Why do AI tools hallucinate citations?

Two failure modes. First, the model generates a plausible-sounding source for a claim it's already produced (citation-after-the-fact). Second, the model finds a real source but the source doesn't actually say what the model claims it says (the citation exists but doesn't support the claim). Both are common; both have specific prompt-level fixes.

Which tool is best for cited research?

Perplexity by a margin — it's designed citation-first. ChatGPT search and Gemini have improved a lot but still trail on claim-to-source accuracy. See our comparison post for the data.

How much of an AI-assisted research draft can I trust without checking?

None of it, for facts that matter. Treat AI research like a starting research assistant — fast at finding sources you'd have missed, unreliable at summarizing what they actually say. Always click through and read the source for any claim that goes in your final piece.

How to use AI for research without losing the citations

AI research has a discipline problem. The default workflow — "summarize what's out there on X" — gets you a fluent paragraph that's 80% right and 20% subtly wrong, and you can't tell which 20%. This guide is the discipline we use to keep AI-assisted research trustworthy: citation-first prompts, source verification habits, and the difference between asking for a synthesis and asking for actual quotes.

The default workflow that doesn't work

You open Perplexity (or ChatGPT, or Gemini), type "what's the current consensus on X", read the paragraph that comes back, and paste it into your draft. The paragraph cites sources. You assume the citations support the claims.

This fails in two specific ways:

The synthesis blends claims that come from different (sometimes conflicting) sources. The citation footnotes attach to particular sentences, but the synthesis-as-a-whole reads as if all sources agreed.
The model paraphrases too aggressively — the source said "X may contribute to Y under conditions Z", the summary says "X causes Y".

The fix is to ask differently.

The citation-first prompt

I'm researching <topic> for <purpose>. I want you to find sources,
not write a synthesis.

For each source you find:
- Quote the specific sentence or paragraph that supports a claim.
- Name the source (publication, author, date).
- Tell me which of the three following positions it takes: pro / con / mixed.

If you can't find a quote that directly supports a claim, say so.
Don't paraphrase to fill the gap.

Return the sources, not a summary.

This produces a different shape of output: quotes you can verify, attributed to sources you can click, sorted by position. You then write the synthesis yourself — which is the part where your judgment actually matters.

The verification habit

Even with citation-first prompts, verify. Two-minute discipline:

Click through to every source the AI cited.
Find the quoted text in the source. (Cmd-F.)
Read the sentence in context. Does it say what the AI claimed it said?

About 1 in 10 citations fail this check on hard queries even from Perplexity (which is the best tool for this). The failure rate is much higher when you use general chat tools (ChatGPT, Claude) for research.

When to use which tool

For deep research with citations: Perplexity by default. Free tier is real; Pro adds better models and longer answers.

For "while I'm at it" research mid-conversation: ChatGPT search (it lives in your chat, so the research feeds the broader work).

For quick factual lookup: Google AI Overviews. Don't trust them blindly, but for "what year did X release" they're usually right and fast.

For research that involves a long source document you already have: Claude 4.7 with the PDF attached. Use the file-attachment prompts to keep the citations honest.

Quotes vs syntheses

The cleanest rule we know: ask for quotes when you'll cite the result, ask for syntheses when you're orienting yourself.

Orienting ("what's the lay of the land?"): syntheses are fine. You won't quote them.
Writing ("I'm going to cite this in my piece"): always quotes, always verified.

The default chat workflow blurs this line. Naming the mode in your prompt protects you from the blur.

A worked example

Bad: "What does the research say about AI's effect on writing quality?"

Better: "I'm writing a piece arguing that AI-assisted writing improves first drafts but harms revision. I want to find sources that support, contradict, or complicate this claim. For each source, quote the specific finding (with sentence-level quotes), name the source, and tell me whether it supports / contradicts / complicates my argument. Don't synthesize. Don't paraphrase to fill gaps."

The "better" version produces a research scaffold you can actually use. The "bad" version produces a paragraph that sounds authoritative and isn't.

Where this fits

If you're combining AI research with the portable system prompt template, add "for research questions, find sources with quotes; don't synthesize unless I ask" to the constraints section. The discipline becomes default.

More practical workflows in Playbooks. Or try oran.chat free and switch between Perplexity-style research and other models in one conversation.