CiteOwl

Blog/Writing your paper

How to write a research paper (step by step)

The CiteOwl Team· Updated 20 September 2026· 32 min read

Nine steps, and the first five happen before you write a paragraph. Papers fail before the keyboard. The ones that struggle began drafting on a topic the size of a textbook, with sources nobody had opened. Revision does not repair that. Reading is where the paper is written, so the first hour goes on narrowing the question, whether you have two weeks or one night. Then open every source before you cite it.

Step 1. Understand the assignment and its requirements

Before anything else, read the brief twice and the rubric once. You are looking for the constraints that will shape every later decision: the required length, the citation style (APA, MLA, Harvard, Chicago, or something your department prefers), the minimum number of sources, whether they must be peer-reviewed, and the due date. Note whether the question is assigned or whether you choose it. If a marking rubric exists, it tells you exactly what earns points, which is the most honest map of what your professor cares about.

Is this a research paper or a term paper? If your assignment is the end-of-course paper tied to one class, the method is the same, but the prompt, the rubric, the reading list, the page band and the marker are all decided for you. How to write a term paper covers those constraints, with published rubric weights and what a late penalty costs.

Two requirements quietly decide how you work. The citation style decides how you record sources from day one, so pick it now and stay consistent. The source count tells you how much reading to plan for. If anything is ambiguous, ask your instructor early; a two-line email now is cheaper than a rewrite later. University writing centers like Purdue OWL are a good reference for the format your style requires.

Step 2. Choose and narrow a topic into an original research question

If the topic is yours to pick, choose something you can stay interested in for two weeks, because you will be reading about it for a while. Then narrow it, which is the step that saves the most time. A broad topic forces you to read everything; a narrow one lets you read a manageable, relevant slice and actually say something specific.

The move is to turn a topic into a question that is yours rather than a summary of what everyone already knows. "Ocean plastic pollution" is a topic the size of a textbook. "How concentrated are riverine plastic emissions in a small number of rivers, and why do published estimates disagree?" is a question you can answer in a paper. Five tests tell you whether a draft is ready, and the one it fails names the repair.

• Specific. You can name what is being measured or examined, in what setting. An "and" joining two separate inquiries means you have two questions.
• Researchable. Evidence, not values, would settle it. "Should cities ban cars from their centres?" is a debate. "Has the car ban changed footfall on the streets it covers?" is a question.
• Open. A reasonable person could expect a different answer than you do. If it resolves in one lookup, there is no paper in it.
• Feasible. Answerable in the length and the time you actually have. Ambition is where most projects lose their first fortnight.
• Grounded. Published work already engages with it. This is the one students assume rather than check, and checking takes two minutes.

The two-minute literature dipstick

Before you commit to a question, search its exact phrases in a free academic index and read the number of results as a signal about the question itself. Not the papers, the count. It tells you in seconds whether you are holding a topic, a workable question, or a phrase the field does not use. We used OpenAlex, which is free, needs no account and has an open API, searching titles and abstracts for the phrases as phrases. Every count below was run on 11 August 2026 and anyone can rerun it, either as a request to api.openalex.org/works with a title and abstract filter or as the same phrases typed into the search box. Google Scholar works too, with the caveat that its counts are rounded estimates over far more grey material.

The bands are our judgement, calibrated on the searches below, not a rule from anywhere. The last row saves the most pain. Students read a near-empty result set as evidence they have found something original, and in an undergraduate paper it almost never is. It means you wrote "buying behaviour of young people" where the literature says "consumer socialisation", and the fix is fifteen minutes with the vocabulary of the field.

One topic, narrowed

Assigned topic: ocean plastic pollution. Four drafts, with the test each one fails.

Draft 4 earns its space because of what it hands you. A question whose published answers disagree gives the paper its structure for free: here is estimate A, here is estimate B, here is what the two models assume differently, here is what would settle it. You are not required to resolve the disagreement. You are required to explain it, which is a paper an undergraduate can genuinely write.

Every edit between those drafts was one of five moves. When a question is stuck, run down the list and try each one on it.

