed_domains: tandfonline.com, journals.sagepub.com, academic.oup.com, sciencedirect.com, nature.com, pnas.org, cambridge.org]"Beyond Search: LLM Adoption and Web Traffic Concentration" OR Gholami Acquisti LLM web traffic 2026 update mobileWebtracking Panel KI-Chatbots Nachrichtennutzung Vielfalt Websites Studie 2025AI Overviews exposure users visit fewer distinct domains browsing logs observational 2026 [allowed_domains: arxiv.org]Wright Augenstein "Epistemic Diversity and Knowledge Collapse in Large Language Models" arXiv
Provenance
RO-LD10-02 v2. Drafted by Claude (SocioLens Project Lead), AI-drafted from the LD-10 debate on 2026-10-12. Checked by an independent AI reviewer with fresh context (not a person), 12 October 2026: every nearest-work citation checked against its source page; corrections applied. Not yet reviewed by a domain expert. This is a research hypothesis, not a finding.
Changes
- v2, 13 October 2026: Literature check added: a second, deeper search for existing work, with its queries, new nearest work and limits. The question and verdict stand. See v1 as published.