Generative Engine Optimization (GEO): What Is Proven

What is Generative Engine Optimization?

Generative engines are systems that answer a question by writing their own response in full sentences. Behind them are large language models (LLMs). Examples include AI chat assistants, Google Gemini and the AI features in Google Search such as AI Overviews and AI Mode. Many of these services search the web before they answer and then show links to their sources.

Generative Engine Optimization, or GEO for short, covers every measure meant to make a website appear more often in such answers: as a cited source, as a linked page or as a named brand. You will also hear it described as optimisation for generative AI.

GEO, AEO and LLMO: three terms, one topic

• GEO (Generative Engine Optimization): the term that comes from research.
• AEO (Answer Engine Optimization): refers to content that answers a question directly and briefly. People used this term before the GEO paper appeared, and today it is often treated as a synonym for GEO.
• LLMO (Large Language Model Optimization): puts the focus on the language models themselves, including what a model "knows" about a brand from its training data.

None of the three terms has a fixed definition, and providers use them in different ways. So always ask which concrete measures someone actually means.

Where the term comes from: the GEO paper

The term goes back to an academic paper. A team made up of Pranjal Aggarwal, Vishvak Murahari, Tanmay Rajpurohit, Ashwin Kalyan, Karthik Narasimhan and Ameet Deshpande submitted it to arXiv on 16 November 2023 (number 2311.09735). It was published at the KDD conference in 2024.

The researchers built their own test dataset, GEO-bench, and used it to try out how changes to a text affect its visibility in generated answers. Among the things they tested were adding source references, quotations and statistics. Their finding: GEO can raise visibility in the answers by up to 40 per cent. That figure applies to the study's experimental set-up. How strongly the methods worked varied from one subject area to another.

For context: the study looked at its own test system. It does not show that AI assistants or Google Search react in the same way. If a provider advertises "40 per cent more AI visibility", that is a transfer of the result the study does not support.

How AI search engines build their answers

Many AI services use a method called Retrieval Augmented Generation, or RAG. In simple terms, it works like this: the service takes the question, looks for matching documents in a search index and hands them to the language model as a basis. The model writes an answer from them and refers to some of the sources.

One basic consequence follows from this. An AI service can only use your page if it can find it and is allowed to read it. With Google, the sources come from the normal search index. Other providers run their own crawlers or draw on search indexes from partners. Most providers do not disclose exactly how they choose their sources.

What is different from classic SEO?

In classic search, Google shows a list of links. The aim is a good position so that people click. An AI answer, by contrast, is a finished text, and links only appear as source references. A user therefore often gets the answer without visiting a website at all. For site owners the question shifts: alongside the ranking position, it matters whether your page is named as a source and whether your brand appears in the text.

The basics stay the same, though. A page has to be technically reachable, cover its topic well and come across as trustworthy. AI services with web search rely on search indexes, and pages with solid search engine optimisation tend to do better there. What is new is mainly measurement, control over individual AI crawlers and the question of how your brand is described outside your own website.

What Google officially says about AI search

AI Overviews have been available in Germany since 25 March 2025, in German and English, for signed-in users aged 18 and over. In its documentation and in a separate guide, Google has explained what website owners need to do. The main points:

• No extra requirements: according to the page "AI features and your website", there are no special requirements for AI Overviews and AI Mode. You do not need new machine-readable files, AI text files or special markup. Special schema.org markup is not required either.
• Basic condition: a page has to be indexed and eligible to appear with a snippet in search. Google makes no promise that it will be shown.
• It is still SEO: in the guide "Optimizing your website for generative AI features on Google Search" (as of 10 July 2026), Google answers the question of whether SEO is still relevant with "In short, yes!". The AI features rest on the core ranking and quality systems. In Google's words, optimising for generative AI search is still SEO.
• What you can ignore: Google lists splitting texts into small pieces ("chunking"), unnecessary AI text files such as llms.txt, and artificially created mentions.
• No insider tools: "No third-party tool has access to our internal ranking or AI systems." So no provider can look inside Google's AI systems.
• Mass-produced pages are spam: creating large numbers of pages for possible sub-questions that an AI might ask in the background breaks the policy on "scaled content abuse".

