Generative Engine Optimization

What it is

Generative engine optimization is the practice of changing web content to increase how prominently a source is cited in an answer generated by an AI search system. The term comes from a paper by Aggarwal and colleagues, first posted in November 2023 and published at KDD 2024, which formalised the generative engine — a system that gathers and summarises sources to answer a query — and proposed GEO as a way for content creators to "increase their visibility (or impression) in generative engine responses". Its headline number is one metric's ceiling: the abstract claims GEO "can boost visibility by up to 40% in generative engine responses", while the body reports "a relative improvement of 30-40% on the Position-Adjusted Word Count metric and 15-30% on the Subjective Impression metric". Both are impression metrics the authors defined, and neither is a citation rate. In the experiment "only the top 5 sources are fetched from the Google search engine for every query", and that set is then held fixed, so what improved was the prominence of a source already in front of the model, not the chance of being retrieved at all. The authors state that they "didn't evaluate how GEO methods affect search rankings".

Why it matters

The tactics are young and contested, which is why this entry is flagged for review rather than treated as established practice. C-SEO Bench, accepted at NeurIPS Datasets and Benchmarks 2025, retested Aggarwal's own content transformations across "two search tasks, question answering and product recommendation, with three domains each" and with several sites competing at once, and found most of them "not only largely ineffective but also frequently have a negative impact on document ranking", with conventional retrieval-focused SEO doing considerably better. Its breadth is in the second task and the competition, not in topical coverage: Aggarwal's benchmark already spanned "25 diverse domains such as Arts, Health, and Games". How far that generalises is itself disputed: GEO-Bench — an arXiv preprint by Nimase and colleagues with no stated venue, posted in May 2026 and substantially revised in October 2026 (v3), and not the GEO-bench Aggarwal’s paper introduced — scores 15 methods under one protocol, using “every item in five datasets as a target (729 targets)”, and C-SEO Bench is only one of those five. Earlier versions scored everything against a single fixed open-weight ranker; in the revision the main ranker is still Llama-3.1-8B-Instruct, but every result is re-scored on three further LLM rankers, among them the proprietary GPT-4o-mini, and the paper reports that “The order of methods is consistent across LLM rankers.” Its caution about datasets softened in the same revision — “the order of methods is moderately stable across datasets, but the size of the gains is not” — and neither paper measures whether an unretrieved page gets retrieved at all. The underlying shift is real: visibility is increasingly decided inside a generated answer rather than on a page of ranked links. That does not mean ranking stops mattering — C-SEO Bench's own conclusion is that "C-SEO will not replace SEO, but will complement it". Either way, a citation is only worth having if the sentence being cited is accurate. Knowledge Company sells the accuracy rather than the visibility for that reason. Optimising to be quoted more often while your claims go unmaintained simply distributes the error faster.