What is GEO? Generative engine optimization explained for business owners

GEO stands for generative engine optimization. A generative engine is a system that writes an answer to your question, where a classic search engine shows a list of links. ChatGPT, Gemini, Claude, Perplexity, Microsoft Copilot and Google's AI Overviews are all generative engines.

GEO is everything you do so that those written answers include your business. That means three things: the assistant can find you, it describes you correctly, and it names you when a customer asks for a recommendation.

You will also see the phrase GEO optimization. The O in GEO already stands for optimization, so the phrase says it twice. People search for it anyway, and it means the same work.

Picture a customer who needs a kitchen fitter. Five years ago she typed two words into Google and scrolled through ten links. Today she asks an assistant a full question: which kitchen company near me is good with small spaces and fair on price?

The assistant replies with a short text and three or four company names. Each name comes with a reason, such as strong reviews or a specialty in compact kitchens. Under the answer you see the sources it read.

So what is generative engine optimization for this kitchen fitter? It is the work that puts his name in that short list, with the right reason next to it and a link to his site underneath.

The name comes from an academic paper titled GEO: Generative Engine Optimization. Six researchers published the first version on the research archive arXiv on 16 November 2023. The paper was accepted at KDD 2024, a leading conference on data science.

The authors are Pranjal Aggarwal, Vishvak Murahari, Tanmay Rajpurohit, Ashwin Kalyan, Karthik Narasimhan and Ameet Deshpande. The published version names Princeton University and IIT Delhi as their institutions. That is why people call it the Princeton paper.

Their starting point was simple. Generative engines decide which websites appear in an answer, and website owners have little control over that. The researchers wanted to test which changes to a page raise its visibility inside an AI answer.

The team built a test set called GEO-bench with 10,000 search questions from many fields. They then rewrote web pages in nine different ways and measured how often and how prominently each version showed up in generated answers.

1. What worked bestAdding statistics, adding quotations from relevant people and citing sources. The paper reports visibility gains of up to 40 percent for these methods.
2. What worked poorlyKeyword stuffing, the old habit of repeating a search term many times. In the test it gave little to zero benefit.
3. What variedThe best method differed per subject. A change that helped a history page did less for a legal page.

Read these numbers with care. They come from a research setup built in 2023, and today's assistants are different products. The direction still holds up in practice: concrete facts with a named source get quoted more often than vague claims.

It helps to split GEO into three levels. Each level has its own test, and each one asks for different work.

Our own research shows how far apart these levels are. Claude described 88 of 100 businesses correctly when asked by name. It recommended 34 of those 100 when the question was about their trade and town.

Most businesses already pass level one. The gains in GEO sit at levels two and three.

An assistant runs on a large language model, or LLM. That is software trained on a huge amount of text to predict which words come next. GPT, Gemini and Claude are examples of such models.

The model has two ways to know something. The first is its training: everything it read up to a fixed date. That knowledge is frozen and changes only when the company releases a new model.

The second is live search. The assistant searches the web at the moment you ask, reads a set of pages and writes its answer from them. The technical name is retrieval, and grounding means tying the answer to those retrieved pages.

For questions about local companies, prices, opening hours and current products, assistants lean on live search. That is good news for a small business. A page you publish this month can appear in an answer soon after search engines index it.

A customer asks one question. Behind the scenes the assistant often runs several searches. Google's Search Central documentation calls this query fan-out: AI Overviews and AI Mode may issue multiple related searches across subtopics and data sources to develop a response.

Take the kitchen question from earlier. The system may search for kitchen companies in the town, for reviews of the top names, for compact kitchen specialists and for typical prices. It then combines what it found into one answer.

The lesson for you: a single keyword is only one of the doors. A page that covers the related questions around your service, with clear facts for each, has more ways to be picked up.

After the search, the assistant selects passages from the pages it read and writes the answer in its own words. Most assistants then show where the information came from, as numbered references, link chips or a sources list.

Those citations are your route to a visitor. A customer who wants to check the answer clicks through to the source. Google's documentation says clicks from results pages with AI Overviews are of higher quality, meaning visitors tend to spend more time on the site.

There are two ways to appear. The assistant can cite your page as a source, and it can name your business in the text. The strongest result is both: your name in the answer and your site among the links.

