What is generative engine optimization? Origin, meaning and limits

Where does the term come from?

From Aggarwal et al., "GEO: Generative Engine Optimization", a paper first posted in November 2023 (its arXiv number, 2311.09735, carries the month) and presented at KDD 2024, the ACM conference on knowledge discovery and data mining. Its authors, several of them at Princeton, coined the term for the search products they called generative engines: systems that answer a query with a written response synthesized from retrieved sources, rather than with a ranked list of links.

The name was built to sit beside SEO. Search engine optimization is work on a page for the benefit of a search engine's ranking; generative engine optimization is work on a page for the benefit of a generative engine's answer. The object is the same, a page. What differs is what the system does with it: a search engine ranks the page, and a generative engine reads it and may or may not use what it says in an answer. The glossary entry on generative engine optimization holds the short definition.

What did the paper test, and what did it find?

It changed source pages and measured how visible each became inside generated answers, on a benchmark of 10,000 queries the authors built for the purpose. The generative engine in the experiment retrieved sources for a query and wrote an answer citing them. The authors then applied one change at a time to a source, adding statistics, adding quotations, adding citations to other sources, changing the tone or the vocabulary, and measured how much of the answer drew on the changed source and how prominently it was cited.

The result that gave the field its number: adding citations, quotations and statistics raised a source's visibility by up to about 40% for the best methods on that benchmark. Keyword stuffing, the one old lever they included, did not help. The finding is about the source, not the query. The authors did not change what people asked or which engine answered; they changed what a page said and watched whether the answer used it. That is the whole premise of GEO as work: the page is the lever, and the lever moves.

Three limits of the paper, which the field tends to leave out. The benchmark was queries, not buyers; the sources were web pages in general, not small-business sites; and 40% is a relative gain on a visibility metric inside an experiment, not a share of anything a business counts. We cite it for the mechanism, that a source's content changes whether an answer uses it, and for nothing else. Our own numbers come from our own audit.

What does GEO mean for a business?

Being one of the names when a buyer asks an assistant, rather than holding a position on a page of links. A buyer who asks ChatGPT, Claude, Gemini or Perplexity for an electrician in their town gets two or three names and a sentence on each. There is no page two. A business that is in that answer is in the running; one that is not is not, whatever it ranks on a results page the buyer did not open.

The consequence is a different unit of measurement. SEO is measured in positions and clicks. GEO is measured in whether the business is mentioned, whether it is recommended, whether it is named first, and who else was named; the reading is called AI visibility. In our audit we put one buyer question per business to Claude, with web retrieval on, for 9,240 businesses: 14.9% were named, and 17.7% among the 5,928 answers that gave a differentiated list rather than a hedge. One assistant, one question each, one day, so a direction and not a rate, and a Claude number rather than anyone else's. It is the kind of reading GEO exists to move.

How does GEO differ from SEO, AEO and LLMO?

GEO, AEO and LLMO name the same practice from three angles; SEO is the older, different practice they sit beside. The full comparison is on its own page and the site's explainer covers the practice at length, so this table is only the terms.

The distinction that matters is not between the three names. It is between ranking and being read. A search engine can rank a page it has not understood. A generative engine has to read the page and decide whether what it says is worth repeating. So the work of GEO leans toward the content of the page (does it state the answer, the name, the place, the price, the date) and the consistency of the business across the web, where SEO leaned toward links and keywords. AEO is filed in the glossary as the same practice under a second name.

What does GEO work consist of?

Five steps, in a loop: measure, make the site readable and consistent, write the answers, earn mentions, measure again. The loop is the method, because none of the middle three can be judged without a reading on either side of it.

1. Measure. Fix a panel of the questions your buyers type, put each one to the four assistants, several runs each, and record whether you are mentioned, recommended or first, and who else is. This is the baseline. Our methodology describes the narrower version we run: twelve questions put to one assistant, Claude, and a visibility score from 0 to 100 that is the share of them in which the business was named.
2. Make the site readable and consistent. The answer-side crawlers allowed in robots.txt, no bot wall in front of them, the text in the HTML rather than in a script or a PDF; one name, one address, one phone number on every page and in LocalBusiness markup. In our audit, 43.1% of sites failed the consistency test on their own three main pages, and 7.2% sat behind bot protection that served a challenge page.
3. Write the answers. On each service page, the questions buyers ask as headings, each answered in one complete passage that names the business, the place and the price range, with a date and an author. 61.5% of audited sites had nothing of the kind.
4. Earn mentions. Reviews quoted as text on your own pages, with the platform named; listings that agree with the site; mentions in the places the assistants read, which for a local trade means the review platforms and directories buyers already use, and any press or association page that names you. A mention needs no link.
5. Measure again. The same panel, the same assistants, a month later. What moved, on which questions, against whom. The next month's work is whatever that reading says it is.

That is also the shape of what we sell, with the measuring done on one assistant, Claude. The audit is the first reading and the list of what to fix; optimization is steps two to four, with a fixed scope set by the audit; monitoring is step five, monthly, on the same twelve questions.

What GEO cannot do

It cannot guarantee a place in an answer, and anyone who promises one is describing a lever that does not exist. No agency has access inside ChatGPT, Claude, Gemini or Perplexity. We have no arrangement with OpenAI, Anthropic, Google or Perplexity, and neither does anyone else selling this. The assistants choose what to name from what they can read. GEO changes what they can read about you, and measures whether the choice changed.

It cannot make an assistant say something your pages do not support, and it should not try. A passage that overstates what you do is a passage the assistant may repeat to a buyer who then finds out. The work is to make the true facts easy to find, quote and verify, which is duller than the promise and holds up better.

And it cannot be judged from one reading. Answers vary between runs of the same question on the same day, so a single screenshot of an assistant naming a competitor is a data point and not a verdict. We date every reading and read the pattern across twelve questions, and with monitoring across months, not any one answer. The measurement is the product. The rest is what the measurement tells you to do.