Background:
Population-level cancer surveillance and early awareness efforts increasingly rely on digital traces of health information seeking, particularly web search behavior. The rapid adoption of large language models (LLMs) introduces a potential shift in how individuals access and interpret health information. Unlike traditional search engines, LLMs generate synthesized responses that may contain inaccuracies or hallucinations, raising concerns about misinformation exposure in early-stage cancer information seeking.

Objectives:
To assess whether the widespread availability of LLMs is associated with substitution away from traditional web search for cancer-related information seeking, and to discuss the implications of such shifts for misinformation risk and cancer surveillance.

Methods:
We analyzed weekly Google Trends data from 2020 to 2025 for multiple high-frequency cancer types at the global level. Cancer-related queries were categorized into informational queries reflecting symptom interpretation and early information seeking, and transactional queries reflecting care access intent. In parallel, we constructed an index of LLM-related search interest. Using anchor-based normalization and counterfactual forecasting based on pre-LLM dynamics, we quantified post-period deviations in search behavior. Change-point detection, lagged correlation analyses, negative controls, and placebo intervention tests were used to assess temporal alignment and robustness.

Results:
Globally, informational cancer-related search demand exhibited systematic post-period deviations below counterfactual expectations, while transactional search demand remained stable or increased. These informational declines temporally aligned with increases in LLM-related search interest and were characterized by negative lagged associations, indicating that rising interest in LLMs typically preceded reductions in traditional symptom-oriented search.

Conclusions:
The findings are consistent with a shift in early-stage cancer information seeking from traditional web search toward AI-mediated information sources. While observational, this shift has important implications for misinformation risk, as LLM-generated health information may differ in accuracy, uncertainty expression, and evidence grounding compared to traditional sources. Cancer surveillance and public health strategies that rely on search-based indicators may need to adapt to evolving AI-mediated information pathways and incorporate safeguards to mitigate potential downstream effects of misinformation and hallucinations.