AI prompting is erasing the web-browsing data demographers rely on

A quiet change in how people look for information online is starting to undermine one of the most valuable data streams in the social sciences. A perspective article published in Discover Data argues that the migration from open-web browsing to prompting large language models inside closed platforms constitutes a structural break for web-tracking-based demographic inference, not a marginal methodological nuisance. As users increasingly ask conversational agents for direct answers instead of exploring multiple websites, the volume, diversity, and richness of externally observable behavioural signals captured through conventional tracking infrastructures are substantially reduced.

Web-tracking data record how people search, navigate, compare sources, and interact with digital environments. Technologies such as cookies, tracking pixels, browser instrumentation, server log files, and embedded analytics frameworks register page visits, clicks, scrolling behaviour, dwell times, search queries, and navigation paths. Unlike censuses, household surveys, and administrative registers, which are collected episodically and face rising costs, declining response rates, reporting bias, and long delays, digital behavioural traces are generated passively and continuously as a by-product of routine online activity. That high temporal resolution and fine granularity has made them attractive for capturing short-term fluctuations, emerging trends, and responses to external shocks such as economic crises, armed conflicts, or public health events, often at lower marginal cost than traditional collection methods.

The authors, Alberto Martinez-Serra and Alejandro De La Fuente-Cuesta, stress that these traces mainly capture behavioural signals rather than demographic attributes. Age, gender, education, or socioeconomic status usually have to be obtained through surveys, panels, registered user information, or statistical inference, which makes demographic inference from web-tracking data inherently model-dependent. Large language models are themselves emerging as tools for data enrichment, for instance by classifying cryptic URLs into topic categories or inferring the sentiment of search queries at a scale manual coding could never achieve. These techniques may be combined with causal inference frameworks to assess behavioural responses to policy interventions, social shocks, or changes in platform design, and counterfactual fairness frameworks can test whether inference models produce systematically different predictions when protected attributes such as age, gender, or ethnicity are hypothetically altered.

Model performance, however, remains highly sensitive to the quality and representativeness of training data, feature selection, and assumptions about the stability of behavioural-demographic relationships across contexts and over time. Digital behavioural traces tend to overrepresent groups with higher digital access, literacy, and platform engagement while systematically excluding others, and data sourced from opt-in panels of incentivised participants can skew samples toward particular socioeconomic profiles. A further distortion, which the authors call compliance bias, arises when participants in tracked panels shift sensitive or private activities to secondary, untracked devices while keeping the tracked device active mainly for benign ones, creating a gap between observed behaviour and total digital consumption. Contemporary social media interfaces add another blind spot: infinite scrolling and dynamic content loading often do not trigger URL changes, so content exposure, scrolling depth, and micro-interactions remain partially or entirely unobserved, causing recorded events to underestimate actual engagement.

In practice, web-tracking data are rarely used alone. They are linked to survey or administrative records through probabilistic or deterministic matching, or integrated at aggregated levels such as geographic regions, demographic strata, or time intervals, as in studies that calibrate digital proxies against census benchmarks to estimate migration flows or monitor demographic disparities. Post-stratification, raking, and model-based weighting align digital samples with known population benchmarks, while machine learning models fuse behavioural features with demographic variables, using traditional sources as anchors for training, calibration, and validation. In EU jurisdictions, individual-level linkage would normally require a structured governance workflow, including a Data Protection Impact Assessment under GDPR Articles 35-36, purpose limitation, data minimisation, and a valid legal basis for processing under Article 6, and secure-access arrangements at national statistical institutes such as Statistics Denmark and Statistics Netherlands show how sensitive microdata can be made available under strict conditions.

The disruption now under way stems from the interaction model of AI-mediated interfaces. Conversational agents, integrated search assistants, and personalised recommendation systems increasingly powered by large language models compress behavioural pathways that were previously distributed across the open web into a single interface, reducing the observed incidence of exploratory browsing, source comparison, and sequential navigation. Many traditional indicators, including navigation depth, cross-site transitions, search refinement patterns, and browsing diversity, are becoming less informative or less frequently observable. The authors focus on centralised platforms such as ChatGPT, Gemini, and Claude, in which interactions are routed through proprietary servers and retained within the provider’s infrastructure; domain-level access may still be observable, but the content and sequence of interactions are not. They also flag a more extreme scenario: decentralised or locally deployed models, run for example via Ollama or similar open-weight frameworks, process data on-device and generate no server-side traces at all, so widespread adoption would make behavioural data not merely inaccessible but entirely absent.