• Name the population or the setting. Which people, which organisations, which places, which period.
• Replace the vague verb with a measurable one. "Affect" and "impact" hide the decision about what you are counting. "Lower daytime air temperature" cannot.
• Swap "whether" for "how much" or "under what conditions". Yes-or-no questions are usually answered already. Magnitude and conditions usually are not.
• Surface the ambiguity instead of resolving it privately. If a key term has three meanings in the field, the question is which meaning changes the answer.
• Find the disagreement. Two credible published answers to one question is the best gift a topic can give you, and turning one up is half of what the dipstick is for.

This is a fine place to use AI as a brainstorming partner: ask it to break your broad topic into five narrower sub-questions, then pick one and put it through the dipstick yourself. Just do not ask it for sources yet. Naming the question is your job, and so is everything that follows it.

Step 3. Find and vet real sources

Now you go looking for evidence. Start with academic databases rather than a general web search: your university library catalog, Google Scholar, and open indexes like OpenAlex surface peer-reviewed work instead of blog posts. Use the keywords from your research question, plus synonyms and author names, and follow the reference lists of the best papers you find to reach the sources they relied on.

Then vet what you find. Skim each result against your question and a simple quality bar: is it peer-reviewed, recent enough, and actually relevant? A useful rule of thumb is to favor the last 5 to 10 years and reach further back only for foundational papers that newer work keeps citing. Discard aggressively. A focused paper built on twelve strong sources beats a sprawling one padded with thirty weak ones, and how many sources a paper needs is decided by your brief and your question rather than by a round number. Where to look, and how to judge what comes back, are both in how to find sources for a research paper.

Start that vetting by checking what a result even is, because a results list is not a list of papers. On 11 August 2026 we put one narrow query, urban heat island tree canopy, through the open OpenAlex index, limited it to work published since 2021, and grouped the 430 hits by document type (the query is public, so you can re-run it). Journal articles came to 259 of them, 60 percent. The other 40 percent was 41 conference abstracts, 30 preprints, 27 conference papers, 20 dissertations, 17 book chapters, 12 datasets and a tail of editorials and reports. Every one of those can be worth citing, and none of them is the peer-reviewed journal article your sentence will silently imply if you cite it like one. Read the type before you read the abstract.

One non-negotiable: confirm every source is real before you trust it, especially anything an AI tool hands you. A general chatbot will happily invent a reference that looks perfect, with a believable author, journal, and DOI, that points to nothing. Search the exact title in Google Scholar, paste the DOI after https://doi.org/ to confirm it resolves to that paper, and check the lead author exists. We lay out the full method in how to check if a citation is real.

Step 4. Read and take notes, building an evidence base

Reading is where the paper is really written; the drafting just transcribes it. But you do not read a paper the way you read a chapter, and the arithmetic is why.

On 11 August 2026 we took the full text of 80 open-access research articles from PLOS ONE, indexed under the subject terms Climate change and Transportation and published since January 2024, and counted the words in every section. The median article ran 5,912 words of body text, with a 238-word abstract, 5 figures, 3 tables and 52 references. At the 200 words a minute this blog uses to estimate its own read times, reading that median paper straight through is half an hour before you have understood anything, and technical prose is slower than blog prose. Twelve sources at that rate is six hours before a word of your own is written. Selective reading is not laziness, it is the only way the numbers work.

So decide which of three jobs you are doing before you open anything. Screening, meaning is this relevant at all, runs three minutes. Extracting, meaning what did it find and can I describe it accurately, runs fifteen to twenty. Trusting, meaning should my argument lean on this, runs thirty to forty-five, because it is the one where you have to read the methods. Those budgets are ours rather than a rule, but the ratio holds: the mistake that eats a week is doing the third job on papers that needed the first, and the mistake that costs marks is doing the first and writing as though you had done the second, which is citing a paper you never read.

