Generative Engine Optimization (GEO)

People now ask an AI system a full question and read the answer before they visit any site, if they visit one at all. A page can rank well in classic search and still be missing from that answer. For a shop or a SaaS company, the answer to a buying question is where the shortlist is made.

The term comes from a research paper of the same name, first published in 2023, which tested ways of rewriting source pages and reported visibility gains of up to 40% in generative engine responses on its own benchmark^[1]. That is a result on a test set, not a promise for a live site, but it showed that what a page says, and how, changes how often a generative engine uses it.

GEO shares most of its ground with good SEO. Google says its AI features need no special optimization: a page has to be indexed and eligible to be shown with a snippet^[2]. Crawlable pages, clear entities, accurate facts and content that answers the question directly do most of the work. What changes is the unit of success, from a position in a list to a place in an answer.

Access comes first. AI search products read the web with their own crawlers. OpenAI documents OAI-SearchBot, for ChatGPT search, and GPTBot, for model training, as separate agents a site controls separately^[3], and Perplexity documents PerplexityBot, which surfaces and links sites in Perplexity’s search results^[4]. In QueryRev’s study of the top sites, 896 of the 6,294 sites whose robots.txt could be read, 14.2%, block at least one AI crawler by name^[5].

Against a fixed set of prompts, asked on each AI surface at regular intervals. For every answer, record whether the brand is mentioned, whether one of your pages is an AI citation, and which competitors are. The headline figures are citation rate and mention rate. Answers vary from run to run, so the trend over many snapshots matters more than one reading.

Read access beside the answers, split by what each crawler is for. A site that blocks training crawlers and allows search crawlers can still be cited; one that blocks a search crawler removes itself from that engine’s answers. Across the top sites, training crawlers are blocked by name more often than search crawlers: 7.9% against 4.8% on average^[5].

Sample data: an invented site, to show the method.

A furniture shop tracks four buying questions on four AI surfaces: sixteen answers. Its pages are cited in eleven and the brand is named without a link in three. The two answers that leave it out share a cause: the sizing guide buries its answer in a long paragraph.

The fix is an answer-first rewrite of that guide, read again against the same sixteen answers in the snapshots that follow. One run is not a result: answers vary between runs, so the change is judged over several.