Generative engine optimization (GEO)

The term was popularized by a 2023 research paper, "GEO: Generative Engine Optimization" (Aggarwal et al., arXiv:2311.09735), which proposed methods for making content more visible in generative search responses and tested them on a benchmark. Since then it has become a common label in marketing and communications. Related labels include answer engine optimization (AEO) and AI search optimization. They overlap, and no standard definition governs any of them.

AI systems produce answers in two ways. Some answers come from what the model absorbed during training. Others come from live retrieval, in which the system searches the web, reads pages and cites them. Different products and different questions use these in different proportions, and the mix changes over time, so any tactic should be tested rather than assumed.

GEO work has three parts. Measurement: ask a fixed set of benchmark questions to each model on a schedule, and record whether the brand appears, what is said and which domains are cited. Diagnosis: find which sources the models draw on for the topic. Action: improve the quality, clarity and availability of the information on those sources, including the organization's own site, and correct inaccurate claims at the origin.

Practical levers commonly discussed include stating facts plainly and near the top of a page, dating and sourcing claims, keeping information consistent across owned and third-party pages, allowing reputable crawlers to read the pages, and earning coverage in the outlets that models already cite. None of these guarantees a result, and the models are not transparent about how they select sources.

PeakMetrics' free AI Perceptions check ran category questions across eight sectors on 28 September 2026, asking each question to ChatGPT, Claude, Gemini, Perplexity and Grok (25 answers per sector, 200 in total). Those answers cited 2,365 links across 1,389 unique domains. The single most-cited domain, forbes.com, accounted for 40 citations, or 1.7%. In other words, citations were spread over a long list of sources rather than concentrated in a few, and there is no single outlet to target.

The same snapshot found that press volume does not settle the outcome. In the airlines sector, Alaska Airlines appeared in 14 of 25 sampled answers and American Airlines in 4 of 25, while American had roughly ten times the news mentions of Alaska over the prior 30 days (32,299 versus 3,138). See AI visibility for the full comparison and its limits.

These figures cover corporate and reputational category questions, the five assistants PeakMetrics benchmarks, and one point in time. Live model answers vary between runs, so a snapshot is not a trend. The citation-sources study also covers consumer shopping questions and Google AI Overviews, where the source mix is different.

People increasingly ask AI systems questions that used to go to a search engine or a reporter. The answer often names a handful of companies, so a brand that is absent may be excluded from consideration before a customer, journalist or investor visits its site. Comms teams own the sources models learn from: press coverage, executive statements, corporate pages. That makes GEO a communications problem before it is a technical one.

GEO is not a replacement for search engine optimization. Many AI systems retrieve pages through search indexes, so the fundamentals still apply. It is also not prompt manipulation or hidden text; attempts to game a model tend to be discounted or reversed. And a single benchmark run does not show improvement. Because results move between runs, meaningful claims need repeated measurement with the same questions.

Nor is it a fixed discipline. Model behavior and product features change, so results should be dated. For the measure GEO tries to move, see AI visibility.