GEO: What Generative Engine Optimization Really Does

A domain, a landing page, and thirty days

On July 2, a marketer bought a domain for $11.25. Morrowen.com became a landing page with formulation notes and a comparison page about magnesium versus baking soda for sensitive skin. A single launch essay went up on a Substack under the Morrowen name, and the whole footprint was built in under an hour before the builder went on vacation for thirty days. No reviews, no paid coverage, no outreach — she deliberately skipped fake testimonials and Reddit spam.

The result was narrower than the headline version of this story usually suggests. The first hit came on day 21, and by day 30 ChatGPT with browsing enabled named Morrowen in four out of four answers to two specific questions, listed first each time. On thirteen broader queries — best, cheapest, aluminum-free, what people actually recommend — nothing. Claude and Gemini never mentioned the brand once, on any setting. And the incumbent category leader wasn't dethroned: it was still named 99 times out of 180, while the fake brand picked up what its creator called a dusty spot at the back of the store.

That's the real shape of the experiment being cited across marketing circles as proof that GEO can manufacture instant authority. It can — but only inside a narrow, unbranded, low-competition slice of queries. The gap between that nuance and the pitch being sold around it is the actual story.

What changed, mechanically

Search used to return a ranked list; a generative engine returns one answer, assembled from sources it chooses to trust. The academic term for optimizing toward that outcome comes from a 2023 paper by Pranjal Aggarwal and coauthors, later published at KDD 2024, which introduced GEO-bench, a large benchmark of queries and sources, and demonstrated that GEO can boost visibility by up to 40% in generative engine responses. The gains weren't uniform: classical SEO signals like keyword density showed minimal influence on citation probability, while statistics, source citations, and quotations from credible authorities were what moved the needle.

An independent field test from SE Ranking, run across a fictional brand, put numbers on how narrow that opening really is. The team tracked 825 prompts across query types, generating 15,835 AI answers in the first month. On queries only the fake brand could realistically answer, it beat established competitors by as much as 32x and reached near-exclusive visibility inside thirty days. But 96% of all AI visibility came from branded searches — a new domain had little chance competing for broader, non-branded topics even in a low-competition niche. Non-branded informational queries produced just 4% of AI answers, and even those mostly ran through supporting test domains. That 96/4 split is close to the exact pattern Morrowen's builder found by hand.

The verification gap

The mechanism that makes GEO work is the same one that makes it risky. Generative engines are optimized to sound complete, not to be correct, and a widely cited Ahrefs experiment that seeded a fictional brand with detailed, false product claims found the model rewarded confident detail over accuracy — the researcher behind it concluded "in AI search, the most detailed story wins, even if it's false."

Independent accuracy research backs the general pattern, if not the precise figures circulating informally. A Tow Center for Digital Journalism study that tested eight AI search engines on citation accuracy found AI search engines were inaccurate 60% of the time, with wide variance by platform — Grok had the highest error rate at 94%, while Perplexity had the lowest at 37%. Separately, Ahrefs' own analysis of cited URLs found ChatGPT's hallucination rate for links was nearly 3x higher than its competitors. None of this means GEO doesn't work. It means the layer deciding which brand gets named is demonstrably capable of naming the wrong one with full confidence — which is precisely the condition under which a three-page site with no product can out-cite an established competitor on a narrow question, and the condition regulators are now moving to address.

Disclosure is arriving before measurement is settled

The EU AI Act's transparency provisions become enforceable this year. Starting August 2, 2026, deployers of AI systems that generate or manipulate content meeting the Act's deepfake definition must disclose the content's artificial nature. The same obligation extends to AI-generated or AI-manipulated text published to inform the public on matters of public interest, unless meaningful human editorial review has taken place. Guidance published ahead of the deadline clarifies that the rule is narrower than a blanket AI-content label — the operative test is whether the AI element would affect a consumer's understanding or decision-making, in which case it counts as material information that should be disclosed. For GEO specifically, the open question is whether a brand quietly engineering its own citation into an AI answer — without disclosing that engineering — eventually gets treated the same way as an undisclosed ad. Nothing in the current text says so explicitly. That is exactly the kind of gap a reader should watch, not assume is already closed.

A market with no verified size, and a acquisition that says something anyway

There is no credible, sourced figure for how large the GEO market is today; treat any specific dollar total circulating online as unverified. What is verifiable is that money is moving inside the adjacent categories — SEO tooling, PR measurement, brand monitoring — toward AI-answer visibility as a line item. Ahrefs added Brand Radar, describing it as the largest AI visibility database, tracking hundreds of millions of monthly prompts across AI Overviews, AI Mode, ChatGPT, Copilot, Gemini, and Perplexity. SE Ranking built a parallel product, SE Visible, that collects real AI responses through a browser interface rather than simulated or API-based results, refreshed daily.

The clearest signal of category-level conviction is a transaction, not a survey. Scrunch AI, a GEO-specific monitoring startup founded in 2023, raised a $15 million Series A in July 2025 led by Decibel, with participation from Mayfield and Homebrew, and was reporting an average 40% increase in referral traffic and up to 4x improvement in visibility across generative platforms among its customers. Less than a year later, Sitecore, an enterprise content and digital-experience platform, acquired Scrunch in a deal valued at around $225 million, per Bloomberg. Acquisitions at that size are a stronger signal of category conviction than any unverified market-size figure — but they say a buyer believes in the category, not that any given GEO tactic reliably works.

On the traffic side, the most rigorous available dataset undercuts some of the more dramatic numbers being quoted informally. First Page Sage's own longitudinal panel — session-level data from 218 client websites across 12 industries, covering roughly 3.4 billion sessions between January 2023 and July 2026 — found that AI platforms now account for 6.24% of all sessions, up from 0.11% in the first quarter of 2023. That's real, fast growth. It is also a different order of magnitude from the double-digit AI-traffic-share claims sometimes attributed to the same firm without a clear source — a reminder to check the underlying dataset before repeating any single-brand statistic.

What to check before the next pitch

The operating discipline here is simple to state and hard to hold to: separate what a controlled experiment demonstrated from what a case study implies, and treat anything without a traceable dataset as a claim, not a fact. Concretely, a reader evaluating a GEO vendor or agency pitch should ask for the query set behind any visibility number — branded versus non-branded, as the SE Ranking data shows these behave completely differently. Ask which AI platforms were tested, since Grok, Gemini, Perplexity, and ChatGPT show materially different citation behavior and error rates. Ask whether the disclosure question has been addressed at all, given that EU enforcement begins in August. And watch whether category-level acquisitions like Scrunch's continue, since consolidation activity is currently a more reliable signal than any circulating market-size figure.