Deep Research

Deep research is a type of agentic search feature in AI assistants. Instead of answering immediately, the system plans a research task, runs many searches, reads and compares sources, and after several minutes produces a long report with citations. Google introduced Deep Research in Gemini in December 2024; OpenAI and Perplexity launched their own deep research modes in February 2025, and Google added a Deep Search mode to AI Mode in 2025.

How it works

A deep research agent usually combines a reasoning model with search and browsing tools. It breaks the question into sub-questions, a heavier form of query fan-out, then iterates: search, read, note gaps, search again. Google describes Deep Search in AI Mode as using the same fan-out technique “taken to the next level”, able to issue hundreds of searches; Perplexity says its mode performs dozens of searches and reads hundreds of sources. The final step synthesises findings into a structured, cited document.

How it’s used in SEO and marketing

• Research tasks: market and competitor overviews, literature scans and briefing documents.
• Visibility: deep research reports cite many more sources than a quick answer, so being a well-structured, verifiable source on a topic creates more chances to be cited.
• Content strategy: the sub-questions these agents explore show which angles a complete resource on a topic should cover.
• Measurement: visits from research agents and their users may show up as AI referral traffic or in server logs as user-triggered fetches.

Example

A marketer asks a deep research agent for “a comparison of EU VAT rules for selling digital courses”. The agent searches tax authority pages, accountancy guides and recent news, notes contradictions, searches again for the latest thresholds, and returns a structured report with dozens of citations. An accountancy firm with a clear, dated, well-sourced guide on the topic is far more likely to be among those citations than one with a vague overview.

MLforSEO’s comparison of NLP APIs and generative AI included Gemini 1.5 Pro with Deep Research; on that entity extraction task it returned no entities at all, a reminder that research agents are built for synthesis rather than structured extraction.

Limitations and things to watch

• Reports can still contain errors and misattributed citations; check the sources before relying on them.
• Features, limits and availability change quickly and differ by plan and country.
• Agents favour accessible pages. Content blocked in robots.txt, behind logins or rendered only by heavy JavaScript is easy to miss.
• Research agents are slower and costlier per query than standard answers, so they are typically used for complex questions.

Learn it on MLforSEO

• Comparing NLP APIs for Entity and Sentiment Analysis (vs Generative AI)
• How Entities Power Query Fan-Out in AI Search
• Your audience is no longer only human
• AI Search Optimisation & Agentic SEO (course)