Prompts that make AI cite its sources, and admit when it can't

Ask ChatGPT for the source of a statistic and it will usually give you one. Whether that source exists, and whether it says what the answer claims, is a separate question.

There's a reason for the confidence. OpenAI's own researchers argue that hallucinations persist because training and evaluation "reward guessing over acknowledging uncertainty" (Kalai et al., Why Language Models Hallucinate, September 2025). A model that says "I don't know" scores zero on most benchmarks. A model that guesses sometimes scores a point. So models learn to guess, fluently.

You can't retrain the model from a chat window. You can change the rules inside one conversation, though, and that's all a grounding prompt does: it makes "I don't know" an acceptable answer, and it makes every claim point at the text it came from.

One warning before the techniques. A citation is not proof. In a 2026 study of 14 models, the best ones produced working, relevant links more than 94% of the time, yet only 39–77% of the claims were actually accurate (Onweller et al., Cited but Not Verified, May 2026). A link that opens is the easy part.

Anthropic, OpenAI and Google publish their own prompting guides, and on grounding they say almost the same thing. When three competitors converge, it's worth listening.

1. Restrict the model to the sources you supply. Anthropic calls it "external knowledge restriction": tell Claude to use only the documents provided, not its general knowledge. OpenAI's sample instruction is "Only use the documents in the provided External Context to answer the User Query." Google's version: "rely only on the facts that are directly mentioned in that context."

2. Give it a way out. Each guide includes an explicit line for when the sources don't cover the question. Anthropic: "I don't have enough information to confidently assess this." OpenAI: "I don't have the information needed to answer that." Anthropic says giving Claude permission to admit uncertainty "can drastically reduce false information". Without the out, the model fills the gap.

3. Sources first, question last. For long inputs (Anthropic's threshold is around 20,000 tokens) put the documents at the top and your question at the end. Anthropic reports that queries at the end "can improve response quality by up to 30 percent in tests"; Google says to place instructions "at the very end of the prompt". OpenAI's older GPT-4.1 guide hedges: instructions at both the start and the end.

4. Quote first, then answer. Ask for the exact sentences that support the answer before the answer itself. Anthropic goes one step further: after drafting, the model should find a direct quote for each claim, and "if you can't find a supporting quote for a claim, remove that claim".

5. Let tools handle facts and maths. Google recommends switching on search grounding "whenever the model may need to know obscure or recent facts" and code execution "whenever the model needs to perform any kind of arithmetic, counting, or calculation". A model reciting a number from memory is the riskiest thing it does.

Even with all five, a citation is a claim about a source, not proof of it. Count connects to Claude and shows the reasoning chain behind each claim, so you can open the source and see whether it says what the answer says.

This combines all five. It works in Claude, ChatGPT and Gemini; the XML tags are Anthropic's recommendation, and Google's guide says XML tags or Markdown headings both work as long as you're consistent.

<documents>
<document index="1">
<source>[title, author, date]</source>
<document_content>[paste the text]</document_content>
</document>
[add more documents the same way]
</documents>

Answer the question at the end using only the documents above. Do not use your general knowledge.

First, copy the sentences from the documents that are relevant to the question, word for word, into <quotes> tags, numbered. If there are none, write "No relevant quotes found" and stop.

Then answer. After each claim, give the number of the quote that supports it, like this: [2]. If a claim has no supporting quote, leave it out.

If the documents don't contain enough information, say: "I don't have enough information to answer this from these sources."

Question: [your question]

Each line is there for a reason. The documents block keeps sources separate from instructions. "Do not use your general knowledge" closes the side door. The quotes step forces the model to find evidence before it commits to an answer, rather than finding evidence for an answer it has already chosen. The out stops it improvising.

When there are no documents, only the web

If you're relying on the model's web search, the shape changes but the principle doesn't:

Search the web to answer this. For every factual claim, give the URL and quote the sentence on that page that supports it. Prefer the primary source (the original report, filing or dataset) over articles about it. If you can't find a page that states something, say so rather than answering from memory.

Question: [your question]

OpenAI's GPT-5.5 guide adds two lines worth borrowing: "Use the minimum evidence sufficient to answer correctly, cite it precisely, then stop", and "Absence of evidence shouldn't automatically become a factual 'no.'" The second one matters more than it looks. A model that can't find something will happily tell you it doesn't exist.

The core prompt travels well. The details don't, quite.

If you build on the Claude API, use the Citations feature rather than asking for citations in the prompt. Anthropic says it is "significantly more likely to cite the most relevant quotes from documents than purely prompt-based approaches", and it can't point at text that isn't in your documents. For more on Claude specifically, see writing a system prompt for Claude.

The studies on AI citations make uncomfortable reading. When Columbia's Tow Center asked eight AI search tools to identify where 200 news excerpts came from, ChatGPT Search got 134 wrong and declined to answer none of them (Tow Center, March 2025). A Salesforce study of deep-research tools found citation accuracy between 40% and 80%, with large fractions of statements unsupported by the very sources cited next to them (DeepTRACE, September 2025).

So treat each citation as something to check, not as the check itself. We've written about why the fluent, well-referenced answer is the one people check least in the hidden AI hallucination.

A two-minute check

1. Open three citations at random. Does the quoted sentence appear on the page, word for word?
2. Read one sentence either side. Does the source say what the claim says, or just mention the same topic?
3. Follow it one step back. Is this the original study, filing or dataset, or an article about it? Numbers drift with every retelling.
4. Ask the model: "Which of your claims are not directly supported by one of your quotes?" Then ask what it would say without them.
5. Run the same prompt twice. Anthropic suggests comparing repeated runs: "Inconsistencies across outputs could indicate hallucinations."

If that sounds like a lot of clicking for every answer, it is. That's the job Count does: connect it to Claude or ChatGPT, and it splits the answer into claims and labels each one as a citation you can open, an assumption, the model's own memory, or a hallucination, with the reasoning chain behind it.

Grounding prompts reduce errors. They don't remove them, and the vendors say so: Anthropic's own guide notes that these techniques "don't eliminate them entirely. Always validate critical information."

Three situations where a better prompt won't save you:

• Long agent runs. In the Cited but Not Verified study, factual accuracy fell by about 42% as the number of tool calls grew from 2 to 150. The more an agent searches, reads and summarises, the more room there is for a claim to drift from its source. Anthropic calls the underlying problem "context rot": recall falls as the context grows. More in context engineering.
• Calculations. A grounded quote doesn't make the arithmetic on top of it right. Use the model's code execution, and check the formula it chose, not just the result. See Claude vs ChatGPT for data analysis and Claude data analysis.
• Synthesis. Each sentence can be sourced while the conclusion that strings them together isn't. That's the hardest error to spot, because every individual citation checks out.

A good grounding prompt changes what the model is willing to say. Checking the claims is still your job.

How do I get ChatGPT to cite its sources?

Why does ChatGPT make up sources?

Is there an AI that provides real sources?

Can a prompt stop AI hallucinating?