Each assistant has its own way of searching. This table is based on the documentation each company publishes for website owners, checked in October 2026.

The pattern is clear. Every assistant stands on a search index. If search engines can read and index your site, the assistants can use it.

SEO, or search engine optimization, aims for a high position in a list of links. GEO aims for a place inside a written answer. The two share most of their foundations and differ in what counts as a win.

People sometimes write GEO SEO as one phrase. That fits the reality well. GEO is a layer on top of SEO, and a site with weak SEO has little for an assistant to read.

The field has many abbreviations, and speakers at conferences define them differently. Here is how the terms are commonly used in 2026.

In the AEO vs GEO debate, the practical gap is small. Answer engine optimization focuses on the question and answer format. GEO adds the wider picture of sources and reputation. The daily work overlaps for the most part.

Our advice: pick one word and spend your energy on the work. We use GEO on this page since it is the term from the original research.

The letters geo have an older meaning in marketing. Geo marketing, often written geomarketing, is marketing based on location: showing ads or offers to people in a certain area. Geo targeting belongs to that same family.

Since 2023 the phrase GEO marketing also means marketing aimed at generative engines. When an agency offers GEO marketing or GEO SEO, ask which of the two it means. This page covers the AI meaning.

The two do meet for local businesses. An assistant that answers a question about a plumber in your town uses location as well. Your town, your service area and your address need to be stated clearly on your site and profiles.

Google published a guide in May 2026 on optimizing for generative AI features in Search. Its position is plain: optimizing for generative AI search is optimizing for the search experience, and therefore still SEO.

The guide also clears up several GEO myths for Google's own AI features. Three points stand out.

1. Special files are optionalGoogle says Google Search itself leaves llms.txt files unused. You are free to create one for other services.
2. Write for peopleGoogle says a special writing style for AI is unnecessary, and so is chopping your content into tiny pieces.
3. Structured data is a plusGoogle says structured data is helpful for rich results and is optional for generative AI search.

Google's main recommendation is content that people find unique, helpful and reliable. Keep in mind that this guide speaks for Google Search. ChatGPT, Claude and Perplexity are separate products with their own crawlers and their own choices.

The shift from links to answers is measurable. These figures come from the companies themselves and from independent research. We list the source and the year for each.

Two conclusions follow. A large audience now reads AI answers every week. And when an answer appears, fewer people click through to a list of links, so the names inside the answer carry more weight.

Keep perspective as well. Regular search still sends most visits to most websites. GEO is an addition to your SEO, and the two reinforce each other.

On 28 September 2026 we asked Claude, with web search switched on, two questions about 100 Dutch small businesses. The sample covered 5 businesses in each of 20 industries. It is a one-day snapshot, and each question was asked once.

Three findings matter for anyone starting with GEO. First, access is rarely the problem. All 100 websites allowed the search crawlers of the AI assistants, and 97 placed their text directly in the HTML.

Second, facts go wrong often. In 29 of 100 answers the assistant stated a fact that conflicted with the business's own website, usually an address, a phone number or opening hours.

Third, the assistant looks beyond your site. It consulted 12.1 websites per answer on average. In 81 of 100 answers at least one was a review site, a directory or a list site. The business's own website was among the sources in 43 answers.

We repeated the questions in ChatGPT for 24 of the businesses. ChatGPT recommended 13 and Claude 9, and 8 businesses were recommended by both. Each assistant makes its own selection.

The companies behind the assistants keep their exact selection rules to themselves. Their documentation, the research and our own measurements point to the same six factors.

1. A crawlable and readable siteThe crawlers of the assistants need access, and your text needs to sit in the HTML of the page. Content that appears only after scripts run may stay unread.
2. Clear factsState who you are, what you do, where you work, your hours and your prices or price ranges in plain sentences. An assistant quotes exact facts and skips vague slogans.
3. Mentions on other sitesAssistants check what independent sources say. Directories, industry guides, local news and best-of lists all feed the answer.
4. ReviewsReviews on platforms the assistants read give them reasons to choose you and words to describe you.
5. Structured dataCode that labels your business details for machines. It helps systems read your name, address and hours the same way every time.
6. FreshnessLive search favors current information. Keep prices, hours, team and services up to date, and show a date on articles.

