Generative Engine Optimization: Get Cited by ChatGPT, Gemini, Grok

What is generative engine optimization (GEO)?

GEO is everything that makes a brand findable, readable and credible to AI assistants, so they name it and cite its pages when a prospect asks them a question. It is SEO applied to written answers instead of a list of links.

A small business owner asks ChatGPT: “What invoicing software should a 10-person agency use?” The answer names three brands, with two links to a comparison article and a forum. Yours is not there, even though your pricing page ranks well on Google. That prospect will never see your site.

The term comes from a research paper, “GEO: Generative Engine Optimization”, published in November 2023 and accepted at KDD 2024. This guide is the pillar page of a series: each section links to the guide that covers the topic in depth.

What is the difference between a mention and a citation?

A mention is your brand name in the text of the answer. A citation is a link to your page in the answer’s sources. Measure them separately, because they do not always show up together.

In practice there are four cases. Named and cited: the best case. Named without a link: the assistant knows you, often thanks to third-party sites. Cited without being named: your article helped build the answer, but the recommendation goes to a competitor. Neither: you are missing from the answer.

The third case is the most frustrating, and often the easiest to fix: your page gets read, but it lacks a clear statement of what you sell and who it is for. The guide why ChatGPT recommends your competitor walks you through a step-by-step diagnosis.

How does an AI assistant build its answer?

An assistant answers either from memory, using its training data, or after a web search whose results it reads before writing. On buying questions (“best tool for…”, “which provider in…”), assistants often run a web search (this is observed, not documented), and that search is what produces the citations.

flowchart TD
  A[Prospect question] --> B{Web search?}
  B -->|No| C[Model memory]
  B -->|Yes| D[Multiple queries]
  D --> E[Search index]
  E --> F{Page accessible?}
  F -->|Crawler blocked| G[Page skipped]
  F -->|Yes| H[Pages read]
  C --> I[Written answer]
  H --> I
  I --> J[Mentions in the text]
  I --> K[Citations in sources]

What each provider documents about web search:

For Grok, xAI documents its API’s Web Search tool, which “enables Grok to search the web in real-time and browse web pages to find information.” As of October 8, 2026, no official xAI page describes a crawler to allow in robots.txt: the names going around come from third-party directories. On the Gemini side, the API documentation says grounding “connects the Gemini model to real-time web content” and lets it “cite verifiable sources.”

What do providers actually document?

Three things: your pages must be accessible to their search crawlers, indexed, and your important content must exist as text. Everything else is observation.

Crawler access. OpenAI clearly separates training (GPTBot) from search (OAI-SearchBot), and says what happens if you block the second one:

So blocking OAI-SearchBot takes you out of ChatGPT search answers. Blocking GPTBot only signals that your content should not be used for training. OpenAI also notes that a robots.txt change takes about 24 hours to be reflected in search. The guide GPTBot, OAI-SearchBot and robots.txt covers each crawler and the Cloudflare trap: on July 1, 2025, Cloudflare announced it was changing its default setting to block AI crawlers. If your site runs through Cloudflare, check which crawlers that setting stops: a block applied by default does not necessarily show up in your robots.txt.

Indexing. For AI Overviews and AI Mode, Google asks for nothing beyond what classic search requires:

The only technical requirement: a page that is indexed and eligible to be shown with a snippet. Watch out for the Google-Extended token: according to Google, it changes nothing in Google Search, but it controls whether your pages are used to train Gemini and for grounding in the Gemini apps. Blocking it can therefore cost you citations in Gemini.

Content in text. Google recommends making sure important content is available in textual form, that structured data matches the visible text, and that Merchant Center and Business Profile information is up to date. A price shown only in an image, or an offer described only in a video, may slip past a crawler that reads text.

What is observed but not guaranteed?

Several patterns keep coming up in measurements and research, without any provider putting them in writing. Treat them as hypotheses to test on your own set of questions.

• Third-party sources carry a lot of weight. A study by Chen et al. (2025) (a preprint, not peer reviewed) finds that AI search engines “exhibit a systematic and overwhelming bias towards Earned media (third-party, authoritative sources),” at the expense of brand-owned content and social media. An independent comparison that names you often counts for more than one more post on your blog.
• Answers vary. Ask the same assistant the same question twice and it can name different brands. A single screenshot proves nothing: only a set of questions replayed over time shows a trend.
• Server-rendered HTML is safer. OpenAI, Anthropic and Perplexity do not say in their documentation whether their crawlers execute JavaScript. Content present in the HTML the server sends does not depend on the answer.
• Specific pages win. An “invoicing software for agencies” page answers the prospect’s question better than a generic homepage. Answers at the top of each section, comparison tables and public prices lend themselves well to citation.

