What Is a Generative Engine?

A generative engine is a system that answers user queries by retrieving relevant web sources and using a large language model to compose — generate — a unified response from them, typically with citations. The term comes from the academic literature: Aggarwal et al. formalized it in "GEO: Generative Engine Optimization" (KDD 2024), the paper that also named the discipline of optimizing content for these systems.

How does the original paper define it?

The paper models a generative engine as a pipeline: a query triggers one or more searches; a retrieval system returns a set of source documents; and a generative model synthesizes them into a personalized, precise response with inline attributions. Two properties distinguish it from prior search architecture. First, the response is composed, so multiple sources blend into single sentences. Second, visibility becomes fractional and attributional — the paper measures it via impression metrics like citation frequency and word count attributed to each source in the generated answer, not via ranking positions.

Where do brands get named or dropped?

Inside the synthesis step. The retrieval stage assembles maybe 5-20 candidate sources; the generation stage decides which claims survive into the answer and which sources anchor them. Content that is self-contained, evidence-dense, and directly responsive to the query tends to be quoted and attributed; content requiring surrounding context to make sense gets paraphrased anonymously or discarded. This is the mechanical basis for GEO's core craft rules — answer-first passages, statistics, quotations — which the paper's experiments validated with a 30-40% visibility lift across its 10,000-query GEO-bench benchmark.

Which production systems are generative engines?

All the majors match the definition: Perplexity (retrieval-first with numbered citations), ChatGPT with Search (October 2024 onward), Google AI Overviews (rolled out from May 2024) and AI Mode, Microsoft Copilot, and grounded modes of Gemini and Claude. They differ in retrieval depth, source counts, and citation UI, but share the retrieve-then-generate skeleton — meaning content optimized for the architecture travels across engines.

Example

Asked "how does serverless pricing work," Perplexity retrieves nine pages, then generates four paragraphs in which one vendor's pricing-docs passage supplies the opening definition (citation [1]) and two others contribute supporting details. Three sources achieved visibility; six were retrieved and contributed nothing — the generative engine's selection in action.