Existing evidence on AI use is recent and uneven, but it points in a consistent direction. Survey research indicates rapid adoption of generative AI in both work and non-work contexts, workplace studies show conversational AI being incorporated into routine information-processing tasks, and large-scale usage data from ChatGPT identify information seeking, practical guidance, writing assistance, and technical help as prominent categories. Research on perceived credibility suggests AI-generated content can be judged as credible as human-authored content, while users appear to verify AI outputs only selectively, with verification varying by domain, task familiarity, perceived risk, and digital literacy. One widely cited behavioural indicator is Google search interest in the word “translate”, a task historically performed through search engines to reach translation sites: after a long period of growth, a noticeable decline follows the public release of ChatGPT, consistent with some tasks shifting from search-based browsing to direct interaction with conversational systems.

The downstream effects on web traffic are already measurable at the publisher level. According to data from the analytics provider Chartbeat published by the Reuters Institute for the Study of Journalism, Google organic search traffic to more than 2,500 news sites fell 33 percent globally between November 2024 and November 2025, and 38 percent in the United States, while traffic from Google Discover dropped 21 percent globally and 29 percent in the US over the same period. Across all referral channels combined, total referrals fell 24 percent globally and 28 percent in the United States between May 2023 and November 2025. The declines are not uniform: Chartbeat data reported by Axios show search referral traffic fell 60 percent for small publishers attracting 1,000 to 10,000 daily page views over the two years to March 2026, compared with 47 percent for medium-sized outlets and 22 percent for large ones, leaving smaller outlets more exposed because larger, branded publishers can absorb the shock through direct and internal traffic. The authors describe these figures as a structural shift rather than a cyclical dip, and as direct macro-level evidence of the behavioural transition reshaping the web-tracking data landscape.

They are careful not to overstate the case. Web-traffic indicators point to ChatGPT as one of the fastest-growing AI services, yet search-market indicators still show traditional engines dominating, suggesting coexistence rather than full replacement: many users continue to rely on open-web search while a growing share of tasks is routed through conversational systems. Quantifying the shift at the micro level would require tracking data from the same users before and after the diffusion of large language model interfaces, and because tracking data are usually anonymised, such public longitudinal data do not exist, an absence the authors count as part of the methodological problem itself. Aggregate indicators of website browsing, search demand, and AI-platform traffic can still indicate the direction and scale of the transition, even if they cannot resolve individual behaviour.

The implications reach beyond methodology. As observable online behaviour declines across the open web, demography faces a narrowing empirical window into information-seeking, labour market exploration, consumption patterns, and social engagement, and analysis risks becoming dependent on data selectively released by private actors whose commercial incentives may diverge from public-interest objectives. Reduced visibility also complicates policy design, targeting, and evaluation in areas such as digital exclusion, health behaviour, educational engagement, and labour market transitions, particularly as platform-level choices such as encrypted URLs prevent even high-level inference about engagement content. Existing EU regulation offers partial mechanisms: the Digital Services Act creates data-access obligations for designated very large online platforms and search engines under Article 40, which could become relevant if large AI services were designated, and the AI Act introduces transparency and risk-management requirements for providers, though neither solves the problem by itself. The authors call for new governance frameworks, independent public bodies with audit powers, legal authority, and sanctioning capacity, and hybrid data strategies combining surveys, administrative records, and emerging digital sources, warning that without such adaptations, demographic research and social policy risk losing empirical grounding in an era where digital behaviour is both widespread and increasingly hidden within private technological infrastructures.

Subject of Research: How the shift from web browsing to AI prompting reduces the observability of digital behavioural traces used for demographic inference

Article Title: The shift from browsing to AI prompting challenges web-tracking data

Article References: Martinez-Serra, A., & De La Fuente-Cuesta, A. (2026). The shift from browsing to AI prompting challenges web-tracking data. Discover Data, 4(1), Article 23. https://doi.org/10.1007/s44248-026-00122-1

Image Credits: AI Generated

DOI: 10.1007/s44248-026-00122-1

Keywords: web-tracking data, digital behavioural traces, large language models, AI prompting, demographic inference, computational social science, data governance, ChatGPT, publisher web traffic, selection bias, data integration, EU regulation