The idea of reading in passes is S. Keshav's, from a two-page note, How to Read a Paper (ACM SIGCOMM Computer Communication Review, 2007), that nearly every reading guide online descends from. His budgets were set for a graduate student reviewing conference submissions: a first pass of "about five to ten minutes", a second of "up to an hour", a third of "about four or five hours for beginners". Run those against a reading list and the problem is plain, since twelve second passes is twelve hours before a word is written, which is why the budgets above are shorter and the first job is a screen.

The three-minute screen

Read four things, in this order, and then decide. Do not scroll to the top and start.

• The title and abstract (60 seconds). The median abstract in our sample was 238 words, under 4 percent of the paper, and it carries the question, the approach and the headline result. If it does not touch your question, you are done.
• The figure captions (60 seconds). Not the figures, the captions. In an empirical paper the findings live in the figures, and the caption states what each one shows.
• The last paragraph of the discussion (30 seconds). Where the authors say what they think it all means, in plainer language than anywhere else in the paper, usually with the hedges the abstract left out.
• The limitations (30 seconds). Usually the paragraph before the conclusion, and the highest-value paragraph per word in the whole paper, because it tells you exactly what you can and cannot cite this paper for.

Then discard it, file it as background you will cite once, or promote it to a real read, and write that decision beside the reference. A paper you screened and did not record is a paper you will screen again in a fortnight. On a thirty-paper reading list, expect roughly twenty discards and background files, eight worth an extraction and two worth the trust pass, and read the most recent review first if the list has one, because it tells you which of the other twenty-nine matter.

When you need one fact from a paper

This is the most common reading job in student work and no guide covers it. You have written a sentence, or you are about to, and you need the finding that supports it. Four moves, about four minutes.

• Search, do not read. Open the PDF and search for the unit, the number or the distinctive term, and search the tables as well as the prose. The median paper in our sample carried three tables, and a specific value sits in one more often than in a sentence.
• Read the sentence before the one you found. The qualification that makes a number honest is almost always in front of it: the sample it applies to, the conditions, the model it came out of. Take that sentence with you or you will cite a general claim the paper never made.
• Check the paper owns the fact. If the sentence carries a citation of its own, the paper is repeating someone else's result. Follow it and cite the original, because citing a paper for something it cited is how errors propagate and it is trivially caught by anyone checking.
• Prefer the discussion's wording to the abstract's. Abstracts overstate, discussions hedge, and the hedged version is the one the evidence supports.

Done properly that takes four minutes and produces a citation you can defend. Done from memory a week later it produces the most common citation failure there is, a real paper attached to a claim it does not quite make.

Notes that survive to the draft

Keep notes claim-shaped, not paper-shaped. Instead of "Smith 2021 summary," write the specific finding you might use ("Smith 2021 found heavy users reported lower grade averages, n=400") under the theme it belongs to, and group notes by theme as you collect them, so the structure of your argument appears before you outline. One rule covers the rest: no note exists without its source and its location attached. Author, year, DOI, page or figure, and quotation marks around anything you copied verbatim, put there at the moment you copy it, because in three weeks you will not remember which words were yours. A reference manager such as Zotero keeps the paper and the note together and exports a formatted bibliography at the end; a plain document works if the rule holds. When you write, turning a note into a cited sentence should be transcription, not detective work.

AI can help you summarize a paper you have actually downloaded so you triage faster, but read the parts you plan to cite yourself. You cannot defend a sentence built on a summary you never checked.

Step 5. Outline: a working thesis and a section map

With evidence in hand, write a working thesis statement: one or two sentences stating your specific, arguable answer to your research question. It is "working" because you will refine it once the draft tests it, but having a target now means every section has a job. A thesis is not a topic; it takes a position. "Using social media while studying is linked to lower grades among undergraduates, though the effect depends on what the student is using it for" gives the whole paper something to prove. If yours feels soft, we cover what makes a claim defensible in how to write a thesis statement.

Then build a section map: the headings in order, with a one-line note under each saying what it argues and which sources support it. For a standard paper that is usually introduction, a few body sections grouped by theme or sub-argument, and a conclusion; empirical papers often follow introduction, methods, results, discussion, and the methodology section is the one most students underdescribe. Order the body so each section sets up the next, building toward your thesis rather than listing facts. A good outline is the moment the paper stops being scary, because the hard thinking is done and what remains is filling in prose.

