How Does ChatGPT Choose Sources to Cite? Our 20-Question Test

/ AI Visibility 
Nov 19, 2026 | 7 min read
AI-generated illustration — Empty grey archive room with rows of identical filing cabinets and one folder pulled out and placed on top of a cabinet
AI-generated image
Olga LescuyerFrontend Developer,SEO specialist

How does ChatGPT choose sources to cite?

ChatGPT does not look for the best website on a topic. It turns your question into narrower search queries and cites the pages that answer those queries directly, usually in the words the question uses. When we asked ChatGPT 20 real buyer questions, it named Pixerry 0 times. The pages it cited instead were written for that exact question.

What happens between a question and ChatGPT’s citations?

ChatGPT rewrites the question before it searches. According to OpenAI’s help page on searching the web with ChatGPT, ChatGPT search sometimes works with other search providers. When it does, it typically rewrites your question into one or more targeted queries. In OpenAI’s example, a question about drugs that target CCR8 for cancer becomes the search “CCR8 immunotherapy drug development 2025”. After the first results, ChatGPT may send a narrower query, such as “CHS-114 conference 2025”.

The rewrite also uses context. ChatGPT collects a general location from your IP address, so “good restaurants near me” can become “top restaurants San Francisco”. OpenAI says it does not share the IP address or account details with the search providers. When memory is on, saved memories may also shape the rewritten query.

Google’s AI search works in a similar way. Google calls it query fan-out: a set of related queries run at the same time to fetch more relevant results. In Google’s guide to optimizing for its generative AI features, “how to fix a lawn that’s full of weeds” fans out into “best herbicides for lawns”, “remove weeds without chemicals” and “how to prevent weeds in lawn”. Google’s systems then review specific information from the retrieved pages and link the ones that support the answer.

Researchers describe the same pipeline. The GEO paper by Aggarwal and colleagues (KDD ’24) (GEO stands for generative engine optimization) models a generative engine in four steps. A model breaks the query into simpler queries, a search engine retrieves ranked sources, a model summarizes each source, and a final model writes the answer with inline citations. The authors say this closely matched BingChat’s design when they wrote the paper.

StepWhat the AI engine doesWhat it means for your page
RewriteTurns the question into one or more targeted search queriesYour page competes on the rewritten query, not on the reader’s original wording
RetrieveA search provider returns ranked pages for each queryThe page has to be indexed and reachable by crawlers
ReadA model reviews or summarizes each sourceThe answer has to be easy to find in the text itself
CiteThe final answer links the pages that support specific statementsOnly a page that backs a clear statement gets the link

OpenAI does not publish how ChatGPT picks which retrieved pages to cite. No OpenAI page we read lists ranking factors. So the useful question for a business is narrower: which page matches the rewritten query best?

What did ChatGPT cite when we tested 20 buyer questions?

In November, Olga, Pixerry’s founder, asked ChatGPT 20 questions our buyers really ask, such as “How much does a professional website cost in Switzerland?” and “Which web agency in Lausanne can redo my site?”. ChatGPT named Pixerry 0 times out of 20.

The cited pages taught us more than the score. For the Swiss cost question, all 4 sources were pages written for that one question: three were called “website-cost-switzerland” and one “web-design-prices-switzerland”. For the Lausanne question, ChatGPT cited pages like “refonte-site-web-lausanne” and “agence-web-lausanne”. When we asked how to get cited by ChatGPT, it cited OpenAI’s and Google’s own documentation, not a single agency blog.

Perplexity named Pixerry 3 times. Each time, it cited only our Russian home page.

“ChatGPT doesn’t pick the best website. It picks the page that answers this exact question, in the words the question uses. We didn’t have those pages yet.” Olga, founder of Pixerry

Why does a page written for one question win the citation?

A page written for one question matches the rewritten search query almost word for word. Say a buyer’s question about Swiss website prices becomes a query close to “website cost Switzerland”. A page with that phrase in its address, heading and first paragraph is an easy match. A studio home page that mentions prices somewhere further down is not. This is our reading of the test, not a rule OpenAI has published.

