The first step to being found: being understood.

There is no shortcut to visibility in AI search, but there is a clear order: make the content understandable, then crawlable and quotable, then measure it. This guide covers what AEO and GEO mean, how the major AI search systems choose sources, and the steps we have implemented on globalmeta.net.

What are AEO and GEO?

Answer Engine Optimization (AEO) is the practice of preparing content so that it can be selected as a direct, accurate answer to a question. Its roots lie in featured snippets and voice assistants; its core question is whether a page answers a question clearly.

Generative Engine Optimization (GEO) aims to be cited as a source in answers generated by search experiences built on large language models, with the brand represented accurately. The term spread through an academic study published in 2023.

Neither replaces SEO. Most AI search systems run a classic search and crawl step before generating an answer, so a page that is not crawled, not indexed or unclear finds no place in either. What has changed is where content is read: as part of an answer, often before anyone clicks.

How do AI search systems choose their sources?

The internals are not public, but documented behaviour draws a shared picture: answers are compiled from web content accessible at that moment, and sources are shown as links.

Google AI Overviews and AI Mode

According to Google, for a page to appear as a supporting link in AI Overviews or AI Mode, it only needs to be indexed and eligible to be shown in Google Search with a snippet; there are no additional technical requirements. Google also notes that these features may use a “query fan-out” technique, splitting a question into subtopics and running several related searches, so pages that answer a topic’s sub-questions can surface. Preview controls such as nosnippet, data-nosnippet and max-snippet apply here too.

ChatGPT

ChatGPT search retrieves current web results and shows source links within the answer. OpenAI uses the OAI-SearchBot crawler for this, separate from GPTBot, which it uses for training. You can stay closed to training and open to ChatGPT search; blocking OAI-SearchBot keeps your site out of those answers.

Perplexity

Perplexity searches the web for every question and answers with numbered sources, building its index with PerplexityBot. To make the list, a page must be crawlable and contain a section that answers the question directly.

Claude

With web search enabled, Claude cites the sources it used. Anthropic defines separate crawlers: ClaudeBot for training, Claude-SearchBot for search and Claude-User for fetching pages at a user’s request. Which you allow decides which uses your content is open to.

The common denominator: a system cites a page only if it can reach it, understand it and trust it.

Content format: answer first, detail second

AI answers usually quote a paragraph or a list, not a whole page, so each section should make sense on its own.

• Answer-first paragraphs: if a heading is a question, the first sentence or two should answer it; context and exceptions come after.
• Question-style headings: headings in the language people type into a search box or an assistant.
• Information as text: address, contact details, scope and dates should not live inside an image.
• Frequently asked questions: short, self-contained answers to the most common sub-questions.
• Date and author: publication and update dates show how current the content is.
• Internal links: link a concept to the page that explains it where you first use it.

This article follows the same rule: the first paragraph under each heading is its short answer.

Structured data: the machine language of visible content

Structured data describes what the information on a page is, using the schema.org vocabulary, usually in JSON-LD. Google says no special schema is needed for AI features, but structured data removes ambiguity about what an organisation, an article or an FAQ is.

On globalmeta.net, every page is linked to a single Organization record that describes the company; Insights articles are marked up as Article, the FAQ page as FAQPage and glossary terms as DefinedTerm. The rule is simple: markup only repeats what is visible on the page.

llms.txt, Markdown versions and MCP

These three tools give AI systems a cleaner way to read your site. None guarantees visibility, and Google states that no such file is needed for its AI features. But agents increasingly come straight to a site to read it, and keeping that path tidy is a cheap, low-risk investment.

llms.txt

llms.txt is a Markdown file, proposed in 2024, that sits at the root of a site: who is the site for, what does it cover, and where are the most important pages? Our llms.txt lists services, work, products and the glossary in both languages, while llms-full.txt gathers the full text of the site in a single file.

Markdown versions

Every page on the site has a Markdown copy; adding .md to the end of the address is enough (for example, the Markdown version of this article). The HTML page points to this copy with a rel="alternate" type="text/markdown" link, and a request sent with an Accept: text/markdown header receives Markdown directly. The Markdown copies are marked noindex, so they do not enter the search index as a second copy of the HTML.

MCP

The Model Context Protocol (MCP) is an open protocol that lets AI applications connect to tools and data sources in a standard way; Anthropic announced it at the end of 2024. A read-only MCP server runs at globalmeta.net/mcp: an agent can search the site (search_site), list pages (list_pages), fetch a page as Markdown (get_page) and read the contact details (get_contact). The server card is at /.well-known/mcp/server-card.json.

robots.txt and Content-Signal: which uses do you allow?

The quiet setting of AI visibility is robots.txt. Block the bots that crawl for search and even the best content cannot reach answers. Yet “being open to AI” is not one decision: search, answer generation and model training are separate uses.

Our robots.txt states this explicitly with a Content-Signal line: search=yes, ai-input=yes, ai-train=yes. In other words, our content may be used in search results, AI answers and model training; we only ask that globalmeta.net is cited as the source. We do not block any AI crawler. Content-Signal states a preference rather than enforcing it, so review crawler permissions separately too.

How do you measure AI visibility?

AI visibility does not show up in a single dashboard; read several sources together.

• Search Console: impressions and clicks from AI Overviews and AI Mode are included in the Web search type totals, with no separate filter. Track trends by page and query.
• Analytics: group referrals from chatgpt.com, perplexity.ai and claude.ai into their own channel. ChatGPT often adds utm_source=chatgpt.com to its links.
• Server logs: which pages OAI-SearchBot, PerplexityBot and Claude-SearchBot read, and how often.
• Question panel: put a fixed list of customer questions to assistants at regular intervals and record whether your brand is mentioned and which source is cited. Answers vary, so watch the trend, not a single attempt.

In our own panel, Globalmeta Board, a Footprint score does this job; we explained the thinking behind it in The Meeting Where Everyone Is Right.

AI visibility checklist

• Each service and product page says what you do in its first paragraph.
• Key information is in text, not images.
• Headings match questions; first sentences answer them.
• Organisation details are identical on every page and platform.
• Organization, Article and FAQPage markup matches the visible content.
• robots.txt allows search crawlers (Googlebot, OAI-SearchBot, PerplexityBot, Claude-SearchBot) and states the training preference separately.
• llms.txt is current and links to the key pages.
• Markdown versions match the HTML.
• Any MCP or API endpoint for agents is read-only and documented.
• Search Console, analytics, server logs and the question panel are reviewed regularly.

Frequently asked questions

What is the difference between AEO and SEO?

SEO aims for a page to be found and clicked in search results; AEO aims for its information to be selected as a direct answer. Both rest on crawlable, understandable, trustworthy content.

Is an llms.txt file mandatory?

No. llms.txt is a proposal, not a standard, and Google says its AI features do not need it. It is still an easy, useful way to give agents a summary of the site.

Can a place in AI answers be guaranteed?

No. No platform guarantees citation, and answers vary by question, person and time. You can make content accessible, understandable and quotable, and measure the result.

Can you stay closed to model training and still appear in AI search?

Yes. Most providers use separate crawlers for search and training, so you can allow one and block the other in robots.txt, and state your preference with Content-Signal.