Put a word count beside every heading

Of the 80 papers we measured, 58 carried all four IMRaD headings separately, and those 58 say where the words go. The medians: introduction 13.9 percent of the body, methods 25.9, results 20.6, discussion 23.2, conclusion 3.6. Round them and you have a budget to write beside each heading before you draft a word. The shares do not reach 100 because about a tenth of a paper goes on sections IMRaD does not name, which the last row keeps honest.

Keeping that number beside the heading while you draft converts the vaguest question in writing, "is this section long enough?", into arithmetic, and it tells you something before you start. 315 words of results is three or four paragraphs, which is two findings well explained, not seven listed. If your outline has seven findings and a 1,500-word limit, the outline is wrong, and finding that out now costs a minute instead of a weekend.

Three of those numbers are worth more than the table. The introduction is small, under a fifth of the body in 50 of the 58 papers, and student papers routinely run double that, because the introduction is the section you can write from general reading before you have any results. Methods beat results, in 33 of the 58, which inverts what most students assume about where the work goes. And the conclusion is tiny, a median of 190 words; if yours is the size of a body section, it is doing work that belongs in the discussion.

Nothing here is tight, and two caveats keep the table honest. The middle half of those papers put results anywhere between a sixth and nearly a third of the body. PLOS ONE also tells authors "manuscripts can be any length" (PLOS ONE submission guidelines), so read the methods share as the length methods reaches when nothing stops it; a journal with a 3,000-word cap squeezes it first. An argued paper with no methods or results redistributes that 47 percent across its body sections, and the introduction at roughly a seventh and the conclusion at well under a tenth hold in every discipline.

The shape bends too. Of those 80 papers, 11 ran results and discussion under one heading, 17 had no separate results heading at all, 18 carried a literature review section IMRaD does not name, and 67 ended with a conclusion it does not name either. The structure everyone teaches is a median, not a rule, and your brief beats it every time.

If we had one extra hour to hand out across the nine steps, we would put it here and take it back from the introduction. An hour spent deciding what each body section argues removes rework everywhere downstream. An hour spent on the introduction before the body exists is speculative by construction, and most of what you polish early gets rewritten to match what you actually proved. Outline slowly, introduce late.

What each line has to carry

A heading with a topic under it is not an outline, it is a promise to decide later. Each line carries four things: the claim as a sentence rather than a noun phrase, the specific finding or figure you will use as evidence, the real source you have opened with its author, year and DOI written down, and the word budget from the table above. A section with no number attached is the one that eats the paper. Where a line is still missing its source, write the gap note instead of a plausible citation. "Needs a recent study that separates background scrolling from coursework use, not the 2010 correlation paper" is an outline you can hand back to yourself on Thursday. An invented reference is a section that collapses the first time anyone checks it.

Here is what that looks like end to end on the social media example. The outline carries a working thesis and a section per sub-argument, and one of those sections becomes a body paragraph where a claim leads and a real, checkable source backs it.

Notice the order inside the paragraph: the point, then the evidence, then the citation that ties the claim to the source, then a sentence pointing forward. Every body paragraph in the paper repeats that shape, and every citation like the one above has to resolve to a real paper you opened. That one does: it is a genuine 2010 study in Computers in Human Behavior, not a reference a chatbot supplied.

Step 6. Draft section by section

Draft in the order that is easiest, not the order the reader sees. Start with the body, because that is where your evidence lives and you already know what each section argues. Write each body paragraph claim-first: open with the point the paragraph makes, then bring the evidence that supports it, then explain what it means for your thesis. Leading with the claim keeps you from stacking quotes and hoping an argument emerges; the structure forces one.

Write the introduction once the body exists, so it can honestly preview what you actually argued: a little context, your research question, and your thesis. That funnel from context to gap to thesis has a reliable shape, which we lay out in how to write a research paper introduction. Write the conclusion to answer the question and say why it matters, without smuggling in new evidence. And write the abstract last, after everything else is settled, because an abstract is a summary of a finished paper and you cannot summarize what you have not written; here is how to write an abstract when you reach it. Do not polish while drafting. A rough complete draft is worth more than three perfect paragraphs, because you cannot revise a paper that does not exist yet.

