The short answer
The best AI prompts for research state the decision behind the question, name what counts as good evidence, ask for dates and page references, and tell the model to say "not found" rather than guess. Use Perplexity, from Perplexity AI, for finding cited sources (its pricing page describes the free plan as good for limited daily usage), Google's Gemini Notebook for reading your own documents, and Claude or ChatGPT for synthesis and checking. The 15 prompts below cover each stage.
How these research prompts are built
Most weak research with AI starts with a weak question. Typing a topic into ChatGPT, Claude, Gemini or Perplexity gets you a confident overview; asking for a scoped answer, with named source types and a date range, gets you something you can put in front of a manager. Every prompt on this page does four things: it states the purpose and the audience, it tells the model what counts as good evidence, it asks for dates and references so you can check the output, and it gives the model permission to say "not found" instead of filling gaps. That last instruction matters more than any clever wording, because the most expensive research error is a plausible sentence with no source behind it.
The 15 prompts follow the order a real research job runs in: scope the question, find and vet sources, extract what matters from long documents, synthesise, and then check your own conclusion. Copy them from the boxes below and replace everything in [square brackets].
Prompts for scoping a research question
Use Turn a vague brief into research questions when someone hands you a topic rather than a question. Asking the model to question you first stops it from guessing at the decision behind the brief. Separate what I know from what I assume is the most underrated prompt here: it turns your notes into a to-do list and shows that half of what you "know" is untested. Build a search plan before you search gives you queries in the vocabulary professionals in the field actually use, which is how you reach specialist sources instead of generic explainers. Any general assistant handles this stage well, including the free plans of ChatGPT, Claude and Gemini, because no live search is needed yet.
Prompts for finding and vetting sources
This is where tool choice matters most. Run Find primary sources only and Trace a statistic to its origin in a tool that searches the web and cites every claim. Perplexity is built for this: every answer lists numbered sources, and Perplexity's pricing page describes the free plan as good for limited daily usage. ChatGPT, Claude and Gemini can all search the web too, but you will often need to ask for links explicitly. Grade a source before you rely on it works in any assistant, because you paste the source in yourself. For academic literature, Elicit searches more than 138 million papers on its free Basic plan, which beats a general web search when you need studies rather than news. Our pick of the best AI for research compares these options in detail.
Prompts for reading and extracting from documents
When the evidence is already in a pile of PDFs, the job changes from finding to reading. Extract the parts of a long report that matter, Compare several documents in one table and Pull quotes you can cite all ask for page numbers or section references, so you can open the original and check a point in seconds. Gemini Notebook, which Google renamed from NotebookLM on 16 July 2026, is the strongest free choice here: its answers are grounded only in the sources you add, with citations, and the free Standard plan allows 50 sources per notebook. Claude and ChatGPT are better when you want the extract turned straight into a draft, and both accept uploads on their free plans within free-plan usage limits. See our guide to the best AI for summarising PDFs and the step-by-step PDF summary guide.
Prompts for synthesising what you found
Lay out the evidence on both sides is the antidote to research that only confirms what you wanted. Write a one-page synthesis memo asks the model to mark its own inferences, which is the single most useful habit to build into any AI-written research: readers can see which sentences rest on a source and which rest on the model's reasoning. Find where sources contradict each other often surfaces the most interesting finding in a project, because disagreements usually come from different definitions or different years, and spotting that is real analysis. For synthesis, Claude and ChatGPT tend to write the clearest prose. If the work is large, the deep research modes help: ChatGPT includes a limited allowance of deep research on its free plan, Gemini offers Deep Research without a subscription, Claude's Research feature starts on the Pro plan, and Perplexity's Deep Research comes with Pro.
Prompts for checking your conclusions
The last three prompts protect your reputation. Red-team my conclusion asks for the objections a sceptical reviewer would raise. Check a draft claim by claim turns your draft into a numbered list of facts and shows which ones are unsupported. Decide what would change my mind tells you which checks deserve your last hour. A useful trick is to run the red-team prompt in a different assistant from the one that wrote the draft, because a second model has different blind spots. Perplexity's Model Council does this inside one tool by running a question through several models. Perplexity's help centre lists it in Search for the Max and Enterprise Max plans (its pricing page also lists it under Pro, so check in the app), and inside Perplexity Computer it is on all paid plans and uses credits. Our guide on how to check if AI is telling the truth goes further.
How to adapt these prompts to your own work
- Name the decision, not just the topic. "Research heat pumps" gets an encyclopaedia entry. "Research whether our 40 UK offices should switch to heat pumps before the 2027 budget" gets an answer.
- Set a date range. Add "published since [month year]" whenever prices, laws, products or market data are involved. Ask for the publication date of every source so stale pages stand out.
- Define good evidence for your field. A lawyer wants statutes and judgments, a marketer wants first-party data and platform documentation, a scientist wants peer-reviewed studies. Say so in the prompt.
- Keep the "say so if you cannot find it" line. Removing it is the fastest way to get invented filler.
- Save your best version. Once a prompt works, store it as a project instruction in ChatGPT or Claude, a skill in Gemini, or inside a Perplexity Project, so your whole team starts from the same standard.
- Keep confidential details out of consumer plans. ChatGPT, Claude and Perplexity personal plans can use your chats for training unless you switch it off, and Gemini does while Keep Activity is on, so check the privacy section of each tool guide before pasting client material.
