Definition
Deep research is an agentic AI mode that takes a complex question, plans a multi-step investigation, searches and reads many sources, and returns a long, structured report with citations. It trades speed for depth: a task runs for several minutes instead of a few seconds, and the output reads like an analyst's brief rather than a chat reply.
OpenAI introduced deep research in ChatGPT in February 2025, and the format has since become a standard feature across assistants. As of October 2026 the main implementations are:
- ChatGPT deep research, available with limited use on the Free and Go plans and fully on Plus and Pro. It runs on the latest model by default, can restrict or prioritize specific websites, reads uploaded files and connected apps, and lets the user edit a research plan before it starts (OpenAI help center).
- Gemini Deep Research in the Gemini app, plus a Deep Research agent in the Gemini API whose April 2026 versions come in a standard and a 'max' variant.
- Perplexity Deep Research, upgraded in February 2026 to run on Anthropic's Claude Opus models and added to Perplexity's Computer agent in June 2026.
- Research in Claude for paid plans, which runs on Claude Opus 5.5 with Opus 4.8 as a fallback.
- Deep Search in Google AI Mode, Google's search-native version, covered under deep search.
A research agent breaks the question into sub-questions, a heavy form of query fan-out. It issues searches, opens pages, extracts facts, notices gaps or contradictions, and searches again. A reasoning model decides when the evidence is good enough to write. Every step is a retrieval event, so the final source list is the residue of dozens or hundreds of fetch decisions. Our ChatGPT query fan-out data shows how far a single prompt expands before any page is read.
For GEO, deep research is one of the most direct links between a published page and an AI output, because the agent fetches pages live at the time of the task. Pages must be reachable by on-demand fetchers such as ChatGPT-User and Google's Gemini Deep Research agent; a robots.txt block or a JavaScript-only page means the agent never reads you. Reports also reward specifics. Original numbers, pricing tables, methodology sections and dated changelogs give the agent something to quote, while generic overviews get skipped because ten other pages say the same thing. And because long reports cite far more sources than a chat answer, a brand that never appears in short answers can still earn citations in a research report on the same topic.
To measure it, run the research-style prompts your buyers would use (vendor comparisons, market overviews, due diligence questions) and record which of your URLs appear in the sources section. Compare that against short-answer AI search results for the same topic: the gap between the two often shows which pages are too thin to survive many retrieval hops.
Examples of Deep Research
- A procurement lead runs ChatGPT deep research on 'SOC 2 compliant CRM vendors under 50 seats', restricts it to vendor sites plus two review sites, and gets a 20-page report whose sources section lists 64 URLs.
- A SaaS company notices its pricing page is cited in Perplexity Deep Research reports but never in standard answers, because the reports reach the detailed tier comparison that short answers skip.
- An analyst at a fund uses Claude's Research mode with Google Drive connected, so the report blends public filings with the firm's own meeting notes.
- A developer-tools brand finds Gemini Deep Research never cites its docs; server logs show the docs return an empty shell without JavaScript, which the team fixes with server-side rendering.