Step 7. Cite as you write, in one style

Add each citation the moment you use a source, not in a panicked pass the night before. Citing as you go means you never lose track of which sentence rests on which paper, and your reference list assembles itself. Pick one style and apply it consistently: APA and Harvard are common in the sciences and social sciences, MLA in the humanities, Chicago in history and some other fields. Your assignment brief or department decides; when in doubt, ask.

Every citation has to be real and traceable. A citation is a promise that a specific source says the thing you attached to it, and the reader can follow the reference and find it. That promise breaks in two ways: a source that does not exist, and a real source cited for something it never said. Both are caught the same way, by opening the source and confirming it backs your sentence. If you are using AI anywhere in the paper, this is the step where it can hurt you most, because chatbots fabricate references that look flawless. The fix is structural: use tools that retrieve and read real papers before writing a claim, so a citation points at something the tool actually found. That order is the whole reason AI makes up citations in ordinary chatbots and why a source-first tool does not.

Step 8. Revise in passes

Revising and proofreading are different jobs, and doing them together does both badly. Work in passes, biggest problems first.

Structure and argument

Read the whole draft for the argument alone. Does each section earn its place and move toward the thesis? Is every major claim supported by evidence, and is any evidence sitting there unexplained? This is the pass where you cut a section that does not belong, move a paragraph, or strengthen your thesis to match what you actually proved. Changes here are expensive later, so make them now.

Clarity and flow

Now read for the sentences. Tighten bloated phrasing, break run-ons, and make sure each paragraph flows into the next. Cut filler and hedging. Read a section aloud; your ear catches clumsy rhythm your eye skims past. This is also where AI is genuinely useful as an editor: ask it to flag unclear sentences or suggest tighter phrasing, then accept only the changes that keep your voice. You own the words.

Proofread

Last, hunt for surface errors: spelling, grammar, punctuation, and consistent formatting of headings and citations. Proofread when you are rested, ideally a day after the clarity pass, because tired eyes invent corrections and miss real mistakes. Reading backwards, sentence by sentence, breaks the momentum that makes you skim.

Step 9. Final checks before you submit

Before the paper leaves your hands, run a short checklist against the requirements from Step 1. Confirm the formatting matches the assigned style down to the title page, margins, and headings. Confirm the reference list is complete, that every in-text citation has a matching entry and every entry is cited in the text. Then do the check almost nobody does: open the reference list and confirm every source resolves, that each title, author, and DOI points to the paper you claim. One fabricated or garbled reference can put the whole paper under suspicion, and this five-minute pass is the cheapest insurance you will ever buy. A clean export to your required format is the final step; getting it right is the difference between a finished paper and a returned one.

A realistic timeline

For a standard course paper, give yourself one to two weeks of working time, not one heroic night. If yours is a term paper, the clock is set by the syllabus and the late penalty is published, and both are covered in how to write a term paper. A rough split: a day or two to settle the question and gather sources, two to four days to read and take notes, a day to outline, two to three days to draft, and two to three days to revise and proofread, with at least one day of rest between drafting and revising. The order is deliberate: start the source-gathering early, because finding and vetting real papers is the part you genuinely cannot rush. A thesis or capstone scales the same shape up to weeks or months. Whatever your deadline, work backwards from it and protect the revision days, because the last pass is where a decent paper becomes a good one.

When it is due tomorrow

Sometimes the timeline above is not available and the paper is due in the morning. The steps do not change, but the order of what you protect does, and the first hour is not spent writing. It is spent making the paper smaller. Pull the three numbers that bound everything, length, citation style and minimum source count, then narrow until the question points at a specific, well-studied claim rather than a field. "The effects of social media" is a week of reading. "Does time on social media during study sessions track with lower grades among undergraduates?" is answerable by morning, because it names a population, a variable and an outcome, which is what makes a search return the right papers instead of a thousand near misses. Then apply the one filter this night allows: if ten minutes of searching does not surface real papers on an angle, it is the wrong angle for tonight. Findability beats originality when the deadline is in the morning.