Which AI tool works best for research, and why
Pick Perplexity for the finding and fact-checking stages, because citations are built into every answer rather than added on request. Pick Gemini Notebook when the evidence is a set of documents you already have, because it will not wander off to the open web. Pick Claude or ChatGPT for synthesis, memos and red-teaming, because they write and reason over long material best. Pick Elicit for academic literature. A practical stack for most professionals is Perplexity plus one general assistant; our tool stacks show this combination and others, and Gemini Notebook vs Perplexity and Perplexity vs Elicit settle the most common choices. If Perplexity is your main tool, our Perplexity prompts for research use its specific features.
Prompts to copy
Turn a vague brief into research questions
I have been asked to research [topic] for [audience, e.g. our leadership team] who need to decide [decision] by [date]. Before any searching, turn this into: one main research question, four to six sub-questions that together answer it, the type of evidence each sub-question needs (official data, primary documents, expert analysis or customer evidence), and what I can safely leave out. Ask me up to three questions first if the brief is ambiguous.
You get: A scoped research plan with sub-questions and the evidence each one needs
Separate what I know from what I assume
Here is what I currently believe about [topic]: [paste notes or bullet points]. Sort each statement into three groups: known (and how I would prove it), assumed (and what evidence would confirm or kill it) and unknown (and where I would look first). Flag any statement that sounds like a fact but is really an opinion.
You get: A known, assumed and unknown list that shows where the real research effort belongs
Build a search plan before you search
I am researching [question]. Write a search plan: 8 to 10 precise search queries (including the exact phrases and synonyms professionals in [field] would use), the five kinds of sites most likely to hold primary evidence (for example regulators, company filings, standards bodies, peer-reviewed journals), and three kinds of sources I should distrust for this topic and why.
You get: Ready-to-run queries plus a short list of where good evidence lives
Find primary sources only
Find primary sources on [topic] published since [month year]: official documents, original data, regulator or company announcements and peer-reviewed studies. For each, give the title, publisher, publication date, a one-sentence description of what it shows and the link. Leave out blogs, listicles and summaries of other articles. If you find fewer than five, say so instead of padding the list.
You get: A dated list of original sources with no secondhand commentary
Grade a source before you rely on it
Assess this source for use in a professional report: [paste link or text]. Tell me who published it, who funded or benefits from it, when it was published, whether it cites its own evidence, and whether other independent sources support its main claim. End with a grade (strong, usable with caution or do not use) and one sentence explaining the grade.
You get: A quick credibility check with a clear use or do-not-use verdict
Trace a statistic to its origin
This statistic keeps appearing: '[paste statistic]'. Trace it back to the original study, survey or dataset. Tell me who produced it, the year, the sample size or method if stated, and whether the way it is usually quoted matches what the original says. If you cannot find the origin, say that plainly.
You get: The original source of a number, and whether it is being quoted correctly
Extract the parts of a long report that matter
I have attached [report name, number of pages]. I need it for [purpose]. Extract: the five findings most relevant to [my question], each with the page or section number; any numbers I would want to quote, with their exact wording; the authors' stated limitations; and anything the report says that contradicts [our current view]. Quote rather than paraphrase where the wording matters.
You get: A focused extract with page references, so you can check each point in seconds
Compare several documents in one table
I have attached [number] documents about [topic]. Build a table with one row per document and these columns: [criteria, e.g. scope, method, main finding, date, sample]. Use 'not stated' when a document does not cover a column. Below the table, list where the documents agree, where they disagree and which one is most authoritative for my purpose.
You get: A side-by-side comparison table with agreements and conflicts spelled out
Pull quotes you can cite
From the attached [document or transcript], pull every passage that speaks directly to [specific question]. Give each quote word for word, with the page, timestamp or section, and one line on why it is relevant. Do not paraphrase inside the quotation marks and do not include passages that are only loosely related.
You get: Exact, referenced quotes ready to drop into a report
Lay out the evidence on both sides
Using the sources I have provided [or: reputable sources from the last 12 months], set out the strongest evidence for and against [claim or proposal]. For each point give the source, its date and how strong the evidence is (single study, several studies, official data). Finish with what remains genuinely uncertain and what evidence would settle it.
You get: A balanced evidence view that shows its own gaps
Write a one-page synthesis memo
Write a one-page memo for [audience] answering [research question], based only on the sources and notes I have shared. Structure: the answer in two sentences, the three findings that support it (each with a source), the main counter-argument and why it does or does not change the answer, and what we still do not know. Keep it under 400 words and mark any sentence that is your inference rather than a sourced fact.
You get: A decision-ready memo that separates evidence from inference
Find where sources contradict each other
Read these sources: [list or attach]. Identify every point where two or more of them disagree on a fact, a number or a conclusion. For each conflict, quote both versions, give their dates and explain the most likely reason for the difference (different definitions, different years, different samples or an error).
You get: A list of conflicts between sources with the likely reason for each
Red-team my conclusion
My conclusion is: [paste conclusion]. My evidence is: [paste key points]. Act as a sceptical expert reviewer in [field]. Give me the five strongest objections, the alternative explanation I have not ruled out, and any evidence that is weaker than I am treating it. Be direct; I would rather hear it from you than from my boss.
You get: The objections a tough reviewer would raise, before anyone else does
Check a draft claim by claim
Here is a draft [report, slide or article]: [paste text]. List every factual claim in it as a numbered list. For each one, say whether it is supported by the sources I attached, contradicted by them or not covered, and quote the supporting line where there is one. Do not judge the writing, only the facts.
You get: A claim-by-claim audit showing which statements still need a source
Decide what would change my mind
I am about to recommend [recommendation] based on research into [topic]. Tell me the three findings that, if they turned out to be wrong, would most change this recommendation, how likely each is to be wrong, and the fastest way to check each one in under an hour.
You get: A short list of the checks that protect your recommendation the most