A growing share of your customers never see a results page. They ask ChatGPT, Claude, or Perplexity, and they read one answer; Google now does the same above its own links. A machine decides which businesses get named in that answer. You can’t control the decision, but you can influence it, and most sites don’t even try.
That is the work GEO names. Generative engine optimization: shaping your content so an AI assistant cites you when it answers. It is the sibling of traditional SEO, which aims to rank your page in a list of links. The two share most of their foundation and differ in the last step: a ranking, versus a mention inside a written answer.
What follows is what we have learned about earning that mention, including the limits.
How an assistant decides who to name
You cannot see inside a model, but you can reason about how these systems work, because the mechanics are public.
An AI assistant answers from two sources. The first is what it learned during training: a snapshot of public text, frozen at some past date. The second is what it retrieves at the moment you ask. Many assistants now search the live web, pull in a handful of pages, and write their answer from those. This second path, retrieval, is the one you can influence directly and quickly.
When an assistant retrieves, it runs a search, reads the top results, and summarizes them. If your page is in that set, is easy to parse, and states a clear answer to the question, it is a strong candidate to be quoted or named. If your page buries the answer under marketing language, or hides it in an image, the model has nothing clean to lift.
So the first rule of GEO is unglamorous: be retrievable, and be clear.
What moves it
Answer the question in plain words, early. Assistants favour pages that state a fact or a definition directly. A page that opens with “What is X? X is…” gives the model a sentence it can quote. A page that opens with a slogan does not.
Structure the page so a machine can read it. Real headings, short paragraphs, lists where lists belong, and a logical order. The same structure that helps a screen reader helps a language model. There is no separate “AI version” of good structure.
Add structured data. Schema.org markup (Organization, Product, FAQ, Article) tells a machine what your content is, not just what it says. It has helped search engines understand pages for years, and it gives answer engines the same head start.
Be consistent everywhere. Assistants lean on facts they see repeated across independent sources. If your company name, what you do, and your key claims read the same on your site, your profiles, and anywhere you are mentioned, the model has less reason to doubt them.
Publish an llms.txt. This is a simple, proposed file at the root of your site that gives AI crawlers a concise, factual summary of who you are and links to your important pages. It is new and not universally adopted, but it costs little and points in the right direction: give machines a clean version of the truth.
Earn mentions elsewhere. Models weigh what other sites say about you, not only what you say about yourself. A citation in something someone else wrote is worth more than another line on your own homepage.
Two things do not help, whatever you may have been sold. Keyword stuffing fails because assistants read for meaning, and padding reads as noise. Walls of thin, generated content fail because there is already too much of it and models are being tuned to discount it. And nobody can promise you a citation: the systems change often, the ranking signals are not published, and the output is probabilistic by design. You can raise your odds. You cannot control what a model says.
What you can’t measure
GEO has a limit that SEO does not. With search, you can at least see your ranking and watch it move. With an assistant, the same question can produce different answers on different days, for different people, in different tools. Measurement is harder and looser. We instrument what we can. We track whether assistants cite you for the questions that matter, and check your content the way a model would read it. Then we tell you where the numbers stop.
The businesses that get named by AI assistants are, overwhelmingly, the ones that were already clear, credible, and well-structured. GEO is the discipline that has always made a site trustworthy, aimed at a new kind of reader.
Where to start
If you do three things, do these. State your core answers in plain language on the pages that matter. Mark those pages up with structured data. Make sure your facts read the same wherever they appear. Everything else builds on that base. This site does all three, down to the llms.txt. View source and check.