# How does ChatGPT choose sources? Evidence and limits

> Understand ChatGPT source selection through OpenAI documentation and original research, with a practical worksheet for checking claims, citations, and evidence.

Published 2026-09-08. Updated 2026-09-08. Written and reviewed by Shubham Bansal, founder of The Cited Club. Reading time: 11 min read.

Ask ChatGPT to help you choose a business, and it may search the web before answering. It looks for information related to your question, puts an answer together, and can include links to the pages it used. Those links are citations. OpenAI says its search aims to find relevant, reliable information, but it does not reveal exactly how every page gets chosen.

My starting point is practical: what have we published that deserves to be used in the answer? If the page offers another vague claim about expertise, there is already useful work to do. I would fix that before paying for another explanation of the algorithm.

## TL;DR

- ChatGPT does not search the web for every answer. Check whether it searched before drawing conclusions about which websites it chose.
- When it searches, it can look up several related questions to help answer yours.
- A link to your website and a recommendation to buy from you are different wins. Check which one you received.
- Open the links. The page should actually back up what the answer says.
- Start with information your customers need: what you do, who you help, what you deliver, and why they should believe you.

## ChatGPT may search for more than the words you typed

[OpenAI explains](https://help.openai.com/en/articles/9237897-chatgpt-search) that ChatGPT can turn your question into more specific searches, then search again as it learns more. Your location and relevant things it remembers from earlier conversations can affect those searches. Answers can include links, with more available through the Sources button.

Imagine asking for an accountant who understands a growing online shop. ChatGPT could need information about accountants, ecommerce experience, and your location. This is an illustration, but it explains why typing the same question into Google does not necessarily show you the pages ChatGPT used. Save the question and the answer when checking your own business.

Nobody outside OpenAI can point to a complete published recipe for getting picked. I would question anyone selling a guaranteed score, number of links, or page format that supposedly makes it happen every time.

For the implementation checklist, read [How to Get Cited by ChatGPT](/guides/how-to-get-cited-by-chatgpt). This guide explains how to interpret the evidence behind a source choice, including what remains unknown.

## Finding your page is only the beginning

Here is a simple way to follow what happens. It draws on the public explanations and research, rather than access to ChatGPT's inner workings. Some answers involve several searches; others involve none.

| Step | What happens | What to check for your business |
|---|---|---|
| Someone asks a question | Earlier conversation can affect what they mean | What was asked, when, and whether search was used |
| ChatGPT looks for information | It finds pages that could help | Can people and search tools access your pages? |
| It writes an answer | Information from those pages can shape the response | Does it describe your business correctly? |
| It adds links | The reader can open supporting pages | Does each linked page back up the statement? |
| The reader considers a business | The answer may suggest who to contact | Are you recommended for work you actually do? |

If your business is missing, this gives you places to investigate. Perhaps the page cannot be accessed, or perhaps it does not answer the question well. Sometimes the answer alone will not tell you why you were left out.

Start with a weakness you can actually see and fix. A competitor appearing once is not enough to explain exactly why it was chosen.

## What original research tells us about citations

### AI can learn to look things up before answering

[The WebGPT researchers](https://arxiv.org/abs/2112.09332) trained an earlier AI model to browse websites and collect references before answering. People demonstrated the task and gave feedback on the answers. References helped them check whether the answers were correct.

WebGPT demonstrates how an AI can look things up and collect evidence. It is not a behind-the-scenes guide to today's ChatGPT.

The lesson I take from it is to make your claims easy to check. Say what you know, show why you believe it, and explain where it stops being true. Then check whether ChatGPT uses the page.

### An answer with links can still be wrong

[Gao and colleagues' research](https://arxiv.org/abs/2305.14627) asked two useful questions: do the linked pages back up the answer, and are enough of its statements backed up? They found problems even in answers that included references.

You can apply that idea without a research team. Open a link and find the sentence that backs up the answer. Does it say the same thing, or has ChatGPT stretched it? This is a practical check inspired by the paper, not a new study of ChatGPT's accuracy.

### Changing a page can change how an AI answer uses it

The [original GEO paper by Aggarwal and colleagues](https://arxiv.org/abs/2311.09735) tested changes to content and checked how it appeared in AI answers. Different changes worked differently across topics. That is a reason to test what helps your own pages, rather than assume one formula works everywhere.

Use the research to decide what to try. If a page makes a claim without backing it up, add the missing evidence. Adding a quotation or statistic is only worthwhile if it helps answer the customer's question.

