Adaptive Retrieval

Adaptive retrieval is a technique in which an AI system decides, question by question, whether to retrieve external information and how much to retrieve, instead of always fetching a fixed number of documents or never fetching any. A simple or familiar question may be answered directly from the model's memory; a hard, current or multi-part question triggers one search, several searches, or a long chain of searches with reasoning in between.

The idea addresses a weakness of basic retrieval-augmented generation. Retrieving the same top few passages for every query wastes effort and adds noise on easy questions, and under-retrieves on hard ones. Two research papers frame the main approaches. Self-RAG (Asai et al., University of Washington, Allen Institute for AI and IBM Research, 2023) trains a model to retrieve 'on-demand' and to critique both the retrieved passages and its own output using special reflection tokens (Asai et al.). Adaptive-RAG (Jeong et al., KAIST, 2024) uses a small classifier to predict a query's complexity and route it to no retrieval, single-step retrieval or iterative multi-step retrieval (Jeong et al.).

Commercial AI search works the same way, even when vendors do not use the term. Google's Gemini API documentation says the model first 'determines if a Google Search can improve the answer' and then 'generates one or multiple search queries'. OpenAI's help center says ChatGPT may send 'additional, more specific queries' after reviewing its first results. At the deep end, Google's Deep Search in AI Mode 'can issue hundreds of searches' for a single question, and agentic deep research modes keep searching until they judge the question answered.

The result is three broad outcomes for any prompt:

• With no retrieval, the answer comes from parametric knowledge and carries no citations.
• With shallow retrieval, the system runs one or two queries and the answer draws on their top results.
• With deep retrieval, the system runs many rounds, and each round's gaps generate narrower query fan-out sub-queries.

Where a prompt lands depends on the engine, the model tier and the conversation. Profound's analysis of about 700,000 ChatGPT conversations from late 2025 found that first questions in a conversation carried citations far more often than later turns, which suggests the retrieval decision is made fresh at each turn and favors opening questions that need facts.

For GEO, adaptive retrieval explains why visibility differs so much between prompts that look similar. On no-retrieval prompts, only long-term presence in training data helps. On shallow prompts, ranking for the one or two queries the model writes is what counts. On deep prompts, distinctive, specific content keeps getting pulled in as the system digs into narrower sub-questions; generic overviews are satisfied early and dropped. Our ChatGPT query fan-out data shows how many searches a typical ChatGPT answer runs. Teams that track grounding rate by prompt can see which of the three modes each important prompt falls into and plan content accordingly: original data, detailed specifications and pages for specific sub-questions for deep prompts, and strong ranking pages for the queries behind shallow ones.