What are the GEO myths?

The three most heavily sold promises rest on no documentation at all: the magic llms.txt file, mandatory schema.org, and “40% more citations.”

“An llms.txt gets you cited.” llms.txt is a proposal by Jeremy Howard, published on September 3, 2024: a Markdown file that gives agents a summary of the site and links to its key pages. No assistant provider documents using it to choose the sources for its answers, and Google states plainly that you don’t need to create “new machine readable files, AI text files, or markup” for its AI features. It costs little to publish; it is not a documented way to get cited.

“No schema.org, no citation.” Google says there is “no special schema.org structured data that you need to add” for AI Overviews and AI Mode. Markup is still useful for Google rich results and for describing your organization unambiguously, as long as it matches the visible text. The guide schema.org and AI search sorts it out type by type.

“GEO boosts your citations by 40%.” The figure comes from the original paper: on GEO-bench, a lab benchmark, some content rewrites improved visibility “by up to 40%.” That is a maximum, measured with the authors’ own metrics, in a controlled environment. It is not a promise you can apply to your brand.

“Blocking GPTBot erases you from ChatGPT.” False, according to OpenAI: OAI-SearchBot is what governs search. The real risk is the opposite: a robots.txt file or firewall that blocks every AI crawler, search crawlers included.

What action plan should you follow to get cited?

Start with technical access, measure your starting point, then work on content and third-party mentions, in that order.

1. Open the door. Check robots.txt, Cloudflare and your firewall: OAI-SearchBot, Claude-SearchBot, PerplexityBot and Googlebot must get through. Decide separately on training crawlers. See the AI crawlers guide.
2. Get indexed. Verify your site in Google Search Console and check that your key pages are indexed with a snippet. Traffic from AI Overviews and AI Mode shows up there, counted under the “Web” search type.
3. Pick your questions. List 10 to 20 questions a buyer asks without knowing your name: “best… for…”, “alternative to…”, “how much does… cost”. The method is in choosing the questions to track.
4. Measure your starting point. For each question and each assistant: are you named, cited, and who shows up instead of you? AI share of voice sums up these answers in one number you can compare with your competitors’.
5. Write the pages that answer. One page per use case, the answer in the first sentence, your prices and limits in text, honest comparisons. AI can help you write, as long as you follow Google’s rules: see writing with AI without a penalty and a WordPress blog on autopilot.
6. Exist elsewhere. Find the comparisons, directories, forums and media cited in the answers to your question set, and get named there: product listing, customer reviews, guest article, a useful reply in a discussion.
7. Replay and adjust. Ask the same question set at a fixed interval. A question where a competitor gains ground tells you which page to write next.

How do you measure your visibility in ChatGPT, Gemini and Grok?

Put the same generic questions to the same assistants at regular intervals, and count who gets named. By hand, it works for a few questions; beyond that, you need a tracking tool.

The manual method: a private browsing window, logged out, your question set in a spreadsheet, one column per assistant, and for each answer the brands named and the sources cited. Never put your brand in the question: “Is Tonaily a good tool?” only confirms what the question already says. The limits come fast: 15 questions across 3 assistants means 45 answers to read every round, and the variability of answers forces you to repeat.

For AI Overviews and AI Mode, Google Search Console counts clicks and impressions along with the rest of search, under the “Web” type. To see where your industry stands, the AI Visibility Index ranks the brands AI assistants cite most, industry by industry.

Two limits to know before you choose. Tonaily plans do not track Perplexity, Claude or AI Overviews: if those engines matter most in your market, other tools cover them, and the comparison of AI visibility tracking tools says which ones, with verified prices. And the Free account does not track AI visibility.

For a first snapshot without creating an account, the free test puts a question of your choice to ChatGPT, Claude, Gemini and Perplexity, up to three times a day. The result shows whether your brand is named, by which assistant, with which sources, and which brands show up in its place. It is not tracking: nothing is replayed afterward.

If you already work in Claude, ChatGPT or Cursor, you can plug this data into them: the guide MCP and SEO shows how to query your AI visibility from your assistant with Kira MCP (Growth and Pro plans).

Where should you start this week?

With crawler access and a baseline measurement: both are quick, and everything else depends on them.