# What is GEO? Generative Engine Optimization explained > GEO is the work of becoming visible, accurately described and citable in the answers from ChatGPT, Claude and Gemini. Here is the definition, the difference from SEO, how models pick their sources, how AI visibility is measured, what llms.txt is — and seven concrete actions. With numbers from our own dataset of 253 Norwegian prompts. - Published: August 2026 (2026-08-01) - Author: m51.ai Lab - Read time: 7 min read - Series: GEO Series - Canonical HTML edition: https://m51.ai/en/lab/hva-er-geo - Norsk utgave (Markdown): https://m51.ai/lab/hva-er-geo.md - All Lab articles (Markdown index): https://m51.ai/en/lab/index.md --- GEO (Generative Engine Optimization) is the work of making a business visible, accurately described and citable in the answers produced by generative AI models such as ChatGPT, Claude, Gemini and Perplexity. Where classic SEO is about ranking high in a list of links, GEO is about becoming the source the model builds the answer itself on — and about being named when the answer says who is good at something. That is the whole definition. The rest of this article covers why it became its own discipline, how models actually pick their sources, how AI visibility is measured, what llms.txt is, and what a business can concretely do. The numbers along the way come from our own dataset NorGEO-Bench, where we asked ChatGPT, Claude and Gemini 253 Norwegian questions about Norwegian business. --- ## GEO vs. SEO: what is the actual difference? GEO does not replace SEO. The two disciplines share most of the same groundwork — technically sound pages, clear content, external coverage — but they optimise for two different outcomes. | Question | SEO | GEO | | --- | --- | --- | | What is the result? | A ranked list of links | One synthesised answer, with or without sources | | What are you competing for? | Position in the result list | Being named and cited inside the answer | | Who reads the content? | A crawler that indexes | A model that summarises and rephrases | | What decides the outcome? | Relevance, authority, links | Whether an unambiguous, verifiable source exists | | How is it measured? | Position, impressions, clicks | Mention rate, citation share, factual accuracy | | How fast does it change? | Rankings shift over weeks | The answer can change from question to question | **The most important practical difference: in a search result the user can weigh ten alternatives themselves. In an AI answer the user usually gets three to five names — and everything else does not exist. GEO is therefore a contest for a far narrower list.** ## How AI models pick their sources A modern AI model answers on one of two grounds: what it remembers from training, or what it retrieves through web search at that moment. The distinction matters, because only one of them can be influenced quickly. When we tested the three leading models with no tools and no internet access, the best one scored 61 % correct on questions about real Norwegian companies. The models simply know too little about Norwegian business to answer confidently from memory — and they know that they know too little. That is why they search. **The consequence is simple: whatever is available about you online at the moment someone asks is, in practice, the answer the user receives.** ### The sources the models actually lean on NorGEO-Bench logs which domains the platforms themselves report as sources — in the latest measurement (m-full-2026-07-30), 2,735 unique domains across 1,707 answers. Two traits stand out: the sources are Norwegian, and they are barely shared. | Domain | Citations | Type of source | | --- | --- | --- | | finansavisen.no | 123 | Editorial media | | finn.no | 114 | Marketplace | | proff.no | 107 | Company registry data | | smartbyra.no | 79 | Comparison service | | bytt.no | 74 | Comparison service | | lovdata.no | 74 | Government source | 74.7 % of citations go to Norwegian domains, and the three platforms share only 8.3 % of the source base: ChatGPT leans on government sources, Claude on registries and marketplaces, Gemini on editorial media and comparison services. The practical instruction: a correct entry on Proff and Finn, presence on the comparison services in your industry and Norwegian-language content is GEO work — not merely brand hygiene. → [The full source analysis: Which sources do ChatGPT, Claude and Gemini cite? (in Norwegian)](https://m51.ai/lab/hvilke-kilder-siterer-ai) ## How AI visibility is measured There is no ranking list to read off. AI visibility therefore has to be measured by querying the models systematically and counting what happens. In NorGEO-Bench we put 253 prompts to each platform and run every question several times — 1,707 answers in the latest measurement — and have a judge panel from three different vendors score every answer against a shared evidence base. Four measures cover most of the picture: - Mention rate: the share of relevant questions where your business is mentioned at all — measured across repeated runs, not a single pull. - Citation share: how often your own domain is given as a source, versus registries and third parties. - Accuracy (0–3): whether what the model says about you is actually true. - Hallucination per named entity: how much the model invents, divided by how many businesses it names — a long answer should not be punished for saying more. The last two surprise most people. A model that does not know you will not necessarily stay silent — it guesses. | Platform | Accuracy (0–3) | Hallucination per entity | Mentions expected brands | | --- | --- | --- | --- | | ChatGPT (GPT-5.6) | 2.40 | 0.056 | 81.2 % | | Claude (Opus 5) | 2.18 | 0.047 | 96.6 % | | Gemini (3.1 Pro) | 1.86 | 0.084 | 94.6 % | The accuracy ranking is statistically significant and holds under all eight judge configurations we tested. On hallucination, ChatGPT and Claude are indistinguishable — Gemini is worst on both measures. And the models behave differently enough that «are we visible in AI?» is not one question, but three. → [The full method, robustness tests and what we got wrong last round: the main report (in Norwegian)](https://m51.ai/lab/hvilken-ai-er-best-pa-norsk) And visibility is not a locked contest over five seats: the five most mentioned players in each industry take only 6–16 % of all mentions. No Norwegian industry is owned by a few names in the AI answers yet — an opening for whoever makes themselves citable first. ## What is llms.txt? llms.txt is a file placed in the root of your site — yourdomain.com/llms.txt — that describes in plain Markdown what the business is and where the most important content lives. The idea mirrors robots.txt, but the audience is language models rather than crawlers: a short, unambiguous source an AI agent can read without parsing its way through a designed website. ```markdown # Company name > One sentence on what you do, for whom, and where. ## Services - [Service A](https://yourdomain.com/service-a): Short description - [Service B](https://yourdomain.com/service-b): Short description ## About - [About the company](https://yourdomain.com/about): Facts, history, registration number ``` The honest assessment: none of the major model providers have confirmed that they use llms.txt, and an analysis of 300,000 domains found no clear correlation between having the file and being cited. But it takes fifteen minutes to write, and any AI agent that does visit your site will find it. We recommend it as a cheap supplement — not as the measure itself. ## Seven concrete GEO actions - Ask the models about yourself. Put the questions your customers actually ask to ChatGPT, Claude and Gemini, and write down what is wrong. Without a baseline you cannot see progress later. - Clean up the registries first. Proff and Finn are among the most cited sources across all three platforms — and they are free to correct. Errors there become errors in the answers. - Get onto the comparison services in your industry. Smartbyra, Bytt and similar niche sites are cited as authorities on par with national media — and they are far easier to influence. - Write one unambiguous about page. What you do, for whom, where, and since when. Models reproduce what is easy to extract, not what is most beautifully phrased. - Publish numbers and definitions others can cite. Your own measurements, methodology, price levels. A model prefers to cite a source that says something precise. - Get covered by others. Trade press, industry directories and reference works are both training data and search results. - Measure again every month — and measure repeatedly. Identical questions share only about a third of the companies mentioned from run to run; a single query is an anecdote, frequency is a measurement. ## Three common misconceptions - «GEO replaces SEO.» No. The models retrieve from web search, and web search is still ranked. Weak SEO produces weak GEO. - «Schema markup is enough.» Structured data helps the model understand you, but it cites sources it finds credible — not merely machine-readable. - «If the model does not know us, it says nothing.» This is the most expensive misconception. In our tests the most hallucination-prone model constructed an entirely invented history for a name it did not recognise. **AI visibility is not a one-off project. The models change, and so does the picture. That is one of the jobs the AI agents in m51.ai do: track how the brand is described, and propose action when something shifts.** [See how it works](https://m51.ai/#kontakt) --- → [Explore NorGEO-Bench: the interactive benchmark page with company lookup](https://m51.ai/en/lab/norgeo-bench) → [Read part 1 of the GEO series: How well does AI know the Norwegian market?](https://m51.ai/en/lab/geo-llm-norsk-marked) → [Read part 2: Which AI is best at Norwegian? The main NorGEO-Bench report (in Norwegian)](https://m51.ai/lab/hvilken-ai-er-best-pa-norsk) → [Read part 3: Which sources do ChatGPT, Claude and Gemini cite? (in Norwegian)](https://m51.ai/lab/hvilke-kilder-siterer-ai) → [Read part 4: AI visibility is a lottery draw — why one pull is not a measurement (in Norwegian)](https://m51.ai/lab/ai-synlighet-er-en-trekning) → [Related: Can Norwegian-trained AI models compete with GPT and Gemini?](https://m51.ai/en/lab/open-source-markedstest) ## Frequently asked questions ### What does GEO mean? GEO stands for Generative Engine Optimization. It is the work of making a business visible, accurately described and citable in the answers of generative AI models such as ChatGPT, Claude, Gemini and Perplexity — becoming the source the model builds its answer on, rather than just a link in a result list. ### What is the difference between GEO and SEO? SEO optimises for ranking high in a list of links that the user evaluates themselves. GEO optimises for being named and cited inside one synthesised AI answer, where the user typically gets three to five names and never sees the rest. The disciplines share most of the groundwork — technically sound pages, clear content and external coverage — but are measured differently: SEO on position and clicks, GEO on mention rate, citation share and whether what the model says is actually true. ### How do I measure whether my company is visible in AI answers? By querying the models systematically and counting. Define the questions your customers actually ask, run them several times against each platform, and measure four things: how often you are mentioned, how often your own domain is cited, how accurate the model's claims are, and how much it invents per named business. In NorGEO-Bench we put 253 prompts to each of the three platforms with repeated runs — 1,707 answers in the latest measurement — and have a judge panel from three vendors score every answer against a shared evidence base. The repetition is not cosmetic: identical questions share only about a third of the companies mentioned from run to run. ### Which sources do AI models cite most about Norwegian companies? In the latest NorGEO-Bench measurement the top sources were finansavisen.no (123 citations), finn.no (114), proff.no (107), smartbyra.no (79), bytt.no (74) and lovdata.no (74) — taken from the platforms' own citation fields. 74.7 % of citations went to Norwegian domains, and the three platforms shared only 8.3 % of the source base: ChatGPT leans on government sources, Claude on registries and marketplaces, Gemini on editorial media and comparison services. ### What is llms.txt, and does it work? llms.txt is a Markdown file in the root of your site (yourdomain.com/llms.txt) describing what the business is and where the most important content lives — the robots.txt idea, aimed at language models. None of the major model providers have confirmed that they use it, and an analysis of 300,000 domains found no clear correlation with being cited. It still takes fifteen minutes to produce, and an AI agent visiting your site will find it. We recommend it as a cheap supplement, not as the primary measure. ### Can AI models invent things about my company? Yes, and it is the most underestimated risk. A model that does not know you will not necessarily stay silent — it guesses. In the latest measurement we scored hallucination per named business: ChatGPT and Claude are on par (0.056 and 0.047, a difference that is not statistically significant), while Gemini is worst (0.084). And the errors vary sharply by industry — real estate is almost three times as error-prone as e-commerce, likely because less public, structured information exists to build on. --- *This is a machine-readable Markdown edition of the HTML article at https://m51.ai/en/lab/hva-er-geo. The content is generated from the same source as the website.*