What changes on your site, and what does not

What GEO work changes

Generative engine optimization, GEO for short, aims to get a company named or cited in answers written by AI systems such as ChatGPT, Gemini, Perplexity and Google AI Overviews. In practice the changes on a website fall into four groups. The first is access: whether the crawlers of AI systems can fetch the pages, set in robots.txt, a firewall or a CDN, and whether the text is in the HTML rather than loaded by JavaScript. The second is coverage: whether there is a page that answers the question a buyer asks. The third is content: how those pages are written, from putting the answer first to adding figures, quotes and sources. The fourth is markup: structured data such as schema.org, and files such as llms.txt. Only for the first group do the platforms themselves document the effect. OpenAI states that sites that block OAI-SearchBot do not appear in ChatGPT search answers, and Vercel and MERJ measured in late 2024 that the major AI crawlers, apart from Google's, do not execute JavaScript. Google says there are "no additional requirements" to appear in AI Overviews beyond being indexed and eligible for a snippet. For the other three groups the evidence is thinner, and some of it points the wrong way. A Dutch checklist of which improvement has which evidence is in Controlelijst GEO-verbeterpunten.

What controlled studies found negative

Most evidence on GEO content changes comes from benchmark simulations: a fixed set of sources is shown to a language model, one source is rewritten, and the researchers measure whether that source moves up or down in the answer. That is not the same as a live AI system, but it is the closest thing to a controlled test, because everything except the rewrite stays the same. Several of these studies found changes that made visibility worse.

C-SEO Bench, presented at NeurIPS 2025 by Puerto and colleagues, tested ten content optimization methods over more than 1,900 questions and 16,000 documents on four language models and six domains. The authors write that the methods are "not only largely ineffective but can actually have the opposite expected effect". Adding statistics lowered a document's rank in 19 of 24 settings. On Claude Haiku 3.5, for product recommendations, 26 of 30 method-domain combinations had a significant negative effect. Counting the significant results across all their tables, 70 of 240 combinations were negative and only a handful positive, while a classic SEO-style improvement, moving the document higher in the context the model receives, worked far better. One of the tested methods, "LLM Guidance", adds a markdown summary inspired by llms.txt to the top of the document itself; that is not the same as publishing a separate llms.txt file.

The original GEO paper by Aggarwal and colleagues, published at KDD 2024, is often cited for gains of up to about 40% from adding quotations, statistics and sources. The same paper found keyword stuffing lowered a source's share of the answer text by about 8% on its own benchmark and about 9% on Perplexity, although it scored slightly higher on the paper's second, subjective measure; the authors grouped keyword stuffing and unique words under methods that did not perform. It also found that the methods hurt pages that were already the top source: for a source ranked first, citing sources lowered visibility by 30.3%, adding quotations by 22.9% and adding statistics by 20.6%, while the same methods roughly doubled visibility for a source ranked fifth, by 98% to 115%. Another controlled study, by Wan and colleagues at ACL 2024, found that making text more relevant to the question, including by adding its keywords, made it more convincing to the model rather than less, so the keyword stuffing result is not settled.

SAGEO Arena, a February 2026 benchmark from Yonsei University with 2,700 questions and 171,003 documents, tested the full chain of retrieval, reranking and answer generation rather than generation alone. Rewriting only the body text lowered the chance of being cited for all ten strategies it tested, by 6% on average, and by 22% for an automated rewriting method. The authors write that "optimizing body text alone consistently degrades visibility across all stages". E-GEO, an MIT benchmark tested on 2,000 product queries, found that 11 of 15 hand-written rewriting prompts did worse than a simple one-line prompt; an advertising tone, foreign expressions and a one-sentence minimalist style all clearly lowered a product's rank. The authors conclude that "manually instilling 'GEO knowledge' through prompt design can backfire".

The clearest field study with a matched control group is Ahrefs' schema study from May 2026: 1,885 pages that added structured data, compared with 4,000 matched pages that did not, 30 days before and after. Citations in AI Overviews fell by 4.6% relative to the control group, a small but statistically significant change; in AI Mode and ChatGPT the changes were not distinguishable from zero. Ahrefs calls the AI Overview decline "real but unexplained". That is a reason not to expect schema to add citations, not a reason to remove it.

What Google says will hurt

Separate from the studies, Google's own policies name changes that can remove pages from its results. Its spam policies, updated in August 2026, include "attempting to manipulate generative AI responses in Google Search", and list scaled content abuse, such as using generative AI tools to generate many pages without adding value, together with keyword stuffing and hidden text. Pages that break these policies "may rank lower in results or not appear in results at all". Google's guide to optimizing for its generative AI features adds that creating separate content for every variation of a question primarily to manipulate rankings or AI responses violates the scaled content abuse policy, and lists things site owners can ignore, including llms.txt, which it says "will neither harm nor help", splitting content into chunks and rewriting content for AI.

What this means for changes on your site

The evidence gives a clear order of work. Start with access, because that is the only group where the platforms document the effect: check that AI crawlers can fetch your pages and that the text is in the HTML. Then look at coverage: whether there is a page that answers each question your buyers ask, which a site audit alone does not show. Be careful with content rewrites done for AI, because in the controlled studies they helped less often than they hurt, and in the original GEO paper they hurt most on sources that were already ranked first. Do not expect structured data or an llms.txt file to add citations. And measure before and after each change, per question and per AI system, so that a change that lowers visibility is caught rather than repeated across the site.

Where we stand

AIGeschikt changes a client's existing page only where the measurement shows it misses a buyer question, and measures the same questions before and after. We do not rewrite pages that are already cited, and we do not add pages per question variant. Our Dutch checklist of improvements with their evidence is Controlelijst GEO-verbeterpunten; what the most common audit points do, in Dutch, is in 2 van de 20 punten in een GEO-audit hebben een aangetoond effect.