GEO: how to get your company cited by ChatGPT, Claude and Gemini

What is Generative Engine Optimization (GEO)?

Generative Engine Optimization (GEO) is the practice of structuring your content, your data and your brand presence so that AI engines such as ChatGPT, Claude, Gemini, Perplexity and Google AI Mode retrieve your pages, cite them and repeat your facts in their answers.

The term comes from a paper by Pranjal Aggarwal and colleagues at Princeton, published at the KDD 2024 conference, which introduced GEO as a way to improve content visibility in generative engine responses. According to that paper, targeted changes to a page raised its visibility by up to 40%. Visibility here is not a ranking position: it is how much of the final answer is built from your source and how prominently you are named.

The stakes are no longer theoretical. According to Similarweb's July 2026 AI search report, AI platforms received about 9.5 billion visits a month, up 70% year on year. The same report shows the market fragmenting: ChatGPT's share of the traffic AI tools send to websites fell from 76% to 53% between June 2025 and May 2026, while Gemini climbed to roughly 27% and Claude to about 9%. Show up in one engine only and you are invisible to half of the audience.

How is GEO different from SEO?

SEO gets you ranked on a results page; GEO gets you quoted inside an answer, and an answer only has room for a handful of sources.

Most AI engines work in 2 steps. First they retrieve candidate documents, usually from a conventional search index or their own crawler. Then a language model reads those documents and writes 1 answer with citations. That is why SEO is still the entry ticket: Google's own guide for generative AI features says there are no additional requirements beyond being indexed and eligible to show a snippet. If you are not crawlable, you are not in the pool.

The differences start after retrieval. Winner takes most: a results page has 10 organic slots, an AI answer typically leans on 3 to 8 sources. Passages beat pages: the model lifts the paragraph that answers the question, so paragraph structure matters more than the page title. Entities beat keywords: the engine needs to be sure who you are. And a mention is not a click. According to Similarweb, the share of ChatGPT prompts that returned a citation rose from 1.6% in June 2025 to 6.8% in May 2026, so most of the time you are described without being linked. Being described correctly is the goal.

What do AI engines actually cite?

AI engines cite content that answers a specific question directly, with named numbers, quotable definitions and a visible source, on a page they can crawl and already trust.

The Princeton paper tested 9 optimisation methods on the same pages. Adding statistics, quotations from named sources and explicit citations produced the largest gains, roughly 30% to 40% on their position-adjusted metric, while stuffing keywords produced nothing or made things worse. That matches what I see in my companies: a paragraph that says "according to our 2025 customer data, onboarding takes 9 days" gets lifted, a paragraph that says "best-in-class onboarding" does not.

A March 2026 arXiv paper on diagnosing citation failures adds a useful nuance: pages fail to be cited at different stages (not retrieved, retrieved but not selected, selected but not attributed) and each failure needs a different fix. Their targeted approach improved citation rates by more than 40% while changing only 5% of the content, against 25% for generic rewriting. In plain terms: do not rewrite your whole site, fix the specific reason each key page is skipped.

Two more signals matter in practice. Freshness: put a visible date on every article and update it when the facts change. Entity clarity: I write "Xavi Creus, Barcelona" the same way everywhere for a reason. If the model cannot tell your company from a similarly named one, it will not risk naming you.

Do llms.txt and schema.org actually help?

They help machines read you, but neither is a shortcut: Google states explicitly that no special markup or llms.txt file is needed to appear in AI Overviews and AI Mode, while llms.txt is becoming standard for documentation that developers' AI agents consume.

llms.txt is a plain markdown file at the root of your domain, proposed by Jeremy Howard, that gives a language model a short summary of your site and a curated list of links. The specification reached version 2 in August 2026 and notes that OpenAI, Anthropic and Google publish llms.txt files for their own developer documentation, and that Chrome's Lighthouse now audits for the file. If you sell software with public documentation, publish one. If you sell consulting, it will do little on its own.

schema.org structured data (Organization, Person, Article, FAQPage) is the older, more established layer. It tells crawlers who wrote what, when, and how entities relate. Google's guide says it is not required for generative AI features but remains valuable for rich results. My rule: implement both because they cost an afternoon, and expect nothing from them unless the content underneath is worth quoting.

How do I measure AI visibility?

You measure GEO by asking the engines your buyers' real questions on a fixed schedule and recording whether you are mentioned, cited and described correctly.

In my companies we keep a panel of 30 to 50 prompts per company, the questions a buyer would type before a purchase, and we run them monthly in ChatGPT, Claude, Gemini, Perplexity and Google AI Mode. We record 3 things: share of answers that mention us, share that link to us, and whether the description is accurate. Answers vary between runs, so we run each prompt several times and look at trends, not single results.

On the analytics side, segment referral traffic from AI domains (chatgpt.com, perplexity.ai, gemini.google.com, claude.ai) and watch it next to branded search and direct traffic, because a lot of AI influence arrives as someone typing your name into a browser 2 days later. Review the panel with the marketing team every month.

What is the practical GEO checklist?

A practical GEO checklist is 10 items long and none of them requires a new tool: clear questions, direct answers, original numbers, visible sources, consistent entities, dates, crawlability, structured data, llms.txt and third-party mentions.

This is the list we use across the companies I run. It is deliberately boring: the companies that get cited are the ones that are easiest to quote accurately.

• One question per H2 heading, phrased the way a buyer asks it, with a 1-sentence direct answer as the first line.
• Publish original numbers from your own operations, with the date and the sample, because a statistic nobody else has is the most quotable asset you own.
• Name your sources in the text ("according to ...") and link them, so the engine can verify and attribute.
• Write definitions that can be lifted verbatim: "X is ..." in 1 sentence, before any nuance.
• Add a short FAQ to every important page and mark it up with FAQPage schema.
• Use 1 canonical name for the company, the founder and the city everywhere: website, LinkedIn, directories, press.
• Show a visible published and updated date, and refresh key pages at least quarterly.
• Keep pages indexable and fast, with content in the HTML rather than loaded only by scripts, and decide explicitly in robots.txt which AI crawlers you allow.
• Implement schema.org Organization, Person and Article, and publish an llms.txt if you have documentation.
• Earn mentions on pages the engines already trust in your category: review sites, industry directories, trade press and partner sites.