## If ChatGPT can read my website, will it recommend me?

No. Letting ChatGPT read your page gives it access to the information. It still has to choose what to use in its answer.

[OpenAI's advice to website owners](https://help.openai.com/en/articles/12627856-publishers-and-developers-faq) is to allow OAI-SearchBot, its automated website reader, to access content for search summaries. GPTBot serves a different purpose: potential model training. Your website team should check these settings separately.

Ask your website team to check that search tools can read the public page. Your hosting settings can block access even when the page works for you. Content behind a sign-in is another reason the tool may not see what you see.

Then read the page itself. Is it clear, current, and relevant to the question? Making a page accessible fixes an access problem. It does not make the business the best answer.

## Does ChatGPT prefer original research?

Original research gives an answer something it cannot get from another summary. That is a good reason to publish it. It is not proof that ChatGPT always prefers research pages.

A valuable study explains what was examined, how the observations were collected, and which conclusions the method supports. A page with a large number and an unexplained sample is difficult to evaluate, even if its headline calls it research.

Imagine you publish a study of how long it takes customers to get started. Does the clock begin when they sign the contract or when your team starts work? Did you include the difficult projects? Is the reported time an average? Readers need those answers before they can judge whether the result applies to them.

Those details do more work than saying you get customers started quickly. Another writer can judge the comparison and decide whether to reference it. That is the value I want from research, whether or not ChatGPT cites it next week.

Our [Vendor Bias evidence review](/research/vendor-bias-study) explains a correction to our own work: combining percentages that measure different things creates a misleading comparison. Calling a page research does not excuse weak methods.

## Why a citation can still misrepresent the business

### The cited source supports a narrower statement

A page might say you work in one country while ChatGPT says you work worldwide. The link is real; the description is wrong. Save the specific mistake.

If your page leaves this unclear, fix it. If the page is already precise, record the mistake in the answer. Do not change correct website copy to match an incorrect response.

### The source is old or describes a different product

An old service package or a different company with a similar name can lead to a misleading description. Check which business and time period the source describes. Opening an old page today does not make its information current.

On your website, make replacements and material changes easy to understand. Explain what is current and what has been retired. Preserve corrections where an old claim had enough reach that readers still need to understand what changed.

### A source citation is mistaken for an endorsement

An answer may cite a company's educational page while recommending another provider. It may also recommend a business using a third-party source. Record the cited domain and recommended company separately.

Our [AEO vs GEO guide](/guides/aeo-vs-geo) provides a classification table for these outcomes. This distinction matters when a report is being used to decide what to publish next: a missing recommendation can require different evidence from a missing educational citation.

## A source-review worksheet you can reuse

Copy these fields into a document. If an answer makes several important claims, check each one against its link. One correct sentence should not hide a mistake elsewhere.

| Field | What to record |
|---|---|
| Question | The exact submitted wording |
| Where and when | Which app you used, the date, relevant location, and whether it searched the web |
| What ChatGPT said | Copy the exact statement you want to check |
| Source | Exact cited URL, page title, author if available, and publication or update date |
| What the page says | The text that backs up the answer, or a note that you could not find it |
| Does it match? | Yes, partly, no, or unable to check |
| What happened for your business? | Named, recommended, linked to, or left out |
| Next action | Fix your page, check another source, or save a mistake you cannot yet explain |

This is our proposed worksheet, not the result of a new experiment. Use it to turn a confusing answer into a specific publishing decision.

If you cannot open a source, mark it as unable to check. That does not mean the claim is false. An answer can also be correct but link to a page that does not back it up. Note both separately.

For later checks, keep the questions and setup as consistent as you can. If a monitoring tool collected an answer through a developer connection, or API, label it that way. Do not present it as an answer seen in the ChatGPT app without checking that the comparison holds.

Our [Surface Diet review](/research/surface-diet) explains why you need to know which tools were checked and what was counted. It is a guide to making a fair comparison, not a permanent list of ChatGPT's favourite websites.

## My view: a citation strategy needs a publishing decision

The question I want a source review to answer is simple: what should we change or publish because we did this work? A list of domains is only useful if someone can turn it into a decision.

That is central to how we are defining The Cited Club's service. Investigation leads to agreed work, publication, and repeated checks. The intelligence has to help us choose the next action. A founder should be able to see what the team is doing with the findings.

Take a service page that says the company is experienced, trusted, and customer-focused. I would ask the owner what a prospective customer still needs to know. Which problems do you handle? What do customers receive? When are you the wrong fit? What can you show to support the promise?