In May 2025 Google also wrote that clicks from pages with AI Overviews are "higher quality" and that users stay longer on the website. Its advice was to offer unique content that cannot be swapped for someone else's. Structured data has to match the visible content. The article Is SEO still worth it? explains why classic SEO still counts for these reasons.

What may help with other AI services

For other AI services there are no guides as detailed as Google's. The points below follow from the way AI search works and are widely considered sensible in the industry. Their effect has not been officially proven. Apart from the crawler rules, which the providers document themselves, they are reasoned assumptions.

Access for AI crawlers

An AI service can only use pages that its crawler is allowed to fetch. You control this through the robots.txt file. The large providers distinguish between crawlers for training models and crawlers for search:

• OpenAI: OAI-SearchBot collects pages for search in ChatGPT. According to OpenAI, pages that block OAI-SearchBot do not appear in ChatGPT's search answers. GPTBot collects content that may be used to train models. ChatGPT-User fetches pages when a user asks for them in the chat, and OpenAI says robots.txt rules may not apply to these requests. Each setting works independently of the others.
• Google: Google-Extended is a token in robots.txt. With it you decide whether content may be used to train Gemini models and for grounding in Gemini apps. According to Google, Google-Extended has no effect on inclusion in Google Search and is not a ranking signal.
• Anthropic: ClaudeBot collects content that may contribute to training, Claude-SearchBot crawls to improve search results, and Claude-User fetches pages when users ask questions. Anthropic states that these bots respect robots.txt.

This means you can, for example, opt out of training and still show up in AI search. Check your robots.txt as well as the settings of your firewall or CDN. Bot protection and security rules can block AI crawlers without anyone noticing.

Clear structure and direct answers

Pages with clear headings, short paragraphs and a direct answer at the start are easy for people to read. The same probably applies to machines. Content should sit in the HTML itself and should not depend on JavaScript to load, because not every crawler runs scripts. Google explicitly advises against chopping texts up artificially.

Facts with sources

In its test system, the GEO paper found positive effects from source references, quotations and statistics. Quite apart from that, figures with evidence, dates and primary sources help your readers. Do not invent statistics, though, just because they supposedly go down well in AI answers.

Mentions outside your own website

AI services draw on many sources. If your company is mentioned in specialist articles, industry directories, forums or reviews, the chance that it appears in answers goes up. This is classic brand work and press work. According to its own guide, Google may ignore artificially created mentions.

llms.txt: a proposal with disputed value

Jeremy Howard proposed the llms.txt file on llmstxt.org on 3 September 2024, and the proposal was revised on 10 August 2026. The idea is a Markdown file in the root directory of a website that summarises its most important content in a form language models can process easily, with links to further documents.

llms.txt is not an official standard. In its guide, Google names the file as an example of unnecessary AI text files that you can ignore for Google Search. Whether other AI services read the file is not reliably documented. For technical documentation, such as for software, it can help developers who want to pass content to a language model in a targeted way. Whether it benefits a typical company website is unclear. The file does no harm as long as it is kept up to date and contains nothing beyond what the website itself says.

Checking your visibility in AI answers

Precise measurement is difficult. AI answers change depending on the wording, the user, the location and the time. You can still get a picture of where you stand.

Manual tests

• Collect ten to twenty questions your customers typically ask, for example "Who offers X in city Y?" or "How does X work?".
• Ask these questions in several AI services and in Google Search.
• Note whether your brand is named, whether your website is linked as a source and which competitors appear.
• Repeat the test at fixed intervals with the same questions.
• Also look at whether the statements about your company are correct. Wrong details often point to outdated information somewhere on the web.

