What is generative engine optimization (GEO)?

Most marketing terms are coined by vendors. GEO is unusual: it was coined by researchers. In late 2023, a team from Princeton and Georgia Tech posted a paper titled "GEO: Generative Engine Optimization" (arXiv:2311.09735), later published at KDD 2024. The authors asked a simple question with an uncomfortable answer for the SEO industry: if AI engines now write the answer instead of listing links, what makes a page more likely to appear inside that answer?

They tested nine optimization methods across 10,000 queries and measured how each changed a source's presence in the generated response. The headline result: the right content changes can boost visibility in generative engine responses by up to roughly 40 percent. The best-performing techniques were adding quotations, adding statistics, and citing sources. Measured precisely, the top methods improved about 22 percent on Position-Adjusted Word Count, a metric that weights how much of the answer your source occupies and how early it appears, and about 37 percent on Subjective Impression.

Just as important is what failed. Keyword stuffing, the crudest classic SEO tactic, performed near the bottom for generative engines, below doing nothing at all. The paper also found that efficacy varies by domain: what lifts a page about historical facts differs from what lifts a page about software, which is why the authors released GEO-bench, a benchmark for testing tactics by query category. We unpack the full study in our plain-language walkthrough of the Princeton GEO paper, including which findings still hold in 2026.

A discipline only matters if the surface it optimizes has an audience. This one does. ChatGPT reached roughly 900 million weekly active users in early 2026, about double its count a year earlier. Google's Gemini app passed 750 million monthly users. Google's AI Overviews now trigger on roughly 48 percent of tracked queries by early 2026 according to third-party tracking, though the measurement spread is real: some panels report figures closer to 13 to 25 percent depending on the query set. Whichever panel you trust, the direction is the same and the slope is steep.

Behavior followed the surfaces. McKinsey reported in October 2025 that half of consumers already use AI-powered search intentionally as a primary way to find information and make buying decisions. In B2B the shift is sharper still: G2 found that 51 percent of B2B software buyers now start research with an AI chatbot more often than with Google, up from 36 percent just seven months earlier. When the first question in a buying journey is typed into a chat box, the brands named in the reply get considered and the rest do not. That downstream effect is the subject of our primer on AI visibility.

GEO is the supply-side response to that demand shift. If a synthesized answer is the new front page of your category, GEO is the work of earning a place in it: on your own pages, on the third-party pages the engines trust, and in the indexes the engines read.

You cannot optimize a system you misunderstand, so start with the mechanics. Every generative engine draws on two pools of knowledge. The first is training data: what the model absorbed about your brand before its knowledge cutoff. The second is live retrieval: pages fetched from a search index at answer time and summarized into the response, usually with citations. The distinction matters enormously for strategy, because training data changes on retraining timelines you cannot influence or schedule, while retrieval-grounded answers can change within weeks of publishing the right page. We cover the split in depth in training data vs live retrieval.

The retrieval plumbing is expanding fast. ChatGPT's search experience launched on Bing's index, but OpenAI now operates its own crawler, OAI-SearchBot, and according to a Botify analysis has roughly tripled its web crawl since August 2025. Perplexity maintains its own index, reported at more than 50 billion pages, and averages about 8.2 sources per answer, roughly 3.4 times what ChatGPT cites. Google's AI features cite pages from Google's organic index, which means being indexed and rankable remains a precondition there.

One more mechanical fact reshapes strategy: the engines do not share a reading list. Cross-platform citation analyses find that only about 11 percent of domains are cited by both ChatGPT and Perplexity. A single content strategy cannot win every AI surface, which is why serious GEO work tracks each engine separately instead of assuming one score covers them all.

Two rows deserve honest elaboration. Schema markup is genuinely contested. SE Ranking found that about 71 percent of pages cited by ChatGPT include structured data, and about 65 percent for Google AI Mode, but that is correlation, and well-built pages tend to carry schema anyway. Ahrefs ran the cleaner experiment in May 2026 across 1,885 pages: adding JSON-LD schema produced no measurable citation lift on already-cited pages for ChatGPT and AI Mode, and a statistically significant decline in AI Overviews citations. The caveat cuts both ways: those pages were already heavily cited, so schema may still help initial parsing and discovery. Bing's Fabrice Canel has said schema helps LLMs understand content for Copilot. Our read: schema is cheap, keep it accurate, and expect discovery help rather than a citation multiplier.

