What is a generative engine?

A generative engine combines two systems: a retriever that finds documents and a large language model that writes an answer from them. The output is prose with optional citations, not ten blue links.

The term was coined by the original GEO paper, which also named the practice of trying to appear in these answers: generative engine optimization. People also say generative search engine or generative AI search for the same thing.

One detail matters for everything that follows. The engine the paper measured was a two-step system its authors built, with the source documents handed to the model directly. No live index was queried. Even its one Perplexity test uploaded source text as files, because Perplexity did not let users specify source URLs.

A live engine moves through four stages. Each has a different owner.

1. Parametric. The model's built-in knowledge, frozen at a training cutoff. Nobody changes it this quarter.
2. Access. A crawler fetches and stores your page, or it doesn't. You and your CDN control this.
3. Retrieval. For a given question, the engine picks which pages enter the model's context. The operator controls this.
4. Generation. The model decides which retrieved pages to use and cite.

Retrieval is not guaranteed. As of October 2026, ChatGPT search decides per query whether to search, and OpenAI says it rewrites your question into one or more targeted queries. Google says AI Overviews and AI Mode may use different models and techniques, and a page must be indexed and snippet-eligible to appear as a supporting link.

If an engine answers from memory, nothing you publish today reaches that answer. If it retrieves, your page must first pass the access stage. That is where the strongest evidence sits.

• Content that only appears after JavaScript runs may be invisible to most AI crawlers.
• robots.txt rules are per bot. OpenAI's OAI-SearchBot controls ChatGPT search, while GPTBot governs training only.
• A firewall or CDN can block a crawler that robots.txt allows.

The comparison of rewrites and access fixes sets out why these blockers come before any wording change. To see where you stand today, the free checker on this site shows whether AI engines name you for up to three buyer questions.

"The GEO paper measured live search." It measured content already inside the model's context. The critical survey of 45 GEO studies found already-retrieved content can change its citation, but no technique shows a stable, cross-platform effect on discoverability. Read what transfers from lab studies to live engines before acting on any effect size.

"There is one generative engine." Each platform runs its own retrieval, models and citation format. Treat them separately.

"Cited means used." OpenAI exposes retrieved and cited URLs as separate lists, and the retrieved list is usually longer. A citation count is a noisy proxy for influence.

• Generative engine optimization (GEO): the practice aimed at these engines.
• GEO, explained honestly: the full guide to what holds up.
• GEO SEO meaning: why "GEO" also means geographic search, and how to tell them apart.

Is ChatGPT a generative engine? ChatGPT with search turned on fits the definition: it retrieves pages, then writes an answer that can cite them. Without search, it answers from its training data alone, which stops at a fixed knowledge cutoff. As of October 2026, the app decides per question whether to search, so the same brand question can get a grounded answer one time and a memory-only answer the next.

Does a generative engine always search the web? No. Retrieval is per query and discretionary on every major platform. Google AI Overviews often do not appear. ChatGPT search is triggered at the model's discretion, and Claude.ai search is a user toggle. When no search runs, nothing on your site can influence the answer. That is why any single visibility number mixes answers given with and without live retrieval.

Is the claim that GEO gives up to 40% more visibility true? That figure traces to the GEO paper, which reported a 30 to 40% band across three methods. It was measured on the authors' own engine with sources already in context, using a metric that counts words the intervention itself added. It says nothing about getting a page found by a live engine, so treat it as a lab result.