An AI-Assisted Research Workflow: Primary Sources and Citation Checking

Good research is mostly patient, unglamorous work: finding the right sources, reading them properly, comparing what they say and being honest about what remains uncertain. It is tempting to think that AI makes this unnecessary. A "deep research" tool can produce a polished, cited report in minutes. But a polished report with citations is not the same as a reliable one, and if you build your work on it without checking, you inherit its mistakes.

This article describes a workflow in which AI makes research faster and clearer, and you remain the researcher.

• An eight-stage research workflow with AI's role in each stage.
• How to search effectively and evaluate sources.
• How to build and use a source matrix.
• How to treat deep-research and search-grounded tools.
• How to check citations and write up honestly.

Vague questions produce vague research. "Is AI good for education?" is a topic. "In university programming courses, what does published evidence say about the effect of AI coding assistants on student learning?" is a question you can research. Write down:

• The question, in one sentence.
• Why you need the answer, and what you will do with it.
• The scope: population, place, time period.
• What would count as evidence.

Ask the assistant to challenge your framing: "What assumptions does my question make? How could it be narrower or clearer? What related questions am I ignoring?" It is a good thinking partner at this stage because nothing it says needs to be true. It only needs to help you think.

Before searching, you need the vocabulary. Ask for orientation, and treat it as provisional:

I am new to research on [topic]. Give me: (1) a plain-language overview in 150words, (2) the main terms and synonyms researchers use, (3) the main positions ordebates, (4) the kinds of sources that exist (journals, government reports, datasets).Mark any statement you are not confident about. Do not give citations.

The "no citations" instruction avoids invented references. What you want is search terms and a rough map. You will confirm the facts later against sources. Read the overview with some scepticism: it reflects typical views and may skip recent developments.

Find sources through channels that return real, retrievable documents:

• Library databases and scholarly search engines for research papers and books.
• Official sites for laws, statistics, standards and reports (government bodies, regulators, international organisations, product documentation).
• Original datasets where available.
• Reputable news and specialist publications for context and recent events.

Search effectively:

• Use several terms and synonyms that you found in stage 2.
• Combine terms precisely, using quotation marks for exact phrases and filters for date range.
• Check the references of a good source, which lead to more good sources. This "citation chasing" works backwards (what did it cite?) and forwards (who cites it?).
• Note the date of every source, and the date you accessed it.

Several categories of tool now combine language models with web search or with your own documents:

• Search-grounded assistants produce an answer with links to the pages they used. Perplexity's documentation, for example, describes web-grounded answers with built-in citations.
• Deep-research modes in some assistants browse many pages and write a longer report. Google's Workspace announcement of January 2026, for instance, listed Deep Research among the Gemini app's features.
• Source-grounded notebooks answer from the documents you add, as in the NotebookLM product that the same announcement describes as now usable as a source in the Gemini app.

These are valuable as lead generators: they find pages you would not have thought to search for and give you a first synthesis. But treat each output as a draft that points toward sources, not as a source.

• Open every cited link and read the passage. Does it say what the summary claims?
• Check the quality of what was cited. Tools may cite weak, outdated or promotional pages, or pages that are themselves low-quality or AI-generated.
• Look for what is missing. A cited report can leave out contrary evidence.
• Check the date. Browsing tools may rely on older cached pages.
• Reproduce key facts using a separate search and the primary source.

Anthropic's guidance on reducing hallucinations makes a related point about verifying with citations: have claims supported by quotes and retract any that cannot be supported. The same principle works for you as a researcher. No supporting passage, no claim.

For key sources, you read them. AI can make that easier without replacing it:

• Ask it to explain a difficult paragraph or term from a source you paste in.
• Ask for a structured summary of a source you provide: the question, the method, the sample, the main findings, the limitations the authors state. Then check these against the abstract, methods and results.
• Ask what the source does not say, or what claims go beyond its evidence.

• Abstract first, then methods and results. The abstract is a summary and can overstate. The methods tell you what was done.
• Sample and setting. How many people, who, where and when?
• Design. An experiment, an observational study or a review? What can that design show?
• Effect size and uncertainty, not only whether something was "significant".
• Peer review status. Preprints have not been peer reviewed. They can be valuable, and need extra caution.
• Limitations, which authors often state honestly near the end.

