A search engine finds documents. A generative engine writes an answer.
That one difference reorganises everything downstream of it. The old contract was that a search engine identifies candidates and the human does the synthesis by clicking through three of them. A generative engine does the synthesis itself and hands over a paragraph. The links, when they appear at all, are footnotes to a decision that has already been made.
The loop, in four steps
Every generative engine runs some version of the same pipeline. The details differ, the shape does not.
1. Interpret the query. The user’s question gets expanded into machine queries. Google confirms that both AI Overviews and AI Mode may use a “query fan-out” technique, issuing multiple related searches across subtopics and data sources. So the query the engine runs is usually not the query the user typed, and often several of them run at once.
2. Retrieve candidates. Documents come back from an index. That index is built by crawlers, and for most engines it is a fairly conventional search index. Perplexity documents PerplexityBot as the crawler “designed to surface and link websites in search results on Perplexity”. OpenAI documents OAI-SearchBot for surfacing sites in ChatGPT’s search features and recommends allowing it in robots.txt. Google is blunter: robots.txt directives for Googlebot are the control for managing access to AI features in Search.
3. Ground the generation. The retrieved text is handed to the model and the answer is constrained to it. This is the step that decides whose words end up in the output. See AI grounding for what that constraint actually does.
4. Attribute. Some engines link generously. Some link sparingly. Some ground in a source and never show it. Attribution is a product decision made after the answer already exists, which is why citation counts and influence are not the same measurement.
Each of those four is a filter. A brand can be eliminated at any one of them, and the failure modes look identical from outside: you are not in the answer.
What structurally changes for a brand
Ten positions become one answer. Position four on a search results page still gets traffic. There is no position four in a generated paragraph. There is inclusion and exclusion.
Passages compete, not pages. Retrieval and grounding operate on chunks. Your 4,000 word pillar page does not compete as a unit. Its sections compete individually, which is good news if you write in clean self-contained sections and bad news if you write long undifferentiated prose.
The query multiplies. Fan-out means you are being scored against subqueries you never targeted and cannot see. Optimising for one head term is a strategy built for a system that no longer runs.
Being right beats being persuasive. Marketing copy grounds badly. A model synthesising an answer needs assertions it can support. Superlatives are unsupportable by construction.
The click is optional. The engine can use your content fully and send you nothing. This is the part of the model that most damages measurement, because your analytics show a flat line while your influence changes.
What did not change
Plenty, and the overcorrection is expensive.
The retrieval layer is still an index built by crawlers that mostly obey robots.txt and mostly want plain HTML. Google states there are “no additional requirements to appear in AI Overviews or AI Mode, nor other special optimizations necessary”. That sentence is doing a lot of work: it means the eligibility gate is ordinary search eligibility. If you are not indexed, nothing downstream matters.
So the boring technical layer got more important, not less. Crawl access, render-ability, canonical hygiene, internal linking. Most AI crawlers do not document JavaScript execution at all, and neither Perplexity’s crawler documentation nor OpenAI’s mentions rendering. If your content needs a browser to exist, check what those crawlers can actually see.
Generative engine, answer engine, AI search
Three names, heavy overlap, minor differences in emphasis.
- Generative engine emphasises that the output is written.
- Answer engine emphasises that the output resolves the question, which predates LLMs and included featured snippets.
- AI search is the loose consumer term for the same category.
The optimisation acronyms follow the same pattern. GEO and AEO describe substantially the same work with different framing. Do not let vocabulary shopping distract from the pipeline, which is what actually determines outcomes.
The short version
A generative engine retrieves then synthesises. Four steps, four places to be eliminated, and only the last one is visible to you.
Work the chain in order. Confirm a search crawler can reach you, using the AI crawler directory to check which agent does which job. Confirm your content exists in raw HTML. Then make each section answer one question specifically enough to be quoted.