A buyer asks ChatGPT which tool to pick. Another asks Perplexity to compare two vendors. A third Googles your category and reads the AI Overview without scrolling to a single blue link. In each case a model reads many sources, writes one answer, and decides which names to cite. When your brand is missing from that answer, nobody outranked you. You were never in the room.
The work of earning those mentions has a name now: generative engine optimization, or GEO. It rhymes with SEO on purpose. The goal stays familiar, which is to be the source a discovery surface trusts. How you get there has changed, and for the first time there is real research to point at instead of vibes.
How GEO differs from SEO
Classic search is a ranking game. You compete for a slot on a results page, and a person decides what to click. Answer engines collapse that funnel. The model reads a set of sources, writes one reply, and shows a few citations or none. There is no page two to climb. You are part of the answer or you are invisible.
That shift is already mainstream, not a forecast. BrightEdge tracked Google AI Overviews growing from 31 percent of queries in early 2025 to 48 percent a year later, and most US searches now trigger one. The old assumption that a top-ten ranking guarantees a citation is breaking too. In mid-2025 roughly three quarters of pages cited in AI Overviews also ranked in the organic top ten for the same query. By early 2026, depending on whose data you read, that figure had fallen to somewhere between 17 and 38 percent. Ranking still helps. It no longer decides the answer.
What the research actually proves
The term comes from a 2024 KDD paper by Aggarwal and colleagues at Princeton, IIT Delhi, Georgia Tech, and the Allen Institute for AI. They built a benchmark called GEO-bench, ran roughly 10,000 queries, and tested nine content edits to see which ones got a page cited more often in generated answers. This is the rare case where a marketing tactic has a controlled study behind it.
Three edits moved the needle, and they share a theme. Adding statistics with named sources lifted visibility by about 41 percent. Citing credible outside sources and adding direct quotations from named experts produced gains in the same band, with the paper reporting up to a 40 percent lift overall. The pages that benefited most were the ones outside the very top of the rankings, which is exactly where most brands live.
The failures are as useful as the wins. Padding a page with more words did roughly nothing. Keyword stuffing, the reflex SEO trained into a generation of marketers, performed about 10 percent worse than the untouched baseline. It was the one tactic that actively hurt. The authors are also explicit that results varied by query domain, so the right move for a debate-style question is not the right move for a factual lookup. Treat the numbers as direction, not a recipe.
What correlates with being cited
The lab study tells you what to do on a page. Field studies of live engines fill in the rest, and they keep pointing at the same handful of signals.
Be quotable and structured. Engines extract claims, so a clean heading, an explicit definition, and a direct question-and-answer passage let a model lift your sentence without rewriting it. Claude in particular leans toward structured, bullet-pointed content. This overlaps with plain technical SEO and clean semantic HTML. It just matters more when a machine, not a human, is doing the reading.
Get mentioned, not just linked. Ahrefs studied 75,000 brands and found brand web mentions correlated with AI visibility at 0.664, against 0.218 for backlinks, roughly three to one. As their content director Ryan Law put it, this is not a content-creation arms race, and building links for volume does not move the needle. One confident page about your category is weak. The same framing echoed across your docs, third-party write-ups, and community threads is strong.
Stay fresh and crawlable. Recency is weighted heavily, especially on Perplexity, where recently updated content earns several times more citations than stale pages. None of it matters if the engine cannot read you. ChatGPT search draws on Bing's index, so a site with Google coverage but Bing gaps can be invisible to it regardless of how good the page is. Schema.org structured data and clean, server-rendered HTML keep you parseable.
You will hear that an llms.txt file is the key to AI visibility. Be skeptical. In July 2025 Google's Gary Illyes said flatly that Google does not support it and is not planning to, and John Mueller likened it to the long-discredited keywords meta tag. Adoption studies put it on roughly one site in ten, and no major answer engine treats it as a ranking input. Spend the hour on quotable pages instead.
Why you cannot just buy a citation dashboard
The honest constraint of GEO is measurement. The engines disagree about almost everything. One analysis of 100,000 prompts found that ChatGPT and Perplexity share only about 11 percent of their cited domains, and that 71 percent of cited sources appear on a single platform. ChatGPT favors Wikipedia, Perplexity leans hard on Reddit, and Google AI Overviews still track organic rankings more closely than either. Optimizing for one tells you little about the others.
Worse, the answers are not stable. SparkToro's volunteer study found the same prompt run twice in one day can return different citations, which is why serious audits run a hundred prompts per query just to get a baseline. So when a vendor offers a tidy dashboard of every URL a model cited, distrust it. Nobody has a reliable global feed of which pages which model named.
What teams actually track instead is indirect and honest: referral traffic from AI engines in analytics, lift in branded search volume, and periodic manual prompt audits where you ask the real engines your buyer's real questions and write down who got named. It is more work and less satisfying than a number on a screen. It is also the only thing that is true today.
A checklist for this week
You do not need a GEO department. You need your most important answers to be easy to find, easy to parse, and consistent everywhere they appear.
- List the ten questions a buyer asks before choosing a tool like yours, and name the canonical page that should answer each. Gaps are your first work.
- Open each page with its claim in the first paragraph, then back it with a specific statistic and a named source. That single edit is the best-supported tactic in the research.
- Add one direct, quotable sentence per page that a model could lift verbatim.
- Mark up pages with schema.org structured data and confirm the HTML renders without JavaScript.
- Verify your full site is indexed in Bing Webmaster Tools, not just Google, or ChatGPT cannot see you.
- Run a manual prompt audit: ask your top ten questions in ChatGPT, Perplexity, and an AI Overview, and record who gets cited.
- Chase mentions, not just links. A few credible third-party write-ups outweigh a pile of backlinks.
The message-matched page that wins a campaign click also works as a source an engine can cite. Iterant builds pages that way by default: clean semantic markup, schema.org structured data, citable passage structure, and crawler-level access control. The model drafts, you approve, and nothing ships on its own. We make pages eligible to be cited. We do not promise a citation, and you should distrust anyone who does. The surface is new, the competition has not fully arrived, and the work is the kind you can start this week.