How ChatGPT decides which websites to recommend

Two ways ChatGPT answers

ChatGPT answers a question in one of two ways.

From memory. The model answers from what it learnt in training. That knowledge stops at a cutoff date, and you cannot add your new page to it.

From the web. When a question needs current information, ChatGPT searches, reads a handful of pages and writes an answer with links to its sources. OpenAI calls this ChatGPT search. This is the part a website owner can do something about.

The four OpenAI crawlers

OpenAI runs four user-agents, and much advice about ChatGPT mixes them up:

• GPTBot collects pages that may be used to train future models.
• OAI-SearchBot crawls pages for ChatGPT search.
• ChatGPT-User fetches a page when a person asks ChatGPT to look at it. OpenAI says robots.txt rules may not apply to it, because a person started the fetch.
• OAI-AdsBot checks pages submitted as ChatGPT ads. OpenAI says what it collects is not used for training.

For appearing in ChatGPT's search answers, OAI-SearchBot is the one that counts. A site that blocks GPTBot and allows OAI-SearchBot keeps its text out of training and can still be a source in search answers. A site that blocks OAI-SearchBot is left out of those answers, by OpenAI's own account.

So check your robots.txt for a User-agent: OAI-SearchBot group with Disallow: /, and for a User-agent: * group that blocks everything. Either one shuts ChatGPT search out.

What makes a page easy to pick

OpenAI has not published how ChatGPT search chooses its sources. What follows is what makes a page easy to use as a source, not a formula that gets you picked.

Answer the question on the page. A search system looks for a passage that answers the query. A heading that asks the question and a first paragraph that answers it is easier to lift than an answer buried under four paragraphs of preamble.

Be specific. "Twelve rooms, free parking, dogs welcome" is something a model can repeat. "A great stay in the heart of the city" is not.

Cover the topic properly. A page of 200 words rarely answers a real question in full. Examples, the obvious follow-up questions and the awkward cases are what separate a source from a filler page.

Say who you are. An Organization or LocalBusiness block in Schema.org, with your name, address, phone and sameAs links to your social profiles, states in machine-readable form what the page text says about your business. OpenAI has not said whether ChatGPT reads structured data, so count this as useful for every system that does, not as a known ChatGPT input.

Keep dates honest. An Article with an accurate dateModified and a sitemap with correct lastmod values tell a crawler what changed. Change the date only when the content changed.

And llms.txt

llms.txt is a short Markdown summary of a site at a fixed address. OpenAI has not said that ChatGPT or any of its crawlers reads it. It costs little to publish, and a tool a person points at your site can read it, but I would not count on it for ChatGPT search.

What AI Boost for Joomla does here

AI Boost for Joomla publishes signals that ChatGPT and other assistants can read. Each works once you switch it on:

• robots.txt answers 25 AI crawlers one by one, GPTBot, OAI-SearchBot and ChatGPT-User among them, and labels which one trains, which one searches and which one fetches a page a person asked for.
• Schema.org JSON-LD describes your business as one of 36 site types, with sameAs links to your social profiles, and each article as an Article or BlogPosting with its author and dateModified.
• FAQPage markup comes from headings on your pages that end in a question mark, and, with the YOOtheme extension, from a YOOtheme accordion whose every title is a question.
• The XML sitemap carries a lastmod for each article, taken from its modified date.
• llms.txt is built on each request from your site's details, menu pages and recent articles.

It does not write your pages, and it cannot make ChatGPT cite or recommend you. That still comes down to whether your pages answer the question.