- A generative engine retrieves relevant source material at query time, then uses a language model to synthesize a written answer from it, a fundamentally different pipeline from a traditional search engine's pre-built index and ranked list of links.
- Retrieval-augmented generation (RAG) is the technical architecture underneath most generative engines, not a separate category. It's the mechanism that lets a language model ground its answer in current, retrieved information instead of only what it memorized during training.
- ChatGPT with browsing/search, Perplexity, Google's AI Overviews and AI Mode, and Claude with web search are all generative engines by the retrieval-plus-synthesis definition. A chatbot answering purely from memorized training data, with no retrieval step, is not. It's generating text, but not functioning as a search replacement.
- The mechanical difference matters for visibility work: a traditional search engine's unit of output is a ranked list a person scans and clicks; a generative engine's unit of output is one synthesized paragraph, where a source can be drawn on and never shown as a clickable link at all.
"Generative engine" gets used loosely enough that it's worth pinning down what qualifies mechanically: what happens, step by step, between a user typing a question and a generative engine producing an answer.
What is the technical definition of a generative engine?
A generative engine is a system that answers a query in two steps: retrieval, then generation. First, it retrieves relevant source material (from a live web search, an internal index, or a document store) related to the query. Second, it hands that retrieved material to a language model, which synthesizes a new, written response drawing on it, rather than simply listing the sources for a person to open themselves.
This two-step pattern has a name in the underlying research: retrieval-augmented generation, or RAG. RAG is the architecture; a "generative engine" is the user-facing product built on top of it. The distinction matters because RAG is what lets a language model answer with current, specific, sourced information. A model's own training data has a cutoff date and no awareness of anything published after it, so a generative engine's retrieval step is what lets it discuss this week's news, a company's current pricing, or a recent product launch at all.
Which products actually count as generative engines?
By this definition, the following are generative engines:
- ChatGPT, when browsing or search is enabled for a given response (not every response triggers retrieval)
- Perplexity, which is built around live retrieval as its core product, not an optional add-on
- Google's AI Overviews and AI Mode, which retrieve from Google's search index and synthesize a summary above traditional results
- Claude, when web search is enabled
- Microsoft Copilot, which retrieves from Bing's index
What does not count, by this same definition: a language model answering purely from memorized training data with no retrieval step at all. That's still generation (the model is still producing novel text), but it isn't functioning as a search replacement, since there's no mechanism pulling in anything current or external to what it learned during training. The practical tell is whether a given response cites live sources or dates itself to a training cutoff; if it's the latter, no retrieval happened for that specific answer, whatever product it came from.
How is this mechanically different from a traditional search engine?
A traditional search engine does its heaviest work in advance: it crawls the web, builds an index, and at query time runs a ranking algorithm against that pre-built index to return an ordered list of links. The output is a results page; the "answer," such as it is, lives on whichever page a person chooses to click.
A generative engine does its heaviest work at query time: retrieval happens live (or against a recently refreshed index), and the language model then writes an answer specific to that exact query, in its own words, often blending information from several retrieved sources into a single paragraph. The output isn't a list to scan. It's a finished answer, with the underlying sources reduced to an optional citation a user may never click.
That's the structural reason visibility work looks different for each: a traditional search engine rewards a page for being the single best match a ranking algorithm can find. A generative engine can draw on a page's information without ever surfacing it as a clickable result at all. A brand can be the actual basis for an answer's content and receive zero referral traffic, a gap covered in what does it mean to be cited by AI?
Frequently asked questions
What is a generative engine?
A system that answers a query by retrieving relevant source material and then using a language model to synthesize a written response from it, rather than returning a ranked list of links for a person to click through. ChatGPT with browsing, Perplexity, Google's AI Overviews and AI Mode, and Claude with web search are all generative engines by this definition.
How is a generative engine different from a traditional search engine?
A traditional search engine indexes web pages in advance and, at query time, ranks and returns a list of the most relevant pages for a person to click through. It doesn't generate new text. A generative engine retrieves relevant material at query time and then uses a language model to write a synthesized answer in its own words, with citations (if any) as a secondary layer rather than the primary output.
Is retrieval-augmented generation (RAG) the same thing as a generative engine?
RAG is the underlying technique, not the product. A generative engine is the user-facing system (ChatGPT, Perplexity, an AI Overview); RAG is the specific architecture many of them use internally to ground a language model's output in retrieved, current information rather than relying solely on what the model memorized during training.
Does every AI chatbot count as a generative engine?
No, only ones that actually retrieve and cite external information for a given query. A chatbot answering purely from its training data with no live retrieval step (no browsing, no search grounding) is generating a response, but it isn't functioning as a generative engine in the search-replacement sense this term is normally used for.
Sources referenced
- NVIDIA: "What Is Retrieval-Augmented Generation, aka RAG?"
- IBM: "What is Retrieval-Augmented Generation?"
For how this technical distinction plays out in what actually gets cited, see what does it mean to be cited by AI? For where "GEO" specifically fits against this pipeline, see what is GEO?
