ABSTRACT. Anti-AI Sentiments about Search
How SERP Placement Shapes Search Behavior and Information Avoidance

Introduction

Recently, Generative Artificial Intelligence (GAI) chatbots were introduced into search engines with the intention of altering user engagement. For example, Google’s GAI (Gemini) is used as an agent to create a “search generative experience” that treats user searches as prompts (Venkatachary, 2024). However, many GAI agents on search pages do not provide the option to be turned off. By forcing GAI as a feature into the search pages, any of the users’ ethical concerns about AI end up being dismissed. While many students recognize the benefits of GAI for information, they also express ethical concerns (Chan & Hu, 2023; Gmyrek et al., 2023; Strubell et al., 2019). This paper was interested in exploring how the GAI alters people’s information-seeking behaviors. Prior research found that interface designs can influence user judgement in search results (Huang et al., 2011; Joachims et al., 2005; Pan et al., 2007). Little research looked into whether people actively change their search behaviors to avoid GAI on search pages. Such behavior may be a novel form of information avoidance, whereby users actively avoid information (Hicks et al., 2025), except that users are avoiding the information generation. This abstract’s research question was: How does the GAI agent affect information seeking and do students want to interact with or avoid it? This research-in-progress report describes an experiment on how the average student interacts with GAI is embedded into a search engine results page (SERP).

Methods

A demographic of students (N=104) between 18-24 participated in the experiment. A simulation software was used to either have the GAI agent present or absent on the SERP or in different locations (on the top or side). The students’ task was to query, search, and write about their findings on the topic of health insurance. To observe how the SERP affected young adults’ search behaviors the data collection tools included: think-aloud protocols (Ericsson & Simon, 1993), eye-tracking, interface interaction logs (Novin & Towne, 2025), and interviews. A textual-analysis was used for the think-aloud and interview portion to code for both avoidance and slight resentments towards GAI.

Findings

Earlier SERP studies show that users are influenced by interface cues (Huang et al., 2011; Joachims et al., 2005; Pan et al., 2007). This abstract’s findings are consistent with these prior patterns in search behavior. With the placement of the GAI, we found that if it appeared on the side of the SERP, many participants avoided it (74%), by skipping its content, especially when advertisement links were also present. However, when placed at the top of the SERP, only 9% of participants avoided the GAI content.

During the think aloud portion, we found that participants frequently claimed to be skeptical of AI, yet their behavior showed there were frequent interactions. While 93% found the AI-bot useful, 67% expected the GAI chatbot to appear, and 93% said they were critical or skeptical of AI-generated information. Nevertheless, 93% conducted their search only within the GAI’s text or the links it provided, and only 27% clicked beyond the GAI’s information to traditional SERP results. This means that participants remained within AI’s own framing of a topic. We also found that when the top four results were labeled “Sponsored” (for advertisements), only 13% clicked one of those first four links; without ads, top-result engagement rose sharply. Yet placing AI at the top can keep users engaged above or near commercial content and reduce movement to diverse external sources.

We found a small but growing minority of students are even resenting the heavy usage of AI in academia (N=15). When we asked for reasons that they may avoid AI, they included distrust, but also extended to concerns about creative copyright, labor displacement, and environmental risks. This group was averse enough to AI that they would purposely and actively skip GAI. Notably, unlike other search features the GAI’s “agentic” role made students feel like an unwanted third-party was now documenting the user’s search actions.

Discussion

This abstract builds on earlier work that showed that search engines structure public knowledge rather than merely reflecting neutral relevance (Gillespie, 2014; Introna & Nissenbaum, 2000; Noble, 2018). The findings suggest that users may believe they are critically evaluating AI while actually allowing the AI interface to define the boundaries of verification. While this may partly be due to student passivity, the interface design itself influences ordinary heuristics. When AI appears at the top, provides a succinct answer, and includes its own links, it provides the false sense of a complete search environment. Users may believe they are critically evaluating GAI while the interface defines the boundaries of verification. While AI may encourage students to passively look at information; it is that interface design that influences ordinary heuristics. With GAI producing a new SERP that is curated by AI within the traditional SERP, users receive a SERP-within-a-SERP situation. However, students should be aware that AI is not a neutral feature added to a SERP. Its interface position directly influences whether users see and use it.

Finally, the students who actively avoid GAI in SERPs do so for legitimate concerns that GAI cannot always assist them with. Their anti-AI movements should not be perceived as technophobic, but as an ethically serious response to unconsented training and environmental extraction, amongst other concerns. While students may need AI literacy, it should be delivered with respect to the students’ philosophical conflicts and their frustration with not having the option to refuse participation. The future of GAI in SERPs may require features, such as affording users the ability to turn it off when they please. Beyond researching how AI can be used in education and search for information, scholars should also research: “How does the system’s mechanisms limit and frame the information space?”

