# What Is a Generative Engine and How Does It Work?

> What a generative engine is, the parts it is made of (models and a search engine), where each one retrieves pages from and why it is not the same as an LLM.

URL: https://clione.ai/glossary/generative-engine/
Lang: en

"Generative engine" is a research term, and it has nothing to do with image generators or a video game engine. It names **the system behind a written answer in [AI search](/glossary/ai-search/)**: not the model that writes, but the set of models and search engine that decides which pages to read and what to say with them. It is one of the families of [answer engine](/glossary/answer-engine/), the one that writes with other people's pages. A store should care because every piece of that system is a point where its [product page](/glossary/product-page/) gets in or stays out, and knowing the pieces tells you where to look when it does not show up.

## How it works

The term was set by Pranjal Aggarwal and his co-authors in the 2023 paper that christened [GEO (Generative Engine Optimization)](/glossary/geo/). They present it as a new kind of search engine that uses generative models to gather and summarize information, and they formalize it under a single framework: a generative engine takes the user's query, plus personal information about the user, and returns a natural-language answer.[^geo-paper] Inside it has **two components: a set of generative models, each with a purpose (reformulating the query, summarizing), and a search engine** that returns sources for each search.[^geo-paper]

The flow they propose, which according to the authors closely resembled [Bing Chat](/glossary/copilot/)'s when they wrote it, has three model steps:[^geo-paper]

1. **Reformulate.** A model breaks the query down into several simpler searches, which the search engine understands better.
2. **Retrieve and summarize.** The search engine returns a ranked list of sources, and another model summarizes each one.
3. **Answer.** A third model writes a single answer from those summaries, usually structured text with citations inside.

The citations are not decoration: the authors justify them by language models' tendency to make up information, and they ask of an ideal engine that **every sentence be backed by citations and every citation support its sentence**.[^geo-paper] Those two criteria come from a 2023 audit that measured them on four search engines of this kind and found fluent answers with many unsupported claims.[^liu-verifiability] The paper also describes a conversational, multi-turn version, which fits today's assistants.[^geo-paper]

Google describes the same architecture in its own vocabulary. Its AI features use [retrieval-augmented generation](/glossary/rag/) ([RAG](/glossary/rag/)), which it also calls [grounding](/glossary/grounding/), with its ranking systems bringing pages from its index, and [query fan-out](/glossary/query-fan-out/), a set of related queries the model launches at once.[^google-ai-guide] It is the paper's step 1 under another name.

### Where each engine retrieves from

The search engine in step 2 is not the same in every service, and that decides which door a store has to keep open:

| Engine                                                                           | Where it retrieves from                                                                | What your page needs                                                                                                            |
| -------------------------------------------------------------------------------- | -------------------------------------------------------------------------------------- | ------------------------------------------------------------------------------------------------------------------------------- |
| [AI Overviews](/glossary/ai-overviews/) and [AI Mode](/glossary/google-ai-mode/) | Google Search's index                                                                  | Be indexed and eligible to be shown with a snippet[^google-ai-features]                                                         |
| ChatGPT with search                                                              | The sites it surfaces with [OAI-SearchBot](/glossary/oai-searchbot/), its search agent | Do not block OAI-SearchBot: sites that exclude it do not appear in its search answers, except as navigation links[^openai-bots] |
| [Perplexity](/glossary/perplexity/)                                              | What [PerplexityBot](/glossary/perplexitybot/) crawls, and visits a user requests      | Allow PerplexityBot; Perplexity-User visits generally ignore [`robots.txt`](/glossary/robots-txt/)[^perplexity-bots]            |
| Bing's generative search                                                         | Bing's results, combined with language models                                          | Be in Bing's results[^bing-generative-search]           |

Each one can also use different models and techniques to write. Google warns of it even inside its own house: **AI Overviews and AI Mode can give different answers and links** because they do not necessarily use the same models.[^google-ai-features]

## In a store

A **fictional example**: someone asks, "20-liter dry bag for kayaking that can survive being submerged." Here is how that query goes through a generative engine, and what decides at each step whether a store that sells one gets in:

| Step        | What the engine does                                                                             | What decides whether your product page gets in                                       |
| ----------- | ------------------------------------------------------------------------------------------------ | ------------------------------------------------------------------------------------ |
| Reformulate | Launches searches like "20 L dry bag IPX7," "kayak dry bag reviews" or "IPX6 vs IPX7 difference" | That your page answers one of them, even if not the literal sentence                 |
| Retrieve    | Brings in the best-ranked sources from its index                                                 | That the page is indexed and that its [crawler](/glossary/web-crawler/) can read it  |
| Summarize   | Condenses each source                                                                            | That the protection rating, the capacity and the closure are in text, not in a photo |
| Answer      | Writes and cites                                                                                 | That your page backs the sentence that cites it better than the others do            |

The first row is the one that surprises most. **Your page can get in through a search nobody typed**, the one the engine generated when it reformulated, and stay out of the one the shopper did type.


That the engine searches exactly what the shopper typed. The GEO paper describes a first model that **turns the query into several simpler searches**,[^geo-paper] and Google says its AI features launch related queries at once to bring in more results.[^google-ai-guide] Measuring only your page's position for the literal phrase leaves out the searches through which the engine really finds it.


## How to check it

1. **Open your `robots.txt`** and confirm you let OAI-SearchBot and PerplexityBot through. Changes take up to about 24 hours to show in both.[^openai-bots][^perplexity-bots]
2. **Confirm in [Search Console](/glossary/google-search-console/) that your product pages are indexed** and that nothing stops them from showing a snippet, which is what Google asks for to use them in AI Overviews and AI Mode.[^google-ai-features]
3. **Look in the server logs for visits from each crawler** to your product pages: if an engine does not visit them, it cannot retrieve them.
4. **Write down the searches your customer's question would split into**: the technical feature, the comparison, the reviews. Check that some page of yours answers each one.[^geo-paper]
5. **Run the same question in each engine and note the sources cited**, in new chats and on different days: each engine retrieves from a different place and writes with different models.[^google-ai-features]

## Related concept: LLM

An [LLM (large language model)](/glossary/llm/) is one piece; a generative engine is the system that uses several of those pieces alongside a search engine.[^geo-paper] The difference shows up in the same product: ChatGPT answering without searching is a model answering with what it learned in training, with no sources; ChatGPT with search is a generative engine, which retrieves pages and cites them. **Only in the second case can your page be the source**, and only there does what your page says today count.

To keep going, the [generative summary](/glossary/generative-summary/) explains the text this system produces, and the [AI search reading path](/learn/ai-search/) orders the rest of the terms.