What you get is a sample without statistical weight. It does show trends and gaps, though.

Google Search Console

According to Google, clicks and impressions from AI Overviews and AI Mode count as normal search traffic in Search Console. There is no separate filter for them. Rising impressions combined with falling clicks can be a hint, but you cannot reliably attribute them to AI search.

Bing Webmaster Tools

On 10 February 2026 Microsoft introduced the "AI Performance" report in Bing Webmaster Tools as a public preview. It shows how often a website's content is cited as a source in Microsoft Copilot, in Bing's AI summaries and with selected partners, which pages are cited and which search phrases the AI used to retrieve the content.

Server logs and analytics

Your server logs show which crawlers fetch your pages. In your web analytics, visits from AI services can often be recognised by the referring address. Not every service passes on this information.

First steps for your website

Before you spend money on GEO, you can check a few things yourself:

• Use Google Search Console to see whether your important pages are indexed.
• Open your robots.txt and look at which crawlers you allow or block. Make a conscious decision whether to permit training, search or both.
• Find out whether your content appears in the source code without JavaScript.
• Rework your most important pages: a clear answer at the start, figures with source and date, and current details on services, prices and contact.
• Keep your details on other platforms up to date, for example in your Google Business Profile and in industry directories.
• Start the manual tests with a fixed set of questions and repeat them regularly.

What a GEO agency offers

Many SEO agencies now offer GEO or AEO services, and some call themselves GEO agencies. Typical components are:

• Analysis of which AI answers mention the brand and which competitors are named.
• Technical checks of indexing, robots.txt and crawler access.
• Reworking content: clear answers, structure, evidence, freshness.
• Building mentions through press work, specialist articles and listings in reputable directories.
• Regular reports on samples taken from AI services.

Much of this is classic search engine optimisation under a new name. That need not be a bad thing, but you should know it before you pay extra. The article SEO for beginners shows what SEO covers in general.

Warning signs

• Promises of fixed mentions in AI assistants or AI Overviews: no provider can control what an AI service answers. Google itself says that being shown is not assured.
• Claimed access to Google's systems: according to Google, no third-party tool has access to its internal ranking or AI systems.
• Masses of AI texts for every imaginable question: this can break the spam policy on "scaled content abuse".
• llms.txt as a cure-all: Google says it does not need the file, and for other services the benefit is unclear.
• Bought mentions and fake reviews: they can damage your reputation and break platform policies.
• Research figures presented as a promise: the 40 per cent from the GEO paper applies to a test set-up. Your website is a different situation.

Ask every agency how it measures success, which of its measures are officially documented and which rest on assumptions.

Summary

Generative Engine Optimization is a new name for work on content that AI services can find, understand and cite. For Google Search, Google's own position is that good SEO is enough and that special files and tricks are unnecessary. With other AI services, open crawler access, clear content with evidence and genuine mentions probably help, though this has not been firmly proven so far. Check your visibility with a fixed set of test questions and the reports Google and Bing make available. You will find more on the subject on our topic page AEO & GEO.

Sources: Aggarwal et al., "GEO: Generative Engine Optimization", arXiv 2311.09735 (16 November 2023, KDD 2024); Google Search Central, "AI features and your website" (as of 10 December 2025); Google Search Central, "Optimizing your website for generative AI features on Google Search" (as of 10 July 2026); Google Blog on AI Overviews in Germany (25 March 2025) and on AI in Search (21 May 2025); Google Search Central, "Google's common crawlers" (Google-Extended); OpenAI, "Overview of OpenAI Crawlers" (developers.openai.com/api/docs/bots); Anthropic, "Does Anthropic crawl data from the web?" (support.claude.com); Microsoft Bing Webmaster Blog, "Introducing AI Performance in Bing Webmaster Tools Public Preview" (10 February 2026); llmstxt.org (Jeremy Howard, 3 September 2024, revised 10 August 2026).