ABSTRACT. 1. Objective

This presentation introduces a study examining the search engine visibility of reliable health information websites in Germany across the most prevalent diagnoses encountered in general practice. Grounded in the concept of organisational health literacy, providers of high-quality, evidence-based health information should make accessibility of their contents through commercial search engines as easy as possible (M-POHL, n.d.). By systematically mapping which domains appear in search engine results pages (SERPs) for common health-related queries, this research aims to investigate the extent to which reliable sources compete effectively in the digital information environment.

2. Background

Search engines remain the dominant entry point through which people seek health information online. Finding and identifying reliable health information continues to present significant challenges for lay users. Research consistently shows that people struggle to evaluate the credibility of online health content and often cannot distinguish between high-quality, evidence-based resources and unreliable or primarily commercially motivated sources (Schaeffer et al., 2026).

At the same time, evidence-based health information providers frequently underperform in search engine results, also when their content meets recognised quality criteria. Research shows that websites with lower SEO scores tend to be rated as higher quality by independent evaluators, suggesting a structural disconnect between content quality and search engine discoverability (Schultheiß et al., 2022). This points to a systemic problem: the organisations best positioned to provide trustworthy health information are often not well equipped to ensure that their content reaches users through organic search. This may be due to a lack of technical infrastructure, resources, and expertise (Arnold et al., 2019).

Accessibility is an established criterion within quality frameworks for health information, yet in practice it is rarely operationalised to include search engine visibility. Developers and publishers of patient-oriented health information frequently lack both the institutional knowledge of how search engines function and the practical capacity to implement search engine optimisation (SEO) measures. As a result, high-quality content may be hard to find: medical patient guidelines are published in formats unsuited for indexing (such as PDF rather than HTML), lack optimised metadata, and are misaligned with the language and query patterns that users employ when searching.

This study frames this challenge through the lens of organisational health literacy. While health literacy research has traditionally focused on the capacities of individuals to access, understand, and use health information, the concept of organisational health literacy shifts the focus toward the responsibilities of institutions and information providers. They need to cater relevant information to its target audiences, remove barriers, and design appropriate and modern communication tools. Applied to the digital information environment, this means that organisations providing health information need to understand how search engines work, why and how people use them, and how to structure and publish content in ways that improve its chances of being found by those who are to benefit from it. This study aims to show which types of providers of health information are visible for frequently searched health-related queries.

3. Methods

The study employs a software-assisted, large-scale approach to assess which domains appear in the top ten of search engine result pages – and as sources in AI overviews – for various keyword sets related to the most common diagnoses in German general practice.

Initially, keyword sets are compiled for each included diagnosis, such as sampled queries e.g. on adipositas, diabetes, back pain, hypertension and depression. The keyword sets are identified using software as the Query Sampler tool (Schultheiß et al., 2023) and SISTRIX (2025). Through this keyword research, dozens to hundreds of related keywords are sampled within one set per diagnosis.

In a second step, the top ten organic search results are retrieved for each individual query within these keyword sets. The software used for this is the result assessment tool developed by the HAW Hamburg (Sünkler et al., 2026). This produces a large dataset of domains appearing in SERPs across conditions and query types, enabling an analysis of the health information landscape as users may encounter it.

All domains retrieved from the SERPs are coded according to a classification scheme of publisher types. This classification allows for a structured analysis of which domains and types of publishers achieve visibility across the keyword sets. Main domain categories of health information providers ranking in top ten and being referenced in search engine AI overviews are:

1. scientific: universities / academia

2. journalistic: media outlets

3. commercial: advertisement-financed platforms / industry websites

4. healthcare providers: clinics and practices

5. health insurance providers: public and private insurance websites

6. non-governmental organisations such as foundations, medical associations and patient advocacy groups

Across these categories lies a cluster of pre-defined websites that have been labelled as reliable health information providers based on standardised criteria by non-commercial organisations such as the Stiftung Gesundheitswissen (Engler et al., 2025) and the German Network for Health Literacy (DNGK, 2026).

The combination of query sampling, SERP retrieval, and domain coding enables the study to address two related questions: first, how visible are reliable health information providers for commonly searched health queries? And second, does the distribution of publisher types in the SERPs shift between variables such as topic, search volume and query length?

4. Discussion and Outlook

Findings will identify which domains achieve visibility for the most common health conditions and assess to what degree the predefined reliable sources are represented in SERPs relative to other providers. The results are expected to reveal variation across conditions and domains. Since reliable sources may strategically choose not to create redundant content, the percentage of reliable domains in the overall findings may be low.

These findings will be discussed in the context of organisational health literacy. If high-quality health information providers are structurally underrepresented in search results, this raises the question of how laypeople can be supported in identifying reliable health information if high-quality websites are not transparent or visible in search engine results – especially in light of health information provided by Large Language Models that lack many conventional markers of reliable information such as date of information, authorship and sources (DNGK, 2026).

