Deep Research (OpenAI)

What it does

Deep Research is a feature within ChatGPT that uses OpenAI’s o3 reasoning model to autonomously conduct multi-step web research on a topic and produce a structured, cited report. Unlike a regular ChatGPT prompt — which responds from training data in seconds — Deep Research:

1. Plans a research strategy based on your question (visible as a “thinking” pane)
2. Searches the web dozens of times, following links, reading documents, and refining queries based on what it finds
3. Synthesizes findings across sources into a multi-page structured report with numbered inline citations

A typical Deep Research run takes 5–30 minutes and returns 1,000–5,000 words covering background, key findings, debates in the field, and open questions. Each claim is linked to a source URL.

When it is useful for researchers

Rapid landscape orientation. When entering a new field or reviewing adjacent literature outside your core expertise, Deep Research can produce a useful 20-minute overview that would otherwise take several hours of reading. Treat it as a starting map, not an endpoint.

Grant background sections. The format of a Deep Research report — structured, cited, covering key findings and debates — aligns reasonably well with what a background section of a grant proposal needs. It gives you a draft to edit rather than a blank page.

Identifying key papers and authors. Deep Research reliably surfaces high-citation papers and prominent researchers in a field. Even when the summary of a paper is imprecise, the reference itself is often correct and worth retrieving.

Competitive intelligence on tools and software. For non-technical landscape questions (“what are the main ML frameworks for protein structure prediction?”), Deep Research is faster and more structured than regular search.

Limitations

Sources are web pages, not PubMed. Deep Research searches the web, not academic databases. It will find journal papers when they have public-facing landing pages, but it cannot access papers behind paywalls, search PubMed’s full database, or retrieve PDFs it cannot read. For systematic literature review, this is a fundamental limitation.

Citations require verification. Deep Research’s citations are URLs to source pages, not formatted academic references. The model sometimes misattributes a claim to a source that does not actually say it. Always click through and verify the source text before citing anything from a Deep Research report.

It cannot read PDFs you upload. Deep Research works on live web pages. If the evidence you need is in a paywalled PDF or an internal document, it cannot access it.

Not reproducible. The same query run twice will produce meaningfully different reports depending on what the model searches and finds. This is not a tool for systematic or reproducible literature reviews.

Hallucination persists. The o3 model is better at reasoning than at factual recall. Deep Research can produce confident, plausible-sounding claims about specific statistics, study outcomes, or drug mechanisms that are partially or entirely wrong. The citation may link to a real page that does not support the claim.

Deep Research vs. Elicit vs. Perplexity

For rigorous literature work, Elicit is more appropriate. For fast orientation in a new area, Deep Research produces a more readable starting point.

Related tools

• Elicit — systematic, database-grounded, extractable; better for rigorous literature work
• Perplexity — faster, cheaper, web-grounded Q&A
• ChatGPT — the platform Deep Research runs inside
• Consensus — AI search grounded specifically in peer-reviewed papers