An honest note from our study: a higher technical score showed zero measurable link with being recommended. Businesses above the median were named 35 percent of the time, those below it 33 percent. Technique makes you readable. Reputation gets you chosen.

An assistant that recommends a business takes a risk with its own credibility. It therefore looks for agreement between sources. A company that appears on a review platform, in a directory and in a local guide is a safer pick.

In our Dutch sample the most consulted sources were Wikipedia pages about the town, review and quote platforms, and national business directories. In the ChatGPT subsample, Reddit appeared among the sources in 17 of 24 answers.

Find the equivalents for your country and trade. Ask the assistant your customer's question and read the source list under the answer. That list is your map of where to be present.

Earn those mentions honestly. Google's guide warns that chasing inauthentic mentions across the web is less helpful than it seems, and its spam systems act against manipulation. A complete profile and real customer reviews are the durable route.

You can carry out this plan yourself. Each step takes between one hour and an afternoon. Work in this order, since each step builds on the one before.

1. 1. Run a baseline testAsk ChatGPT, Gemini, Claude, Perplexity and Google two questions: what do you know about my business, and which business in my trade and town do you recommend. Save the answers.
2. 2. List the errors and the sourcesMark every wrong fact. Write down each website the assistants list as a source. These two lists steer the rest of the plan.
3. 3. Check accessOpen yourdomain.com/robots.txt and confirm that the search crawlers of the assistants are allowed. The technical section below names them.
4. 4. Fix your facts on your own sitePut your name, address, phone number, hours, service area and services in plain text on your contact page and in your footer.
5. 5. Make your facts identical everywhereCorrect your Google Business Profile, directories, review platforms and social profiles so they match your website exactly.
6. 6. Add structured dataAdd LocalBusiness or Organization markup with your business details and links to your profiles.
7. 7. Claim the sources from step twoCreate complete profiles on the review sites, directories and guides the assistants consulted, and ask satisfied customers for reviews there.
8. 8. Rewrite your key pages around questionsUse the questions customers ask as headings. Answer in the first two sentences, then add exact facts, a table or steps.
9. 9. Measure again every quarterRepeat the baseline test with the same questions. Compare mentions, errors and sources with the previous round.

For the hands-on detail behind each step, read chapter 12 of our SEO guide. It walks through the work per assistant and includes a 90-day schedule.

A crawler is a program that fetches web pages. Robots.txt is a small text file on your site that tells each crawler what it may fetch. The AI companies run separate crawlers for search, for user requests and for model training.

For visibility, allow the search crawlers and the user crawlers. The training crawlers GPTBot and ClaudeBot and the token Google-Extended are a separate decision. Blocking them leaves your place in search answers intact.

Three details from the official documentation. OpenAI and Perplexity both say a robots.txt change takes about 24 hours to take effect. Anthropic says its crawlers honor robots.txt and support a Crawl-delay line to slow them down.

OpenAI and Perplexity add that their user crawlers act on behalf of a person, so robots.txt rules may apply differently to them. Also check your firewall or security plugin, since those can block crawlers that robots.txt allows.

An llms.txt file is a text file in Markdown format, placed at the root of your site. It gives a short summary of your site and a list of links to your most important pages. Jeremy Howard of Answer.AI proposed it in September 2024.

The idea is sensible: give AI systems a clean, short guide to your site so they can skip menus, scripts and clutter. The proposal is aimed mostly at AI agents and tools that read documentation.

Now the honest status. It is a community proposal, and a formal web standard is still absent. Google's guide says Google Search leaves these files unused. Google's John Mueller said in June 2025 that the consumer chatbots fetch pages, and skip the llms.txt file.

In the crawler documentation of OpenAI, Anthropic and Perplexity that we checked in October 2026, robots.txt is the control they describe. We found zero statement there that their assistants read llms.txt to choose which business to name.

Our own data agrees. In our study 14 of 100 websites had an llms.txt file, and those businesses showed zero measurable advantage in recommendations. Treat the file as optional housekeeping. It takes ten minutes and does zero harm.

Schema, also called structured data, is a block of code on your page that labels facts in a fixed vocabulary from schema.org. It tells a machine: this is the business name, this is the address, these are the opening hours.

