Verify whether OpenAI Deep Research restricts citations to actually visited URLs

Determine whether the OpenAI Deep Research agent (openai-deepresearch) constrains citation URLs to pages that are actually visited during its browsing process, thereby creating tighter coupling between retrieval and generation that could explain its lower hallucinated-URL rate compared to other deep research agents.

Background

In the comparison between deep research agents and search-augmented LLMs, the authors observe that openai-deepresearch has a notably lower hallucinated-URL rate than gemini-2.5-pro-deepresearch, despite both being multi-step agents. They hypothesize that tighter coupling between retrieval and generation, such as restricting citations to URLs actually visited during browsing, may account for this difference.

However, because the internal architectures of these proprietary systems are not publicly documented, the authors cannot verify whether such a restriction is in place. Establishing whether openai-deepresearch implements a visited-URL-only policy would clarify mechanisms that reduce citation fabrication in agentic systems.

References

The lower hallucination rate of openai-deepresearch may reflect tighter coupling between retrieval and generation—for example, restricting URLs to those actually visited during browsing. We cannot verify this without access to internal architectures.

Detecting and Correcting Reference Hallucinations in Commercial LLMs and Deep Research Agents  (2604.03173 - Rao et al., 3 Apr 2026) in Section 4.2 (RQ2: How do deep research agents compare with search-augmented LLMs?)