How GEO differs from SEO in practice
SEO optimises for a ranked list. The buyer sees ten links and chooses one, so position and click-through rate are the governing metrics. A generative engine does something different: it synthesises an answer from several sources and names some of them. The buyer may never see a list at all, which makes "position" the wrong unit and "was the brand named" the right one.
The two surfaces share almost all of their plumbing, which is why treating GEO as a separate retainer usually produces duplicated work. The same crawl, the same server-rendered HTML and the same structured data feed both. What differs is the content standard: a page can rank well while containing nothing quotable, and a page full of specific sourced statements can be cited in an answer while sitting outside the top ten.
The shift also changes what a flat click count means. Ahrefs, studying 300,000 keywords, found that the presence of an AI Overview was associated with a materially lower click-through rate for the top-ranking page Ahrefs, 300,000-keyword study, April 2025. A programme measured only on clicks reads that as a failure, while the brand may be appearing in more answers than before.
What the research actually found
The term comes from a peer-reviewed paper presented at KDD 2024, which tested content-level changes against generative engines and measured the change in visibility Aggarwal et al., GEO: Generative Engine Optimization, KDD 2024. Two findings are worth stating precisely, because both are routinely garbled in secondary coverage.
Adding quotations from credible sources raised visibility in generated answers by roughly 41%. That is quotations specifically, not statistics, and the distinction matters because the usual summary attributes the figure to adding statistics. Separately, adding cited sources produced the largest single gain the study measured for a page at position five, around 115%.
The practical reading is that generative engines reward content that is easy to verify and easy to lift. A paragraph with a named source, a figure and a link gives the system something it can repeat with attribution; a paragraph of unsupported assertion does not.
What works, and what is sold as working
- Works: being reachable. Assistant operators publish their user agents, and a site that blocks them by accident is not a candidate for citation at all OpenAI, bots and crawler documentation.
- Works: specific, sourced, quotable claims. The study measured this directly, and it is also what a human editor would call good writing Aggarwal et al., GEO: Generative Engine Optimization, KDD 2024.
- Works: entity consistency. One description, one organisation, one named person across every surface, so the system can resolve who you are.
- Works: covering the comparison and alternatives questions buyers actually ask, since those are the prompts that produce recommendation answers.
- Does not work: keyword stuffing aimed at models, which Google addresses directly in its spam policies Google Search Central, spam policies for Google web search.
- Does not work reliably: an llms.txt file as a ranking mechanism. It is a useful courtesy file and no major assistant has committed to it as a ranking input llmstxt.org, the llms.txt proposal.
- Does not work: paying for inclusion. There is no advertising slot inside an organic generated answer to buy.
How GEO is measured
There is no Search Console for assistants. Google Search Console reports impressions, clicks and position for Google Search Google Search Console Help, performance report, and nothing comparable exists for ChatGPT, Claude or Perplexity. Referral traffic undercounts badly, because a buyer who reads a recommendation usually searches the brand name afterwards and arrives as organic or direct. Server logs prove a crawler fetched a page, which is a prerequisite for citation rather than evidence of it.
The practical method is direct observation: a fixed panel of buyer prompts, run on a schedule across each assistant, with the answers recorded verbatim and scored for whether the brand was named and linked. The full method used here is published separately, including its limits.
Who should care about GEO, and who should not yet
It matters most where buyers research before shortlisting and where the purchase is considered rather than impulsive: B2B software, professional services, and high-consideration consumer categories. In those markets an assistant answer increasingly performs the function a shortlist article used to.
It matters least where the decision is proximity-driven or transactional. A plumber chosen from a map pack within three miles, or a commodity product bought on price, is not a case where being cited in a generated answer changes much.
It matters not at all if the site cannot be crawled, has no content worth citing, or sells into a category with no search demand. In those cases GEO is the fourth problem, not the first, and anyone selling it as the first is selling the wrong thing.
Frequently asked questions
What does GEO stand for?
Generative engine optimisation. The term was introduced in a peer-reviewed paper presented at KDD 2024 Aggarwal et al., GEO: Generative Engine Optimization, KDD 2024.
Is GEO the same as SEO?
No. SEO competes for a position in a list of links; GEO competes to be the source a generated answer names and cites. They share most of the underlying technical work.
Is GEO the same as AEO?
Not quite. Answer engine optimisation is about being the extracted answer, often a direct passage. GEO is about being named and cited inside a synthesised answer that draws on several sources.
What single change helps most?
In the published study, adding cited sources produced the largest measured gain, roughly 115% for a page at position five, with quotations adding about 41% Aggarwal et al., GEO: Generative Engine Optimization, KDD 2024.
Does llms.txt improve GEO?
There is no evidence it is used as a ranking input by any major assistant. It is cheap to publish and worth having as documentation, but it is not a lever llmstxt.org, the llms.txt proposal.
How long does GEO take to show results?
In engagements run here, a first assistant citation typically appears inside 30 to 45 days once the foundation is in place. That is an observation from a small number of engagements, not an industry benchmark.
Can anyone guarantee a citation in ChatGPT?
No. Answers are non-deterministic and change without notice, which is why the commitment made here is a measurable impression lift plus at least one citation inside 90 days rather than a guaranteed placement.
Does GEO replace SEO?
No. Ranking still drives the majority of discovery for most businesses, and the technical work that makes a page rankable is the same work that makes it citable.