What is generative engine optimization?
Generative engine optimization (GEO) is the practice of making a business, person or page more likely to be retrieved, trusted and cited by AI systems that write answers, such as ChatGPT, Google’s AI Overviews and AI Mode, Gemini, Claude, Perplexity and Microsoft Copilot. Where search engine optimization (SEO) aims for a high position in a list of links, GEO aims for a mention inside the answer itself, and for that mention to be accurate.
The name was introduced in “GEO: Generative Engine Optimization”, a paper by Pranjal Aggarwal, Vishvak Murahari, Tanmay Rajpurohit, Ashwin Kalyan, Karthik Narasimhan and Ameet Deshpande, first posted in November 2023 and published in the proceedings of the ACM SIGKDD conference (KDD ’24) in Barcelona. The paper defines a generative engine as a system that retrieves sources and writes a single synthesised response, and treats visibility inside that response as the thing to optimise.
The same work is also called answer engine optimization (AEO), AI search optimization, or LLM optimization. The labels differ; the job is the same. It is spelled optimisation in British English.
Why does GEO matter now?
Because a growing share of research starts with a question typed into an assistant rather than a search box, and an assistant’s answer behaves differently from a results page in three ways that matter to a business:
- It names one or two, not ten. There is no second page to be on. A business is in the answer or it is not.
- It arrives earlier. People ask assistants while they are still working out what their problem is and who handles it, before they would have searched for a service and a town.
- It reads as advice. A recommendation inside a sentence carries more weight than a position in a list, and it often produces no click at all, so the business never sees the visit.
Search is still the larger channel for almost every business, and nothing here argues for abandoning it. The case for GEO is that the assistant channel is new, its winners are not yet settled in most markets, and the slowest part of the work runs on other organisations’ schedules.
How do AI assistants decide which sources to cite?
Most assistants that answer questions about the current world use some form of retrieval-augmented generation: they look things up, then write from what they found. The details differ by product and change often, but four steps are common, and a business can be dropped at any of them.
- The question is rewritten. A vague question becomes several narrow searches. Google describes this for AI Overviews and AI Mode as “query fan-out”: issuing multiple related searches across subtopics and data sources. Your page competes against those sub-questions, not against the question as typed.
- Pages are fetched. The assistant takes what its crawler or search index can reach and parse. A page blocked in robots.txt, or one whose main text only appears after a script runs, can return nothing.
- Sources are checked against each other. Before naming a business, systems look for agreement between independent records: the business’s own site, its profiles, directories, registers and press. Contradictions make a business harder to confirm, and an assistant that cannot confirm a business tends to name one it can, or a directory instead.
- An answer is written from what survived. Short, self-contained passages that make one clear claim are easy to lift. Sentences that depend on the paragraph above them lose their meaning when quoted alone.
Our register entry How an AI assistant decides which business to name walks through each step with the ordinary reasons good businesses fail at it.
How is GEO different from SEO?
GEO builds on SEO rather than replacing it. Crawlable pages, clean structure and a trustworthy reputation help with both. The differences are in what is ranked, what the user sees and how success is measured.
| Aspect | Search engine optimization | Generative engine optimization |
|---|---|---|
| What is chosen | Whole pages | Passages, facts and named entities |
| What the user sees | A ranked list of links | A written answer naming one or two sources |
| The win | A top position and a click | Being named, cited and described correctly |
| Key signals | Relevance, links, page experience | Consistent records, quotable passages, corroboration, crawl access |
| How it is measured | Rankings, clicks, traffic | Named share and citation share on a fixed question panel |
| Stability | Positions move gradually | The same question can produce different answers on different runs |
For a longer treatment, see GEO and SEO are not the same job.
What does GEO work actually involve?
In practice, GEO is five workstreams, and their order matters. Doing them out of order is the most common way a budget is spent with nothing to show for it.
- Entity reconciliation. Make every public record of the business agree: the exact name, address, phone number, people, services and licence details on the website, Google Business Profile, professional registers and main directories. This is the slowest step, because other organisations make the changes, and it is usually the one that unblocks everything else.
