What generative engine optimization is

Generative engine optimization, shortened to GEO, is the name and framework in the paper GEO: Generative Engine Optimization by Aggarwal, Murahari, Rajpurohit, Kalyan, Narasimhan, and Deshpande. The arXiv record says the venue is the 30th ACM SIGKDD Conference (KDD ’24), 25–29 August 2024, in Barcelona.

The paper describes generative engines as systems that retrieve several sources and use a large language model to summarize them into an answer with source attributions. It defines the creator’s goal as improving the visibility of their content in that answer. Its objective metrics include the share of words tied to a citation, and a position-adjusted word count. Those are the paper’s definitions. OpenAI’s and Perplexity’s product documents do not use them.

The paper says classic SEO targets rank in a search result list. A generated answer weaves sources into one passage, so citation length and position can differ, and rank does not transfer unchanged. Keyword stuffing is one of the paper’s comparison methods, and the paper places it in the group that did not perform well. That statement is about their setup and their metric.

It is not a restatement of Google’s spam policies, and it is not a measurement of ChatGPT as it runs now. Later notes keep those scopes separate.

The abstract says that, in the authors’ experiments, the methods can boost visibility in generative engine responses by up to 40%. The introduction says a Perplexity.ai check showed visibility improvements of up to 37%, and that the effect varies by domain. Section 3.1 describes the experimental engine as taking the top five Google results for a query, then answering with gpt-3.5-turbo at temperature 0.7, with five samples per query. Section 3.2 describes GEO-bench as about 10,000 queries.

Table 1 places Keyword Stuffing in the non-performing group. In that table its position-adjusted overall score is 17.7, against a no-optimization baseline of 19.3. Methods that add citations, quotations, or statistics score above the baseline in the same table. This note paraphrases those cells. It does not reproduce the table.

Lexoford XEO has not rerun the experiment. The 40% and 37% figures are the paper’s account of its own tests. They are not a result a brand should expect in the UAE or Saudi Arabia.

Three limits hold. GEO is a research framework in a public paper, not an official ranking product. A visibility number has to say what is being counted: a mention, a link, or a share of words. Saudi Arabia and the UAE, and Arabic and English, are different question conditions. The paper does not report results for those two countries or for Arabic.