For a business, the useful types are LocalBusiness or Organization for your details, Service or Product for what you sell, and FAQPage for questions and answers. The sameAs property links your site to your official profiles elsewhere.

Be realistic about its role. Google says structured data is optional for its generative AI features. In our study only 11 of 100 websites had complete business details in schema, and schema alone showed zero measurable link with being recommended.

We still advise it. Schema supports rich results in Google and gives every system one clean, consistent copy of your facts. That helps against the conflicting details we found in 29 of 100 answers. Make sure the schema matches the visible text.

GEO has zero fixed rankings to look up. Answers change per moment, per user and per assistant. You measure with three methods side by side.

1. Ask the assistants yourselfKeep a fixed list of five to ten customer questions. Ask them every quarter in each assistant, in a fresh chat, and log whether you are named, how you are described and which sources appear.
2. Read your referral trafficYour analytics tool shows visits from chatgpt.com, perplexity.ai, gemini.google.com, copilot.microsoft.com and claude.ai as referrals. Build one report that groups them.
3. Use the search consolesGoogle Search Console counts AI Overviews and AI Mode inside the regular Web totals of the Performance report. Bing Webmaster Tools added an AI Performance report in February 2026 that shows citations in Copilot.
4. Consider a tracking toolSEO suites such as Semrush and Ahrefs now offer AI visibility reports that run prompts at scale. They are useful for larger brands. A small business can start with a spreadsheet.

One limit to know: many people read an AI answer and visit you later by typing your name. Those visits show up as direct or branded traffic. Watch your branded searches in Search Console as a second signal.

Most mistakes come from treating GEO as a trick. These are the ones we see most often.

1. Skipping the SEO baseEvery assistant stands on a search index. A site that search engines struggle to read gives GEO little to work with.
2. Blocking the wrong crawlerSome site owners block every AI crawler to protect their content and remove themselves from AI search answers in the same move. Decide per crawler.
3. Betting on one fileAn llms.txt file or a block of schema alone changes little. They support the work and they are the smaller part of it.
4. Looking only at your own websiteThe assistants read reviews, directories and lists. Half of the work happens on other sites.
5. Leaving old facts onlineAn old address on a forgotten profile can end up in an answer. Clean up every listing.
6. Testing onceOne answer is a snapshot. Repeat your questions and look at the trend.
7. Buying mentionsPaid lists and fake reviews are a risk for your reputation and your search visibility. Real customers and real sources last.

The honest answer has three speeds, and each one follows from how the assistants work.

Plan for a first round of fixes in a month and a steady routine after that. Measure every quarter. Promises of a guaranteed spot in ChatGPT within a week deserve a critical look, since each assistant decides for itself.

Delta Six is a Dutch web design agency. Optimization for AI search is part of every website we build, next to monthly SEO. You get the foundations from this page as standard.

That means pages with text in the HTML, clear facts in plain sentences, structured data for your business details, open access for the search crawlers and content built around the questions your customers ask. Review follow-up is included too.

We build your complete custom website free first and show it to you in a 15-minute online showcase. You pay nothing upfront and only pay when you're happy. Our page on AI SEO services explains the approach in more detail.

• Ask ChatGPT, Gemini, Claude, Perplexity and Google what they know about your business by name.
• Ask each of them which business in your trade and town they recommend, and save the answers.
• Mark every fact in those answers that differs from your own website.
• Write down every source the assistants list under their answers.
• Open your robots.txt and confirm that OAI-SearchBot, Claude-SearchBot, PerplexityBot, Googlebot and Bingbot are allowed.
• Decide separately whether you allow the training crawlers GPTBot and ClaudeBot and the token Google-Extended.
• Check that your main text is visible in the page source of your home page.
• Put your name, address, phone number, hours and service area in plain text on your site.
• Make those details identical on your Google Business Profile, directories, review platforms and social profiles.
• Add LocalBusiness or Organization schema that matches the visible text.
• Claim complete profiles on the review sites and directories the assistants consulted.
• Set a fixed routine for asking satisfied customers for a review.
• Rewrite your top service pages with customer questions as headings and the answer in the first two sentences.
• Register your site in Google Search Console and Bing Webmaster Tools and submit your sitemap.
• Build an analytics report for visits from AI assistants and repeat the full test every quarter.