- Access. Let the crawlers that power AI search reach your pages, and make sure the main text is in the HTML rather than loaded by a script. Our AI crawler robots.txt generator separates the search crawlers from the training crawlers so you can choose deliberately.
- Answer-first content. Rewrite the pages that answer real client questions so the answer arrives in the first sixty words, in sentences that still make sense quoted alone. Our answer-first checker flags the usual problems.
- Structured data. Describe the business, its services and its people in Schema.org markup that matches what the page visibly says. Its value is machine readability, entity clarity for search engines and rich results, not citations: in an Ahrefs study published in May 2026, adding schema did not measurably change AI citations for pages already being cited, and assistants fetching pages live read the visible text. Put every fact on the page; the schema generator produces a validated starting point for the markup.
- Corroboration and measurement. Get accurate facts stated in sources you do not control, and measure the result on a fixed set of questions run on a schedule, so movement can be shown rather than claimed.
Does GEO actually work?
There is real evidence that the way content is written changes how visible it is inside generated answers, and there is good reason to be careful about how far that evidence goes.
In the original GEO paper, the authors built a large benchmark of real-world queries across many domains, called GEO-bench, and tested nine ways of rewriting source content. Adding relevant statistics and adding quotations from credible sources performed best, improving visibility by roughly 40% on the paper’s position-adjusted word-count measure. Keyword stuffing, the classic SEO shortcut, did not help. The paper’s own summary is that GEO “can boost visibility by up to 40% in generative engine responses.”
Three limits apply. The experiments ran on test engines built for the study, not on the live versions of ChatGPT or Gemini. The products change how they retrieve and attribute sources several times a year. And Google states plainly that there are no additional requirements to appear in AI Overviews or AI Mode beyond being indexed and eligible for a snippet in Search. The defensible reading is that GEO removes the reasons a business is excluded and makes its content easier to quote. Nobody can guarantee that an assistant will name a particular business, and any agency that does is overstating what it controls.
How do you measure GEO?
There is no position one in an answer, so rankings do not translate. Measurement works on a fixed panel of questions that real clients would ask, put to each engine in fresh sessions on a schedule, with every answer stored and dated. Three numbers come out of it:
- Named share
- The share of answers on the panel that mention the business at all.
- Citation share
- The share of answers that link to or cite one of the business’s own pages as a source.
- Accuracy
- The share of mentions that describe the business correctly: right services, right place, right people, still open.
Because the same question asked twice can produce different answers, single screenshots prove nothing. The comparison that means something is the same panel, run the same way, a quarter apart, with competitors measured on the same run.
What doesn’t work?
- Keyword stuffing. It performed poorly in the GEO research, and repetition makes passages less quotable, not more.
- Writing for bots instead of people. Assistants quote passages a person would find useful. Pages written to game a model read badly to the client who arrives.
- Markup that disagrees with the page. Google asks that structured data match the visible text. Markup claiming things the page does not say is a liability.
- Treating llms.txt as a lever. The llms.txt proposal from Jeremy Howard (September 2024) is a sensible courtesy that costs an hour. It is a proposal, not a ranking input you can count on, so it should never be the plan. Our llms.txt generator builds one in a minute if you want it.
- Promises of movement in thirty days. Registries and directories take weeks to propagate a correction. The end of month three is the earliest point at which a before-and-after comparison means anything.
Where should a business start?
Ask the assistants. Open ChatGPT, Perplexity or Gemini in a private window and ask who to call for your service in your town, then ask it to describe your business. What comes back tells you which problem you have: nobody named means an open market; competitors named means you are being found and dropped; you described wrongly means an identity problem in your public records. The six-question self-check on our homepage maps your answers to the step to fix first, and the free visibility sweep measures it properly across forty questions and five engines.
If you run a law firm, GEO for law firms covers bar listings, practice-area pages and advertising rules. Every term on this page is defined in the GEO glossary.