The Lausanne results show the same pattern with place names. The question named the city, and the cited pages carried the city and the service in their address. Our Perplexity result shows the other side. A home page can earn a brand mention, but it cannot answer a specific cost or location question as well as a page built for it.

Do statistics and sources make a page more likely to be cited?

In lab tests, yes. The GEO paper built GEO-bench, a set of 10,000 queries from different domains. It found that adding citations, quotations from relevant sources and statistics raised a source’s visibility in AI answers by over 40%. On Perplexity, a live engine, the authors measured gains of up to 37%, and adding quotations gave a 22% improvement on their position-adjusted word count measure.

Two limits matter. The results vary by domain, and the paper measures how much of a source shows up in the answer, not leads or sales. We treat it as a reason to put a number and its source in each section, not as a promise of citations.

How do you write a section ChatGPT can quote?

Write each section so it answers one question on its own. After the test, we changed how we write Pixerry posts. Most H2s are now the question the section answers, phrased the way a buyer would type it, with the answer in the first sentence underneath. This post follows that rule.

  • Give each important buyer question its own page, with the question’s words in the URL, the title and the first paragraph.
  • Phrase most H2s as questions in the reader’s words rather than clever labels.
  • Answer in the first sentence under each heading, then add the detail.
  • Name the subject in every section instead of “it” or “as mentioned above”, so a quoted section still makes sense alone.
  • Include a number in the section and link where it comes from.
  • Add something only your business can say: a test, a price you have quoted, a project result.

Standalone does not mean tiny. Google says there is no need to break content into tiny pieces for AI to understand it. Its guide also says content people find unique, compelling and useful will likely do more for visibility in generative AI search than any other tip. Google contrasts commodity content like “7 Tips for First-Time Homebuyers” with a first-hand piece like “Why We Waived the Inspection & Saved Money: A Look Inside the Sewer Line”. First-hand data, such as a test of your own buyers’ questions, is the second kind.

Does writing for ChatGPT also help in Google?

Yes, because Google treats it as the same work. Google’s page on AI features in Google Search says there are no additional requirements to appear in AI Overviews or AI Mode, and no special optimizations needed. Google’s guide adds that, from Google Search’s point of view, optimizing for generative AI search is still SEO, built on its core ranking and quality systems.

The technical base still decides whether a page can be cited at all. Google says a page must be indexed and eligible to show in Search with a snippet, and Search Console has a Generative AI performance report to measure results. Google ignores llms.txt files: they neither help nor harm rankings in Google Search. For ChatGPT, check that your robots.txt lets the right AI crawlers in. Our post on which AI crawlers to allow on a business website covers that choice. The wider plan sits in our guide to SEO, GEO and AEO strategy for AI search.

Start with your own list. Write down the 10 questions buyers ask before they contact you, ask ChatGPT each one, and note which pages it cites. Our post on checking whether ChatGPT recommends your business walks through the test. Every question where a competitor’s page is cited and yours is not is a page to write. Buyer questions change, so repeat the list every few months. To see how search and AI engines read your current site before you start, run our free site check.

Run Pixerry's free site check to see how Google and AI assistants like ChatGPT read your pages before you write the next one

Check my site
[ Faq ]

Frequently asked questions

Not always. OpenAI’s help page lets you force a web search by typing / in the message box and selecting Search. You can also regenerate an existing answer with web results by choosing Try again or Search the web, when that option is available.

It can. OpenAI says ChatGPT collects a general location from your IP address and may use it when rewriting the query, so “restaurants near me” can become “top restaurants San Francisco”. It does not share the IP address itself or your account details with search providers. For a local business, this is one more reason to have pages that name the city you serve.

No official source we read says so. Google states that it ignores llms.txt files, so they neither help nor harm rankings in Google Search. The part of OpenAI’s search help page we read does not mention them. Clear pages that answer specific buyer questions are a safer use of your time.

Google says that, for Google Search, optimizing for generative AI search is still SEO. The difference we see in practice is focus: in our 20-question test, AI answers cited pages built for one exact question, not general service pages. The technical base stays the same: pages that are indexed, crawlable and readable.

Still have questions about this topic? Our team is happy to help you apply it to your project.

Get in touch