The definition
A generative engine takes a question and composes an answer from sources it retrieved along the way. Assistants, AI overviews inside classic search, and the research and shopping agents built on them are all generative engines. GEO is the work of making your pages usable as one of those sources.
What changed is not a ranking factor. It is that the reader changed shape. A browser renders a page; a parser does not. A person skims past the navigation to find a price; a parser spends context on every byte of that navigation first, and may give up before it arrives. Nearly everything else in GEO follows from that one difference.
GEO and SEO, side by side
GEO builds on SEO rather than replacing it. Classic search still sends most of the traffic on most sites, and the crawlability work it rewards is the same work a generative engine needs. What changes is which properties of a page matter after it has been retrieved.
| SEO | GEO | |
|---|---|---|
| Who reads it | A crawler, then a person | A parser, under a token budget |
| What you win | A position in a list | Being usable as a source |
| Layout | Helps the person who clicks | Costs context before your facts |
| Client-rendered content | Often indexed | Absent from the response |
| Response size | A speed signal | A hard limit on how much is read |
| Feedback | Weeks | One request |
The naming question follows immediately, so it is worth settling: GEO, AEO and LLMO name overlapping work sold by different people. GEO vs SEO takes the comparison apart in full, and LLMO covers why there are four acronyms at all.
What GEO actually requires
- The fetch is allowed. A
robots.txtline or a CDN bot rule ends it here, silently, with nothing logged on your side. - The response is parseable. Anything appearing only after JavaScript runs does not exist, because the fetch runs none.
- The facts are cheap to reach. The engine reads under a budget. A first fact behind tens of thousands of characters of markup may be dropped for a shorter source.
- The facts agree with themselves. When your prose and your structured data differ, the assistant picks one and you do not choose which.
Notice what is missing: keyword placement, word count and heading density. There is no position in a generated answer to compete for, so the tactics built around occupying one have nothing to act on.