A source matrix puts what you found in one place, so you can see agreement and disagreement.

(The rows are illustrative.) Fill the table from your own reading. An assistant can help you format rough notes into the table, but the content must come from you and the sources.

For each source, ask:

• Who wrote it, and what is their expertise and incentive?
• When was it published, and is it still current?
• Evidence: does it show its data and methods, or only assert?
• Purpose: is it informing, selling or persuading?
• Independence: does it add something new, or repeat another source?

Be careful when many sources say the same thing. They may all trace back to one original, possibly flawed, claim that has been repeated until it feels established. Follow the chain: where did this figure first appear? Count the original, not the echoes. If the original is a press release or a single unpublished claim, treat the claim as weak however many outlets repeat it.

When your sources disagree, do not quietly choose the one you like. Work out why:

• Different definitions or measures (who counts as a "user"?).
• Different dates or populations.
• Different methods, or one is an error.
• Different purposes or incentives.

Then decide: present both, with the reason for the difference; weigh them by quality of evidence and say so; or state that the evidence is mixed. You can ask an assistant, "Here are the quotes from three sources. List the possible reasons they differ", and then test those reasons against the sources.

You form the conclusion, since it must reflect your judgment and your reading. Useful ways to involve AI:

• Challenge your argument: "Here is my argument and my evidence list. What is the strongest objection? Which of my claims is weakest?"
• Spot gaps: "What questions would an expert ask that my notes do not answer?"
• Edit your writing for clarity, under the rules of your institution.

Writing honestly means:

• Match your words to the evidence. One small study suggests; several consistent studies indicate; a substantial body of consistent evidence shows.
• Quote exactly, with quotation marks and a citation, or paraphrase fully in your own words and still cite.
• State limits: what you could not find, what is uncertain, where sources conflict.
• Disclose AI assistance as the rules or norms of your institution or publication require.

Before you finish, check each reference with the three tests from article 26: it exists, the details are right, and it supports the claim. Formatting tools and assistants can help turn verified details into a consistent style, such as APA or MLA. Give them the verified details from the source and ask for formatting only. Never ask an assistant to supply missing details such as the year or page numbers. It may guess.

A final pass: for every sentence in your draft that states a fact, can you point to the source and the page? If not, find one or remove the sentence.

Write a short note on limits: what you could not access, what the evidence cannot show, how sure you are, and what you would research next. This is good scholarship, and it protects readers from over-trusting your conclusion.

Suppose you are asked whether a shorter work week improves productivity.

1. Frame: "What does published evidence say about the effect of a four-day work week on worker productivity and wellbeing, for office-based employees?"
2. Map: the assistant gives terms such as "compressed work week", "reduced hours with no pay cut" and "pilot programme", and the debates (output measures versus perceived productivity).
3. Gather: you search library databases and official sites with those terms, and find reports on pilot programmes, a few academic papers and news coverage.
4. Read: you read the methods of the main studies. Many are pilots with volunteering employers, which limits what they show.
5. Compare: the matrix shows that most reports agree on wellbeing, differ on productivity measures, and that several news pieces echo one report.
6. Synthesise: your conclusion is cautious: evidence from pilots is encouraging but comes from self-selected organisations.
7. Verify: you check that each cited report says what you wrote.
8. Reflect: you note that you found little long-term evidence.

At each stage, the assistant saved you time on vocabulary, structure and clarity. It did not supply the evidence.

• Starting with an AI-written overview and treating it as the research.
• Using citations from an AI report without opening them.
• Counting echoes as independent sources.
• Relying on abstracts and summaries without reading the methods.
• Choosing the sources that agree with you.
• Letting the tool decide the structure of your argument.
• Skipping the limits section.
• Not disclosing AI use where required.

• Next: Privacy and Security: What Not to Paste, Connected Tools and Prompt Injection covers what to protect while you work.
• For using AI as a study partner on research material, see From Notes to Recall.
• For the integrity rules around citations, see Learning with AI Without Cheating Yourself.