The constraint nobody warns you about is that finding a paper and being able to read it are different things, and at 11pm an interlibrary loan request is worthless to you. On 11 August 2026 we ran three narrow questions through the open OpenAlex index, restricted to journal articles published since 2021, searching titles and abstracts, and counted how many results have a free full text on record.

Plan for about a third of your results being unreachable tonight, and note that the share moves by topic rather than staying put. Two caveats: a free full text may be a repository copy or a preprint rather than the published version, and your university login will get you into some of the paywalled rest, which is why the library proxy is worth setting up before you need it. Anyone can re-run this: the query is the public OpenAlex works endpoint with an open-access filter, and the count comes back in the response. Practically, that means sorting for reachability early rather than late. Open every candidate in a tab in the first pass and close the ones that dead-end at a paywall immediately, before you have built an outline that depends on one. If you are stuck on a source you genuinely need, the legal routes are in the paper I need is behind a paywall.

Papers that fail overnight rarely fail at the keyboard. They fail because four hours went into hunting sources and the draft started at 2am with nothing left for the last pass. Fixed blocks stop that, weighted toward the literature:

• Find and triage (60 to 90 min). Four to six real, reachable papers. Stop at six. Check you can open each one before it joins the pile.
• Skim for claims (45 to 60 min). Abstract, then the last paragraph of the introduction and the conclusion, then the figures. Stop as soon as you have the one or two findings you will use. Write each one in your own words with the source attached, so you are not reconstructing the link at 4am.
• Outline (20 to 30 min). Working thesis, then one line per section saying what it argues and which paper carries it.
• Draft (the bulk of the night). Body first, claim first, citing as you write.
• Verify and revise (60 min). One structure pass, one clarity pass, then open every reference and confirm it resolves.

Set a timer per block and move when it ends. A finished rough section at 3am is worth more than a perfect one with three sections missing. The one rule that does not bend under time pressure is that you read the part of any paper you intend to cite: an abstract tells you what a study concluded, not what it measured, and the gap between those two is where a rushed paper picks up a claim it cannot defend.

An essay with no literature to gather is a different night with a different shape, and that one is how to write an essay fast.

The most common mistakes

Most lost marks trace back to a short list of avoidable errors:

• Starting to draft before narrowing the question. A vague topic produces a vague paper. Narrow first.
• Citing sources you have not read. If you cannot say what a paper argues, you cannot cite it safely.
• Trusting AI citations without checking. Chatbots invent references that look real. Verify every one.
• Summarizing instead of arguing. Listing what each source says is a report, not a paper. Make a point and use sources to support it.
• Organising the body by source. A paragraph per paper is a reading list wearing an outline's clothes. Group by claim and bring the sources to it.
• Leaving citation formatting for the end. Cite as you write, in one style, so the reference list builds itself.
• Revising and proofreading in one pass. Fix the argument first, the sentences second, the typos last.
• No time for revision. A first draft is never the paper. Protect the days that turn it into one.

Where a tool carries the slow middle

You can do every step above by hand, and plenty of strong papers are written exactly that way. Where a tool earns its place is the slow, error-prone middle of the nine steps: finding real sources, keeping citations honest, and getting from an outline to a draft without staring at a blank page. CiteOwl is built for that, on a verify-first principle. It searches actual literature and reads what it finds, so the papers behind your citations are real, not generated from memory. It can take your outline to a first cited draft, and every change it makes arrives as a reviewable diff you accept or reject, with version history if you want to step back. When the paper is done, it exports to your required format.

The line it never crosses is doing the thinking for you. The argument, the judgement, and the final words stay yours; the tool just makes doing it properly faster, with sources you can check. If you want to see how it compares to the other options, we ranked them in the best AI tools for academic writing, and if your next assignment is a survey of the field, here is how to write a literature review with AI.