How does ChatGPT choose its sources?

Two ways ChatGPT answers, and why the difference matters

ChatGPT produces answers in two fundamentally different ways, and your visibility works differently in each.

The first is from its training. The model has read an enormous amount of the public web, and when it answers without searching, it draws on the patterns of that reading: which names appear alongside which topics, which explanations recur, which businesses are consistently described as doing what. There is no live lookup happening and no citation to win. Your presence in these answers reflects your footprint across the web over time: how often, how consistently and in what company your business has been written about. That footprint is earned slowly, which is precisely why it is defensible once you have it.

The second is live search. When a question needs current information, ChatGPT searches the web, reads a set of pages, and composes an answer with linked citations to the pages it used. OpenAI operates three distinct agents, documented on its bots page: OAI-SearchBot indexes pages for search-grounded answers, ChatGPT-User fetches a page when a user's question sends ChatGPT to it directly, and GPTBot gathers material for training. Each identifies itself in your server logs and can be controlled through your robots.txt file, which means the first practical check is embarrassingly simple: make sure you are not blocking the ones that carry visibility. Some websites still block everything AI-related by accident, left over from a blanket rule someone added in 2023, and OpenAI is explicit that sites blocking OAI-SearchBot will not appear in ChatGPT's search answers.

OpenAI has also signed licensing agreements with major publishers, announced through its newsroom, including the Financial Times, Axel Springer, News Corp and Reddit, which shapes what the system treats as reliable reference material for news and discussion topics. You cannot become a licensing partner as an SME, and you do not need to: for commercial and local questions, the live-search route is where the winnable ground sits.

What the live-search route actually rewards

When ChatGPT searches, it behaves like a very fast, very literal researcher. It issues queries, receives ranked results, opens a handful of pages and extracts what it needs. Three things decide whether your page is in that handful and whether it survives the extraction.

You have to be findable in ordinary search first. ChatGPT's search retrieves from a web index the way any search engine does; a page that ranks nowhere for the topic is unlikely to be retrieved at all. This is why we hold the position that AI search is not a separate product: the ranking work is the entry ticket. The AI search work sits inside SEO because that is factually where it sits.

The page has to answer, not gesture. The system is assembling a specific answer under time pressure. A page that states plainly what you do, for whom, where, at what standard and with what evidence gives the model liftable material. A page of atmospheric brand copy gives it nothing to quote. Question-led headings, direct opening sentences and specifics beat cleverness every time here.

The entity has to be legible. ChatGPT resolves businesses as entities: a name, connected to a place, services, people and a track record, consistent across your website, your profiles and what others write about you. Inconsistent naming, thin third-party presence or contradictory information makes the entity blurry, and blurry entities get left out of answers because the model cannot be confident what it would be recommending. Structured data helps machines confirm what the words say; the entity work is covered in depth on our GEO page.

What earns the citation

Across the answers we check by hand for clients, the pages ChatGPT cites for commercial questions share a recognisable profile: they rank for the topic, they answer directly high on the page, they carry verifiable specifics (numbers, names, dates), they come from sites whose wider footprint corroborates their authority, and they are recent enough to trust. One of our own clients, a commercial coffee supplier, appears in ChatGPT answers when buyers ask for a UK supplier; the work behind that was structural clarity and earned authority, not a trick, and there is no trick to buy.

Two honest caveats. First, answers vary: the same question asked twice can cite different sources, so presence is a probability you improve, not a position you hold. Second, the mechanics above describe the engine as it works at the time of the last review date on this page; OpenAI changes its products quickly, which is exactly why this guide lives at a stable address and gets updated rather than replaced.

What we do about it, and what you can do today

The work that moves ChatGPT visibility is the work we run inside SEO: entity and structured-data clarity, pages rewritten so they answer, authority earned from sources the engines already trust, and manual checking of the actual answers buyers see, because the tools that claim to automate that measurement lag the engine. If you want the fast version for your own business, the AI-search visibility check is free: we ask the questions your buyers ask and send you a plain-English note on who gets named and why.