Generative engine optimization, GEO for short, is the practice of improving your visibility in the answers generative engines build, so your pages become the sources the engine quotes when someone asks ChatGPT, Perplexity or Google’s AI Overviews a question. A generative engine doesn’t rank pages, it runs searches, reads what comes back and writes an answer from passages it lifts, so your page competes as one passage inside one answer to one question. The term came from a research paper in late 2023, which measured which tactics move visibility and which actively hurt it.
You can point an AI SEO platform at the buyer questions and see which ones you get quoted in, the same way you’d watch rankings. The tactics with measured support are quotations and statistics with named sources, and keyword stuffing measured negative. Freshness moves too, and the measurement is a rate over repeated runs, never a single answer.
What is generative engine optimization
Generative engine optimization is the practice of improving your visibility in answers built by generative engines, the AI systems that answer a question by drawing on multiple web sources rather than returning a list of links. The ground it covers is the ground search work has always covered. The surface is new: a written answer a buyer reads, on your site or a client’s, instead of a list of results.
The name has a paper trail, which is rare for a search label. It comes from “GEO: Generative Engine Optimization”, a paper by Aggarwal and colleagues first posted in November 2023 and accepted at the KDD 2024 conference, which gave the practice its name, built the first benchmark for it, and reported that the right changes to a page could boost its visibility in generative engine responses by up to 40%. So when a claim about GEO circulates, you can usually check it against something that was actually measured.
Usage varies, and it helps to know that up front. Generative search optimization is the same practice under a synonym, and the two names get used interchangeably. Some vendors sell the same work as answer engine optimization (AEO) or AI SEO, and classic SEO is the discipline it sits inside rather than a rival to it.
This guide keeps to GEO, the head term with the paper trail.
What is a generative engine
A generative engine is an AI system that answers a question by running searches, reading multiple pages, and writing an answer that quotes passages from them, with citations back to where each passage came from. The ones buyers use are ChatGPT, Claude, Gemini and Perplexity, plus Google’s AI Overviews and AI Mode, which build answers inside search results.
The mechanism matters more than the label. Because the engine builds an answer instead of ranking pages, your page competes as the passage the engine lifts for one question. Result seven is still a fine SEO ambition, and it no longer decides whether you get quoted.
And the reading goes deep before anything gets recommended. In our Credibility Layer study of how ChatGPT recommends local businesses, 200 reasoning sessions produced 5,933 searches and 1,982 citations, and every recommendation came out of a credibility audit of about 30 searches rather than the first result the model saw.
Reddit was cited zero times. So the engine reads widely, weighs what it finds, and quotes the pages that survive the weighing.
What actually moves visibility in generative answers
In generative engine optimization, the tactics that work are the ones with measured support: quotations and statistics with named sources, plus freshness and quotable structure. Keyword stuffing is the one tactic measured to hurt rather than help, and it’s the cheapest one to fall into. The numbers behind that ranking come from two places, and they agree.
Most GEO numbers you’ll see quoted trace back to GEO-bench, the benchmark that came with the original paper. In that GEO benchmark, adding quotations raised a page’s visibility in the answer by 40.9%, and adding statistics raised it by 30.6%; keyword stuffing came in at -8.3%, the worst result of anything tested. Our study, The 3 Numbers That Decide If AI Cites Your Page, records the figures, from Aggarwal and colleagues, and the computation behind them: a no-optimization baseline of 19.3, gpt-3.5-turbo writing the answers, five averaged seeds on a 1,000-query test split.
It’s one benchmark on one model, so treat it as a lab result rather than a live results page. Our own field work points the same way. In a 39-page sample, the pages that appeared in the AI Overview averaged 26 chunks (headings plus list items) per 1,000 body words, and the pages left out averaged 16, which is a gap you can close with structure alone.
A model scores one chunk of a page at a time, so what you’re really optimising is one passage against the passages of the pages already cited, and the ones that win cleared 105 or higher on chunk score. What made them liftable were extractable units: statistics with a named source, quotations, and claims that stand on their own.
Freshness moves too. In our title-tags study, 1,000 searches across five AI engines, adding the current year to a title tag lifted AI citations by as much as 192%, and title length, brand suffixes and question titles did nothing. In our 10x AI Visibility Refresh, 98 existing posts were refreshed with nothing new published, and citations went up 10x in 91 days.
The convergent read matters more than any single number. A benchmark on one model, a 1,000-search experiment on title tags, and a 91-day refresh all agree that specific, attributed, quotable content wins and stuffing loses, and correlations stay correlations. Treat the numbers as evidence about what to try on your pages.
