Generative engine optimization is the newest three-word phrase in search marketing, and the pitch travels with it: the rules have changed, your SEO is obsolete, here is a new retainer.
Part of that is true. Most of it is advice you were already given, with the nouns swapped. Here is the honest version: what the research measured, what it did not, and which half of the checklist you can safely ignore.
The term comes from a paper, not from an agency
A group at Princeton published GEO: Generative Engine Optimization in late 2023, and it was accepted at KDD 2024. They built a benchmark of 10,000 queries, rewrote the same source pages nine different ways, and measured how much of each generated answer could be traced back to each version.
The headline number: visibility up by as much as 40 percent, and up to 37 percent against Perplexity, an engine real people use.1
Two of the nine rewrites did most of that work. Adding statistics. Adding quotations. Keyword stuffing, the thing the industry spent fifteen years perfecting, did close to nothing.
That is a genuine result and it is worth knowing. It is also much narrower than the way it gets quoted at you.
What the experiment held constant
A survey published in July 2026 read 45 studies on this, the founding paper included, and states the limit plainly: the retrieval set was fixed.2 The page had already been fetched. What the rewrite changed was how much of it got used, and whether it got named, once it was already inside the context window.
So the 40 percent answers one question. Given that your page was retrieved, how do you get quoted from it? It says nothing about the harder question, which is how your page gets retrieved at all.
The same survey reports that rewrites aimed at being quoted sometimes make a page less likely to be retrieved in the first place. This is a pipeline, and a change that helps at one stage can hurt the stage before it.
The broader finding is the one to keep in your pocket. Across those 45 studies, no technique has shown a stable effect on discoverability that holds up over time and across engines. Audits that run the same prompts repeatedly find low overlap between the sources each engine picks, and high variation between two runs of the same question.
Three things that genuinely differ from SEO
There is no ranking function to game
A search engine returns a list, and a list is stable enough to hold a position in. An answer engine runs a sequence: decide whether to search, retrieve, rerank, write, cite. Ask it the same question twice and you can get two different source sets.
Your position is a distribution, not a number. One screenshot of an assistant citing you is not a measurement.
The unit is a passage, not a page
What lands in an answer is a few hundred words, not a URL. That is exactly how retrieval works on our side of the fence too, which is why the advice in how to write documentation an AI can answer from is the same advice here: one question per section, the answer in the first two sentences, and no sentence that only makes sense if you read the heading above it.
The outcome is a citation, not a click
And often not even a citation. The survey has a name for the case where your content shapes an answer without being named: absorption. It is the honest problem with measuring any of this, because the influence is real and it is invisible in your analytics.
What is identical to what you were already doing
Three of those four rows are the same work you have been doing for years. If your documentation is thin, contradictory or unreachable without JavaScript, no amount of generative engine optimization is the thing standing between you and a citation.
Two decisions sit underneath all of it, and both are already written down here: whether to let the crawlers in at all, in should you let AI crawlers index your site?, and whether the file everyone keeps recommending does anything, in what is llms.txt, and should your site have one?. The practical version of this post, pointed at documentation, is how to get your documentation cited by ChatGPT.
How to measure it without fooling yourself
The survey's measurement advice is the most useful part of it, and following it costs nothing.
- Pick the twenty questions you want to be the answer to, in the words a buyer would actually use.
- Ask each one seven or eight times, not once. The variation between runs is the noise you are trying to see through.
- Paraphrase every question two or three ways, because the engine's own search step rewrites it anyway.
- Keep a control: a handful of questions you are not optimising for, run the same way, so you can tell a real change from a model update.
- Record which sources came back, not just whether you were one of them.
Almost nobody does the fifth one, and it is the only one that tells you what to write next week.
The widget on your site answers from your material and cites the passage it used, which is the same mechanism an external engine runs, over a corpus you control. It changes nothing about whether ChatGPT quotes you. Those are two different surfaces, and anyone telling you that installing a chat widget improves your visibility in generative engines is selling something that is not in the box.
The short version
Fix crawlability, structure and the quality of the answer on the page, because that half is settled and it is the half that pays. Add figures and quotations to the passages you most want quoted, because that is the one tactic with a controlled experiment behind it. Treat llms.txt and schema as cheap bets rather than a plan. And measure with repeats, or do not claim a result.
The rest of what is being sold under this name is SEO, and you already knew how to do that.
