How to cite AI-generated answers in academic research

When you cite an AI answer in research, you are citing evidence of what that assistant said at one moment, not a fact about the world. Record the assistant and model, the date and time, the exact prompt, any settings and whether it searched, then keep the full output in a copy you control. Cite the recorded output, and verify any factual claim against a primary source.

Why is citing an AI answer different from citing a book?

A book is the same for every reader. An AI answer is generated for one request and may not come out the same way again. If you ask a second time, the wording, the sources and sometimes the conclusions can differ. See why AI answers change every time you ask.

So the citation has a different job. It cannot say "go here and read the same thing". It has to say "this is what the assistant produced, and here is the preserved copy". That is why the record matters more than the format.

What should you record?

Treat each answer as an observation and capture what a reader would need to judge it.

• The assistant and model name, as the interface shows it, and a version if shown.
• The date and time of the request, with the time zone.
• The exact prompt, word for word, including any earlier messages in the chat.
• Settings you can see: whether search was on, any custom instructions, memory, location, language.
• The full output, not a summary or a screenshot cropped to the interesting part.
• The sources the assistant cited, with their addresses, and the dates you accessed them.
• How you collected it: the consumer app or an API, signed in or out.

Memory and custom instructions deserve special mention. A signed-in chat with memory on may shape its answer by what it knows about you. Say so, or use a clean session and say that instead. See how to test AI answers properly.

How do you preserve the answer?

Save it in a form that does not depend on a company's website staying the same.

1. Copy the text into a file with the metadata above.
2. Keep a screenshot as supporting evidence, never as the only record.
3. Use a share link carefully. Links can expire, change or be removed. Treat them as a convenience, not an archive.
4. Deposit it in a repository your institution supports if the answers are central to your work, so that it has a stable identifier.

Good practice is to keep the raw output unedited and put your own commentary in a separate file.

How do the major style guides handle this?

Several large style guides have published guidance for generative AI, and the details have been changing. In general they agree on a few points: credit the tool or its developer, name the model or version, give the date, and describe the prompt, either in the text or in a note or appendix. Some treat the output as personal communication-like material that readers cannot retrieve, others ask for a link where one exists.

Because the rules move, we do not reproduce a format here. Check the current edition of the guide your institution requires, and follow your supervisor or journal if they have their own policy.

When should you cite the AI, and when the underlying source?

Use two different citations for two different claims.

• "The assistant said X" is a claim about the output. Cite the preserved answer.
• "X is true" is a claim about the world. Cite the primary source that shows it. If the assistant gave a source, open it and read it. Do not cite a source you have not read. Assistants can cite pages that do not support the claim, and can describe a page inaccurately.

Using an AI answer as the only support for a factual claim is weak practice in any field.

What if the research is about AI answers themselves?

Then the answers are your data, and a single citation is not enough. You need a method.

• Fix the questions before you look at results, and publish them.
• Ask more than once, and report how many times.
• Record the same fields for every answer.
• Describe how you decided what counted as a recommendation or a mention.

This is the territory of reproducibility in AI search research and AI answers as a dataset. Citewake publishes its methodology and the answers behind each edition, so that researchers can cite the data and check it. The editions page lists what is available.

How should you describe the method in your write-up?

Put a short, plain methods paragraph in the body or an appendix. Say which assistants you used, how you accessed them, when you asked, how many times, and what you did with answers that failed or refused. State whether search was on, whether you were signed in, and whether any location was set. List the exclusion rules, and give the number of answers excluded. Attach the question list and the raw outputs as supplementary material if the journal allows it. A reader should be able to see exactly what you did and decide how far to trust it. That short paragraph does more for credibility than any particular citation format.

What are common mistakes?

• Citing a screenshot alone. It cannot be searched, checked or reproduced.
• Leaving out the prompt. Without it, the answer cannot be interpreted.
• Paraphrasing and presenting it as a quote. If you changed it, say so.
• Implying stability. Avoid wording that suggests the assistant "says" something permanently.
• Forgetting to name the settings. Search on versus off can change the answer.

A short methods note, covering how you collected answers and what you excluded, answers most of what a reviewer will ask. For a first look at how assistants describe a website, you can also try the Citewake checker.