How to do generative engine optimization
The work starts from the questions buyers ask and works down to the page, its sources and its freshness, in that order.
-
Start from the questions buyers ask. Write down the exact questions a buyer would type into ChatGPT before buying from you, and work those instead of a keyword list. A generative engine answers questions, so the question is the unit of targeting, and a keyword that isn’t a question rarely becomes an answer anyone reads.
-
Make each answer one liftable chunk. Answer the question in the first sentence under its heading, in the searcher’s words, in short paragraphs. Then benchmark the page against the pages already cited for that question: the 26-versus-16 chunk gap and the 105 chunk score threshold are the bar to clear.
-
Add quotations and statistics with named sources. These were the two strongest tactics in the benchmark, and they cost nothing but the discipline of attributing what you say. A claim a model can attribute is also the better page for a human, and the stuffing shortcut is the one that measured negative.
-
Keep it fresh. Put dates on facts that have them, keep the year current in the title tag where that’s honest, and refresh on a cadence rather than publishing your way out of staleness. The 10x refresh got its result from updating existing posts alone, with nothing new published over those 91 days.
-
Be present in the pages the engine already reads. Every recommendation in the credibility study came from a 30-search audit of third-party pages, so off-site presence is part of the work rather than separate from it. That means reviews, industry lists, directories and the publications the engine keeps finding, and it means checking where the engine actually reads rather than assuming.
-
Measure as a rate over repeated runs. One answer is one sample, and one sample is basically a coin flip. The number that means anything is the share of runs that recommend you, which is where measuring starts.
What goes wrong with GEO
The generative engine optimization mistakes that repeat are keyword stuffing, fake freshness, paying for markup Google Search ignores, and measuring badly. Stuffing keywords for the model is the worst GEO tactic the benchmark tested, and it makes the page worse for a human at the same time. Fake freshness is the same mistake wearing a date, and it passes your staleness on to anyone who quotes you.
Then there’s the markup shelf. Google’s own guide to optimizing for generative AI features says to ignore LLMS.txt files and other “special” markup, because doing so “will neither harm nor help your site’s visibility or rankings in Google Search, as Google Search ignores them”. That’s a statement about Google Search, so it says nothing about what ChatGPT or Perplexity do with such files, and it means anyone selling you llms.txt generation is selling something Google Search won’t read, whatever it claims about the other engines.
The instinct behind stuffing is that if the model reads the page, more keywords must help. But the model is deciding whether a passage is worth quoting, and a stuffed passage isn’t worth quoting.
Freshness fails the same way when it’s faked: in the title-tags study, pages that were actually updated won on Google, and just changing the year on a page you haven’t touched didn’t do much. A changed year over quietly stale content fakes the signal.
The last two failures are measurement mistakes. Judging the work from one answer treats a single sample as the verdict, and the answer you happened to see might be the lucky one. And a rank report can’t be the scorecard: in our refresh study, rank and citation moved together basically zero over 91 days, and 41 pages the AI cited had no Google impressions at all.
Measuring generative engine optimization
The number to hold this work to is a recommendation rate over repeated runs. Pick the buyer questions, ask each one many times, and report the share of runs that put you forward as a choice. Keep a few questions untouched as a comparison, so when the rate moves you can show the lift was your work rather than a platform-wide shift.
Omnipresence does this job. It tracks whether ChatGPT, Claude, Gemini and Perplexity recommend you or a competitor for your buyer questions, with each question asked many times per run so the rate reflects what buyers actually see.
It also shows the searches each assistant ran before answering and the pages that won them, so a drop becomes a work list: the query you lost, the page that beat you, the thing to fix. The limits, plainly: it tracks those four assistants and nothing else, you bring your own OpenRouter or OpenAI key and pay the model cost, and it measures recommendations and citations, while the money side stays yours to connect.
If you want to see which questions you’re losing and which pages win them, you can point Omnipresence at one brand for a week and read what comes back.
FAQ
Is generative engine optimization still SEO?
Google says yes, in its own AI optimization guide: “From Google Search’s perspective, optimizing for generative AI search is optimizing for the search experience, and thus still SEO.” The work overlaps almost completely, because the searches a generative engine runs read the same pages a search index serves. The scoreboard is a different matter: rank and citation have stopped moving together, so a rank report won’t tell you whether you’re being quoted, and that’s the honest reason to treat measurement as its own job.
What is generative search optimization?
Generative search optimization is the same practice under a synonym: making your pages the sources a generative engine quotes. The “generative search” phrasing points at the same surfaces, ChatGPT, Perplexity and Google’s AI Overviews, and the work doesn’t change with the name.