The traffic from ten blue links is still there, but a growing share of the questions never reach a results page. They get answered in a chat window, with two or three sources cited at the bottom. Being one of those sources is a different game.
Generative Engine Optimization is the practice of making a page easy for a large language model to cite. The unit of value is no longer the click. It is the citation, and the brand mention that comes with it, even when no click follows.
This sounds like a small shift. It is not. The optimization target moved from "rank in position one" to "be one of the three pages the model paraphrases when it answers."
Crawlable HTML. Clear page structure. A single canonical URL. Honest titles and descriptions. Internal links that actually go somewhere relevant. If any of this is missing, no model will cite you for the same reason no search engine will rank you.
Reputation also stayed the same. Models lean on the same signals search engines do when they decide which sources to trust. Mentions across known publications, consistent entity data, a real author with a real history.
Three things.
First, answer-shaped content wins. A page that opens with a one-sentence answer to a specific question, then expands, gets cited more than a page that buries the same answer in section four.
Second, entity clarity matters more than keyword density. Models reason in entities. A page that clearly defines what it is about, who wrote it, and what its claims are, is easier to ground than a page that gestures at a topic.
Third, brand mentions without links count now. A model that has seen your name in ten unrelated contexts will surface you more often than a model that has only seen one. This is closer to PR than to backlink building.
A one-sentence answer in the lead. A clear definition of every named thing on the page. Schema.org markup for the entity at minimum, plus Article or BlogPosting where it fits. A visible author block with a link to a real bio.
I also write one paragraph per page that I would be happy to see quoted verbatim. If a model is going to paraphrase me anyway, I would rather hand it the sentence than let it invent one.
Ahrefs Brand Radar tracks how often a brand or domain shows up across the major AI answer engines. Watching that number alongside Google Search Console traffic is how I tell whether the GEO work is paying off without confusing it for ordinary SEO.
The metric I care about most is share of voice on the prompts that matter to me. Not how often I am mentioned everywhere. How often I am mentioned in the conversation I want to win.
Stuffing the page with FAQs that no one asked. Schema markup for every possible type. Trying to game one specific model with prompts hidden in alt text. None of this survives the next model update and most of it makes the page worse to read.
The honest play is to write pages that a careful human would want to cite. Models are converging on the same definition of careful.