Generative Engine Optimization

Generative Engine Optimization (GEO) is the discipline of making content, brands and products eligible to be retrieved, cited and recommended by AI systems built on large language models, such as Google's AI Overviews and AI Mode, ChatGPT search, Perplexity, Claude and Gemini, rather than only ranked by a traditional search engine. The evidence published during 2026 changed what the discipline consists of. It is less a set of on-page writing techniques than a matter of retrieval eligibility, third-party presence, upkeep and honest measurement.

Retrieval Eligibility and Query Fan-Out

A page cannot be cited unless it is retrieved, and AI engines rarely retrieve for the query as typed. Through query fan-out, the engine splits a question into several related sub-queries and cites the pages that surface most often across them. This is the leading explanation for a sharp change in Google's behavior: Ahrefs (March 2026; 863,000 keyword results pages and 4 million cited URLs) found that 37.9% of URLs cited in AI Overviews rank in the organic top 10 for the original query, 31.2% rank between 11 and 100, and 31.0% rank beyond 100. The same measurement was roughly 76% in July 2025. BrightEdge (February 2026), using a different method, puts the top-10 overlap at about 17%. The two numbers disagree on magnitude and agree on direction: ranking for the head term is neither sufficient nor strictly necessary, and coverage of likely sub-questions matters more than it did a year ago.

Engines Diverge, and Answers Vary

There is no single AI search to optimize for, even inside one company. A SIGIR 2026 study of 11,500 real user queries (Grossman et al.) compared Google's organic results, the accompanying AI Overview and Gemini for each query and found the retrieved sources "substantially different," with average Jaccard similarity below 0.2; the generative surfaces were more likely to retrieve Google-owned content. Visibility on one surface is weak evidence of visibility on another.

The same engine also gives different answers on different days. Schulte, Bleeker and Kaufmann (April 2026) measured day-to-day similarity of cited sources at 0.34 to 0.42, meaning roughly 65% of sources change daily, while the set of brands named is more stable at 0.45 to 0.59. Their conclusion is that visibility should be characterized "as a distribution rather than a single-point outcome." Repeated sampling across prompts, engines and days, reported as share of voice with a stated margin, is the working method of AI visibility measurement; tracking visibility per engine is the province of AI-search visibility tools like LLM Optimizer.

Selection Is Not Absorption

Being listed as a source and shaping the answer are separate outcomes. Zhang, He and Yao (April 2026; 21,143 citations) found that ChatGPT cites an average of 6.88 pages per prompt, Google 12.06 and Perplexity 16.35, while the mean influence of each cited page on the answer text was 0.27 for ChatGPT against 0.058 for Google and 0.065 for Perplexity. Engines that cite more sources draw less from each one. In the same study, question-and-answer formatting alone did not raise influence, and encyclopedic pages were absorbed far more than news pages (0.21 against 0.07). A GEO program therefore tracks two things: whether a page earns AI citations, and whether the brand's facts and framing appear in the answer itself.

Earned Media, Mentions and Freshness

AI engines lean heavily on third parties. University of Toronto researchers (2025) found that AI search cites earned media 72–92% of the time against 18–27% for brand-owned content. A June 2026 study of about 168,000 citations across 128 brands and 12 markets (Żatuchin) reached a similar split: 85.7% of citations point to sites the brand does not own, and roughly 80% of citations come from about 18% of domains. Ahrefs' analysis of 75,000 brands (December 2025) found that mentions of a brand in YouTube titles, transcripts and descriptions correlate with AI visibility at around 0.74, and branded web mentions at 0.66 to 0.71, while backlinks correlate at only about 0.25 to 0.28. These are correlations, not causes. Well-known brands are mentioned more everywhere, and the same prominence, including training data frequency, may drive both the mentions and the citations.

Freshness is real but works through maintenance rather than volume. Seer Interactive (July 2026; 7,683 pages) found that 75% of the pages LLMs cite had been updated within the last year, and summarized the pattern this way: "The freshness LLMs reward is being manufactured by updates, not by new publishing."

What the Evidence Reset Debunked

The original GEO paper from Princeton and collaborators (2023, published at KDD 2024) reported that adding quotations, statistics, source citations and fluency edits raised a page's visibility in a simulated generative engine, and those effect sizes became the standard GEO playbook. They are best read now as a superseded baseline. A September 2026 re-measurement on ten modern engine families (Bajemon and Rochet, vendor-affiliated authors) found that the same levers "move citation on none," and that a query-agnostic content score has a within-query correlation of only 0.11 with citation order, which the authors describe as "a quality filter rather than a citation predictor." A July 2026 survey of 45 GEO studies (Martinez) concluded that "no reviewed technique shows a stable, longitudinal, cross-platform causal effect on organic discoverability." Schema markup fared no better as a citation lever: Ahrefs' controlled study (May 2026; 1,885 pages adding JSON-LD against 4,000 matched controls) found "no major uplift in citations on any platform." Clear, well-evidenced writing still serves readers; it is not a dial that reliably turns citations up.

Crawler Access in Practice

Eligibility starts with access. AI crawlers now come in distinct types: OpenAI documents GPTBot for training and OAI-SearchBot for its search index, both of which honor robots.txt, and ChatGPT-User for fetches a user initiates. Blocking training is a different decision from blocking search, and the SIGIR study above found that websites blocking Google's AI crawler are significantly less likely to be retrieved by AI Overviews. Publishing these controls is increasingly a platform feature rather than a manual chore — LightCMS, the CMS serving this site, exposes separate switches for training, AI search and user-fetch crawlers. Curated files such as llms.txt belong to the same access layer, though they are read mostly by agents rather than by search engines.

The AI search landscape is still forming, and the durable principles are modest: be retrievable, be present where engines already look, keep pages current, and measure repeatedly. As AI agents begin to act as well as recommend, the stakes of GEO extend beyond visibility into direct commercial impact.