The llms.txt row is simpler. It is an emerging, unratified proposal for a file that summarizes your site for AI crawlers, and no engine has confirmed using it. It costs ten minutes to add and there is no published evidence it changes anything yet. Treat vendors who sell llms.txt as a headline GEO service with suspicion.

The least intuitive part of GEO for SEO-trained teams: much of the battlefield sits outside your domain. 5W Research found that Wikipedia accounts for 13.15 percent of ChatGPT citations in the U.S. and Reddit for 11.97 percent, which means two sites you do not control drive over a quarter of the citations on the most-used engine. Meanwhile the Wall Street Journal, New York Times and Bloomberg do not appear in ChatGPT's top 20 cited domains. Authority as journalists define it and authority as generative engines practice it have quietly diverged.

Perplexity leans even harder on community content. Reddit is its single largest source, with estimates ranging from roughly 17 to 24 percent of citations, and one analysis put Reddit at 46.7 percent of Perplexity's top-10 share. It also skews toward LinkedIn, NIH and G2. Across engines, Reddit's citation share in commercial categories grew about 73 percent through 2025 and 2026. The practical consequence: earning an honest presence in the threads, reviews and comparison pages the engines already trust often moves an answer faster than anything you publish on your own site.

This is why a complete GEO program has two tracks. Track one: publish pages that answer buying questions directly, with the statistics, quotations and citations the Princeton research validated. Track two: get named on the third-party sources each engine cites for your questions. Good tooling shows you exactly which sources an engine cited for each question you lose, which turns track two from guesswork into a pitch list.

GEO has a measurement problem that classic SEO never had: the surface is probabilistic. SparkToro measured under a 1 percent chance that two identical ChatGPT runs return the same brand list. Ask the same buying question twice and you can get different brands, different orderings, different citations. Any score built on a single run is false precision, and any vendor showing you one number without the raw answers behind it is asking for faith rather than offering evidence.

Honest measurement has four properties. Repeated runs on a schedule, so volatility averages into a trend line. Stored raw answers, so every number can be audited by reading what the engine actually said. Literal matching, so a mention only counts when your brand name appears in the answer text. And branded-question exclusion: if the question already contains your brand name, the answer mentions you by construction, and counting it inflates the score. This is the methodology Reachroller runs, documented in full on our methodology page, and it is the standard we would hold any tool to, including ours.

If you want to run the measurement by hand first, our guide on measuring AI visibility without lying to yourself walks through the full protocol, including how many runs you need before a trend means anything.

The label stuck commercially as well as academically. Intel Market Research sizes GEO services at roughly $1.01 billion in 2025, projected to reach about $1.48 billion in 2026, a growth rate around 45 percent. Disclosed venture funding across the tooling category exceeds $200 million. Job listings mentioning generative engine optimization, AI search or LLM visibility grew several hundred percent year over year through 2025 and 2026. Whether or not the acronym survives, the budget line it describes has arrived.

You will also meet sibling acronyms: AEO for answer engine optimization, and LLMO or AIO in some agencies' decks. The terms overlap far more than their promoters admit, and the differences that do exist are worth ten minutes, so we wrote them up separately in our guide to the acronym soup. The short version: the work underneath is nearly identical, so pick the work, and let other people argue about the label.

Day one: build the question list.Write down 20 to 25 questions a buyer would ask an assistant before choosing a product like yours. Real phrasing, unbranded: "best tools for X", "how do I solve Y", "X vs Y for a small team". Ask them in ChatGPT and Perplexity, and record which answers name you, which name rivals, and which sources get cited. That is your baseline, and it is usually a humbling document.

Week one: publish the first fixes.Take the three questions that matter most commercially and publish a page for each that answers the question in the first paragraph, then earns the reader's trust with statistics, quotations and cited sources, per the Princeton findings. Submit each page to Google Search Console and Bing Webmaster Tools, since AI engines read search indexes and an unindexed page does not exist for most of them. The full page anatomy is in how to write content AI engines actually cite.

Week two onward: recheck and repeat. Give the indexes one to two weeks, then re-ask the questions and compare against the stored baseline. This loop is exactly what Reachroller automates: it runs your question set on schedule through official engine APIs, scores only literal mentions with branded questions excluded, generates the publish-ready fix page for every question you lose, complete with slug, title tag, meta description, schema markup and indexing steps, and rechecks whether the answer flipped. Starter is $29 per month, and the three-day trial includes 50 credits and every feature with no card, which covers a full first report and one generated fix on your own brand.