Describe in general terms how assistants such as ChatGPT, Perplexity and Copilot find and cite web pages.
People now ask AI chat assistants such as ChatGPT, Perplexity, Microsoft Copilot and Google's Gemini the kind of question they used to type into a search engine. If you run a small business, this raises two practical questions. How does an assistant decide which businesses to mention, and can you do anything to be one of them? Some of the answer is known and some is not, and it helps to keep the two apart.
Mei Ling (an invented example) is a home baker in Punggol who makes custom cakes. She asked a friend how the friend had found a baker for their son's birthday. The friend had asked ChatGPT for "home bakers in Punggol who do custom cakes". It gave four names with links, and the friend messaged two of them. Mei Ling was not one of the four. She had never thought of an AI chat app as a place where customers would look for her, and she did not know how it chose whom to mention.
An assistant can answer you in two ways. The first is an answer from training data alone. A large language model learns patterns from a vast amount of text collected before a cut-off date. When it answers from that training alone, it produces the answer without looking anything up. An answer like this has no citations, because the model is not drawing on a specific page. It can be out of date, because the training data stops at a point in time. It can also be wrong about specifics, such as a business's opening hours or a scheme's current rules. The course "AI fundamentals: what it is, how it works, where it fails" covers in more depth how models generate text.
The second is an answer with a web search. Many assistants can search the web before they answer. Some do it automatically when a question seems to need current information, and others do it when you ask. In this mode the assistant sends one or more search queries, retrieves pages from the results and reads them. It then writes a summary answer, usually with links or citations to the pages it used. The answer Mei Ling's friend got, with four names and links, was this second kind.
To search the web, an assistant needs a search index. This is a stored copy of the web, the same kind described for Google in lesson 1.1, "Crawling, indexing and serving: how a page gets into Google". Some providers run their own crawlers and indexes, some use a partner's search engine, and some combine both. Microsoft says Copilot draws on Bing. Google's Gemini can draw on Google Search. OpenAI and Perplexity document their own crawlers, and lesson 8.4, "AI crawlers and what you can control", covers those. These details change often, and providers update their documentation when they do.
One general principle applies whichever provider you look at. A page is unlikely to be retrieved if it cannot be crawled, if it is not in the indexes assistants rely on, or if it does not clearly answer the question. A page that is never retrieved cannot be cited.
Providers do not publish how they decide which retrieved pages to cite, or in what order. You will find claims online about how to rank first in ChatGPT, and services that guarantee mentions in AI answers. Treat them with the same suspicion that lesson 1.2, "What Google says its ranking systems look at", recommended for ranking-factor percentages. Nobody outside these companies knows the method, and the companies change their systems often.
Answers also vary. The same question asked twice, or by two people, can give different answers and different citations. Several things change the result: the wording of the question, the user's location and settings, and whether the assistant searched the web. Because of this, lesson 8.5, "Check your visibility in AI answers", asks you to look for trends over several months instead of reacting to one answer.
Most of what helps in ordinary search also helps here. First, the page must be crawlable by the major search engines and, if you choose, by AI providers' own crawlers. Second, it should be indexed in the major search engines, which for most sites means Google and Bing. Bing Webmaster Tools is Bing's equivalent of Google Search Console and works in much the same way. If assistants that use Bing matter to your customers, it's worth verifying your site there. Third, the page should clearly answer a question someone would ask. It should say plainly who the business is, where it is and what it offers.
This is not a separate discipline from ordinary SEO. The pages that do well in ordinary search for a question are usually the ones assistants retrieve when they search the same question.
Mei Ling's homepage said "Handmade with love", and Punggol appeared only in the footer. She could instead have a page whose title and text are about custom cakes from a home baker in Punggol, with her name and her area written into the text. That would give an assistant something to match and quote. Lesson 8.3, "Write passages an assistant can quote", shows how to write pages like that.
You can check what assistants say about your kind of business yourself. Pick a question your customers would genuinely ask, in the words they would use. Ask it of three assistants that can search the web, making sure web search is switched on where there is a setting for it. For each answer, note whether it searched the web, which businesses it named and which sources it cited. The activity below asks you to record those citations side by side.
Ask three AI assistants, with web search on, the same customer question about your field. Write down which sources each one cites.
Junxiong-WFG Organisation is an authorised representative of AIA Financial Advisers Private Limited (Reg. No. 201715016G).