5. References

Arnold, K., Breuing, J., Becker, M., Nothacker, M., Neugebauer, E., Schmitt, J., & Deckert, S. (2019). Entwicklung leitlinienbasierter Qualitätsindikatoren: Eine qualitative Studie zu Barrieren und förderlichen Faktoren aus der Sicht von S3-LeitlinienautorInnen. Zeitschrift für Evidenz, Fortbildung und Qualität im Gesundheitswesen, 147–148, 34–44. https://doi.org/10.1016/j.zefq.2019.09.001

Bertelsmann Stiftung. (2023). InfoCure: Qualität sichtbar machen. Trusted Health Ecosystems. https://trusted-health-ecosystems.org/infoq-qualitaet-sichtbar-machen/

Deutsches Netzwerk Gesundheitskompetenz (DNGK). (2026). Verlässliches Gesundheitswissen. https://dngk.de/verlaessliches-gesundheitswissen/

Engler, A., Randig, J., Lindner, J., & Ströhlein, L.-M. (2025). Wie erkenne ich gute Gesundheitsinformationen? Stiftung Gesundheitswissen. https://www.stiftung-gesundheitswissen.de/gesundheitsinformationen-verstehen/gesundheitsinformationen

Schaeffer, D., Griese, L., Singh, H., Ewers, M., & Hurrelmann, K. (2026). Zusammenfassung Gesundheitskompetenz in Zeiten gesellschaftlicher Unsicherheiten: Ergebnisse des HLS-GER3. Interdisziplinäres Zentrum für Gesundheitskompetenzforschung (IZGK), Universität Bielefeld. https://doi.org/10.4119/unibi/3017049

Schultheiß, S., Häußler, H., & Lewandowski, D. (2022). Does search engine optimization come along with high-quality content? A comparison between optimized and non-optimized health-related web pages. arXiv. https://doi.org/10.48550/arXiv.2301.10105

Schultheiß, S., Lewandowski, D., von Mach, S., & Yagci, N. (2023). Query sampler: Generating query sets for analyzing search engines using keyword research tools. PeerJ Computer Science, 9, e1421. https://doi.org/10.7717/peerj-cs.1421

SISTRIX. (2025). SISTRIX Toolbox [Software]. Retrieved April 30, 2026, from https://www.sistrix.de

Sünkler, S., Bilir, K., Kumar, T., Koop, O., Schultheiß, S., & Lewandowski, D. (2026). Result Assessment Tool (RAT): An open-source toolkit for conducting studies based on search results. In Proceedings of the 2026 ACM SIGIR Conference on Human Information Interaction and Retrieval (CHIIR '26) (4 pages). ACM. https://doi.org/10.1145/3786304.3787925

WHO Action Network on Measuring Population and Organizational Health Literacy (M-POHL). (n.d.). Organizational Health Literacy (OHL). Retrieved July 12, 2026, from https://m-pohl.net/OHL

ABSTRACT. The spread of online misinformation undermines responsible opinion formation, this risk being exacerbated for heavily debated topics where emotionally-driven reasoning and the formation of echo chambers weaken critical scrutiny. Visual misinformation labels which nudge users towards more mindful information-seeking behaviors are common-place, yet, despite its increasing popularity, the podcast medium remains overlooked in these efforts compared to Web search. This work examines the design of auditory misinformation warnings: brief, non-verbal sound cues embedded in podcast audio to indicate that a nearby statement is misleading. To verify the relevant criteria for these signals’ effectiveness and the factors influencing their reception by listeners, two exploratory crowdsourcing studies were conducted. The first assessed 15 auditory icons in terms of recognizability, consistency of conceptual mapping, and induced disruption, isolating the most favorably perceived candidates. The second embedded the latter as purposeful interventions within AI-generated podcast dialogues on three controversial scientific topics interspersed with myths and falsehoods, employing a between-subjects design across four warning-placement configurations (before, after, enclosing, and concurrent with a misinformation occurrence) and treating several psychometric traits as exploratory factors. The findings indicate salient auditory icons that alert listeners are a viable choice provided they do not severely depart from the podcast’s setting or listeners’ expectations of traditional podcast-editing effects; otherwise, they may trigger prolonged dissatisfaction and ultimately break immersion. No conclusive data on the cues’ role in misinformation recognition emerged, although some listeners correctly inferred their intended function when these followed or enclosed a misleading claim. Warnings placed after a falsehood were least disruptive while concurrent placement was associated with the lowest content recall. Regarding contextual factors, listeners’ prior stance on a debated topic and a podcast’s perceived density and pace exhibited significant correlations with several Likert-scale sound-cue evaluations. Participants’ reactions and follow-up attitudes towards the interventions diverged significantly, anticipation, habituation, and active resistance against perceived paternalism all being expressed. These results yield preliminary recommendations for auditory misinformation warnings and, given the many non-trivial trade-offs at play, outline a multitude of directions for future research.