GEO explained: what changes when the answer replaces the link

Every few years the industry gets a new acronym and spends a year arguing about it. GEO — Generative Engine Optimization — is the current one. It is worth understanding precisely, because the narrow version of the idea is real and the loose version is mostly marketing.

Where the term comes from

The phrase entered circulation with a 2023 academic paper of the same name (arXiv:2311.09735), which studied how the content of a source page relates to its appearance inside AI-generated answers. Since then the acronym has been stretched to cover anything involving AI and search. The useful definition is still the original one: shaping content so that a system which generates an answer is more likely to draw on your page — and to attribute it.

SEO competes for a position. GEO competes for inclusion.

A classic search engine returns a ranked list of documents, and you win by occupying a place in it. A generative engine returns one composed answer, assembled from a handful of retrieved sources, and it may or may not name them. That single change moves what you are competing for.

That last row is the uncomfortable one. Search Console reports impressions and average position for Google Search. There is no equivalent public, stable report telling you how often an assistant used your page to answer a question. What most tools sell as AI visibility tracking is repeated sampling: they ask the assistants a fixed set of questions on a schedule and record which sources come back. That is a legitimate method, and it is an estimate — not a number read out of the system.

The retrieval step is where you get in, or don't

Most consumer assistants answer web questions with some variant of retrieval-augmented generation: the question becomes one or more searches, a small set of documents comes back, and the model writes an answer with those documents in front of it. If your page is not retrieved, nothing about your prose matters — it was never in the context window.

Retrieval depends mostly on things SEO has cared about for twenty years: the page can be fetched, it is present in whatever index feeds that assistant, and it contains a passage that unambiguously answers the query. Which is why the honest version of GEO advice looks a lot like good SEO with the emphasis moved — clarity and self-containment over keyword placement.

What is documented, and what is not

Documented

• The major providers publish their crawler user-agents and the robots.txt tokens that control them.
• Several assistants ground their answers in an existing web search index, so classic crawling and indexing still gate access.
• Consumer assistants display citations as links, and those links produce measurable referral traffic in your analytics.

Not documented

• The criteria by which a retrieved document gets cited rather than merely read.
• Whether any specific writing tactic changes inclusion rates in production systems, as opposed to in a benchmark.
• How often answers are recomputed, and how long a change to your page takes to surface.

What holds up either way

Between the two models there is a large, boring overlap. These are the things that serve a ranked list and a generated answer at the same time, and none of them depend on guessing what a model prefers this quarter:

• Content that exists in the HTML the server returns, not only after JavaScript runs.
• A direct answer in the first sentence or two, under a heading that matches the question.
• Explicit entities — who you are, what you do, where you operate — written out rather than implied.
• Facts with their units, dates and sources attached, so a quote survives being lifted out of the page.
• Structured data that matches what a human actually sees on the page.
• A robots.txt that reflects a decision you made, rather than one inherited from a template.

That last one catches more sites than it should. A page cannot be cited by an assistant that is not permitted to fetch it — and it cannot be withdrawn from a model that finished training before you edited the file.