Those answers often require access to the person doing the work. A writer can organize the explanation, but the business has to supply the substance. That is why I see customer expertise as a necessary input to delivery, rather than a task we can replace with more keyword research.

Our own website work brought this into focus. We kept revisiting whether visitors could understand the service and the deliverables. Having a research section and a product demonstration did not automatically make the offer clear. We had to work on the explanation itself.

I would apply the same test to a page we want cited. What does it explain that a reader actually needs? Where does the explanation come from? Can someone inspect the evidence? If those answers are weak, the next job is already visible.

Original research is valuable when it resolves an uncertainty. An expert's account of a difficult decision can be valuable too. They should carry different labels, but both can give a reader something useful that a generic summary cannot. The aim is to contribute a reasoned answer worth referring to.

## Turn the review into a content brief

Start with a specific missing answer. "Improve AI visibility" gives a writer little direction. "Explain what the monthly service includes, using the current agreement and exclusions" gives them a useful job to do.

Ask an accountable subject expert to review the claim before publication. Keep the supporting reference close enough that a reader can check it without searching the entire article. Explain any limitations in ordinary language. Those choices make a page easier to cite responsibly, whether the next reader is a journalist, customer, or answer system.

After publishing, link to the page from a related service page or guide, then open the link to check it works. Save later ChatGPT answers, but do not assume your edit caused every change. Other websites and ChatGPT itself can change too.

If several gaps recur across commercially important questions, a [visibility audit](/audit) can help organize the diagnosis. The useful output is a prioritized evidence plan with owners and concrete page changes. A citation count becomes actionable when it points to work the business can actually do.

## Frequently asked questions

### Does ChatGPT always search the web?

No. Do not assume every response uses web search. Record the visible search state and source links before treating an answer as an observation of search behavior.

### Does ChatGPT publish its source-ranking formula?

OpenAI describes relevance, reliability, query rewriting, and access requirements, but does not publish a complete formula that predicts which page will be cited for a given question.

### Is a citation proof that the linked page supports the answer?

No. Open the page and compare the relevant passage with the answer's claim. Record partial support or an inability to verify rather than assuming every linked reference is sufficient.

### Can original research guarantee a ChatGPT citation?

No. Original research can provide useful, distinctive evidence, but selection is not guaranteed. Its value depends on the question, the method, and the clarity of the claims it supports.

### Is an API monitoring result the same as a ChatGPT result?

Not automatically. Record the product and collection method. An API observation should not be presented as a consumer ChatGPT observation without evidence that the comparison is valid.

## How to cite this guide

Bansal, Shubham. "How Does ChatGPT Choose Sources? Evidence and Limits." The Cited Club, September 8, 2026.

## Sources and references

- [OpenAI: Searching the web with ChatGPT](https://help.openai.com/en/articles/9237897-chatgpt-search)
- [OpenAI: Publishers and Developers FAQ](https://help.openai.com/en/articles/12627856-publishers-and-developers-faq)
- [Nakano et al., WebGPT: Browser-assisted question-answering with human feedback](https://arxiv.org/abs/2112.09332)
- [Gao et al., Enabling Large Language Models to Generate Text with Citations, EMNLP 2023](https://arxiv.org/abs/2305.14627)
- [Aggarwal et al., GEO: Generative Engine Optimization, KDD 2024](https://arxiv.org/abs/2311.09735)
- [The Cited Club: corrected Vendor Bias evidence review](/research/vendor-bias-study)
- [The Cited Club: corrected Surface Diet evidence review](/research/surface-diet)

## About the author

Founder of The Cited Club. Shubham leads the research and delivery behind its GEO, SEO, content, technical, and authority work. [About Shubham](https://thecitedclub.com/about.md#founder).

## Related reading

- [AEO vs GEO](https://thecitedclub.com/guides/aeo-vs-geo.md): Distinguish answer inclusion, citation, and recommendation.
- [How to get cited by ChatGPT](https://thecitedclub.com/guides/how-to-get-cited-by-chatgpt.md): Turn source analysis into practical page improvements.
- [The Surface Diet review](https://thecitedclub.com/research/surface-diet.md): Compare product observations with their context intact.

## Work with The Cited Club

- [Explore GEO and AI Search Growth Services](https://thecitedclub.com/services.md)
- [Run the free AI Visibility Audit](https://thecitedclub.com/audit.md)