In December 2006, medical journals published 72 randomized trials whose primary outcome came back statistically null. A team led by Isabelle Boutron later read every report and tallied the places where the wording implied the treatment had worked anyway. In 42 of the 72 abstracts, the concluding sentences carried that implication, and half the conclusion sections in the main text did the same. The studies had cleared peer review at selective journals. The final paragraphs had quietly stopped reporting results and started defending them.
The conclusion is the last thing a reviewer reads and the thing a paper is remembered by. Reviewers spend the middle of a paper weighing evidence. At the end, they look for what the authors believe they demonstrated. When that belief runs ahead of the data, trust built over four pages of careful results gets spent in one paragraph.
Most guidance on how to write a good conclusion for a research paper treats it as a summary dashed off after the real work. A conclusion is better understood as the place where claims are matched against what the results actually paid for. Our earlier post on the introduction funnel described the promise a paper opens with. The conclusion is where that promise comes due.
Key takeaways
- A conclusion makes three moves: it restates the answer, scales the consequence to the evidence, and marks the limit where the finding’s authority stops.
- Spin appeared in 36 of 72 main-text conclusion sections in Boutron’s sample of null trials, and reviewers now check final paragraphs against the results for exactly that reason.
- The discussion argues and the conclusion settles. New evidence at the end reads as an argument kept in reserve.
- Article conclusions commonly run one to two paragraphs, roughly 150 to 250 words, and open on the answer itself rather than an announcement that it is arriving.
- The margin test labels each sentence answer, consequence, or limit. Unlabeled sentences get cut, and a label with no sentence is a debt still owed.
What is a conclusion in a research paper?
It is the section that states what the study found, what that finding means, and where its authority stops. The reader arriving at it has stopped asking what was done. The question now is what is known that was not known before, and the conclusion answers it in the authors’ own voice. Elsevier’s author services frame the split between abstract and conclusion as a pair of questions. The abstract answers what. The conclusion answers what next. The abstract also sells the paper to readers who’ll never open it. The conclusion serves the ones who’ve just finished it and want the landing to be smooth.
The section takes different shapes. Many Nature-format papers carry no Conclusions heading, because the final discussion paragraph does the work, while clinical journals often require a structured standalone section and a thesis expects a chapter.
The container varies. The moves inside it do not.
How is a conclusion different from the discussion?
The discussion argues, and the conclusion settles. A discussion sets results beside prior studies, weighs rival explanations, admits anomalies, and concedes what the data cannot decide, which is why it carries most of the paper’s citations and hedges. A conclusion states what survived that argument.
AJE, the author-services group, frames its guide to the discussion section around interpretation against prior work, with the conclusion inheriting what’s left standing. The division looks like this in practice:
| Discussion | Conclusion | |
|---|---|---|
| Core question | What do these results mean against the existing literature? | What does the paper establish, and what should change because of it? |
| Evidence | Reinterprets results already reported, introduces no new data | Same rule, stricter: new evidence here reads as a structural error |
| Citations | Carries most of them | Sparse, and none the argument has not already needed |
| Length | Usually the longest interpretive section | A paragraph or two in an article, a chapter in a thesis |
| Failure mode | Wandering, circular, over-hedged | Overclaiming, or a pasted abstract |
Plenty of journals merge the two sections, and Elsevier’s guide to writing the discussion describes the concluding paragraph as the place to name what future work will confirm. That final paragraph inherits every rule in the table.
What are the parts of a good conclusion?
Conclusions that hold up share a small set of moves, and each one fails in a recognizable way when it is skipped. The labels used here, answer, consequence, and limit, are also the labels in the margin test below.
1. The restated answer
The opening sentence converts findings into a statement of what is now established. A results sentence reports that the treatment reduced readmissions by an adjusted 4 percent with a confidence interval crossing zero. An answer sentence says the trial found no reliable effect. The conversion is the whole job of the move, and it’s where Boutron’s spin lived. Her sample held 42 abstract conclusion sections asserting some form of benefit the primary outcome had not shown.
When the move is skipped, the conclusion opens on method or background, and the reader waits out a paragraph to learn what the paper claims. Sometimes the claim never arrives.
2. The consequence
The next move says who changes what. AJE distinguishes implications from recommendations, the impact results have on a field versus the actions those results support, and both belong to the final movement of a paper.
The discipline here is scale. A consequence that outruns its sample becomes the most quoted line in the paper and the fastest way to be quoted against it, because a four-week study of 30 volunteers can’t carry a sentence about transformed practice.
When the move is skipped, the conclusion states the finding twice and calls the repetition significance.
3. The limit and the open question
The last move marks where the authority of the finding stops, then names the question the study leaves open.
A stated limit reads as control. Reviewers treat an author who volunteers the boundary as an author who knows where it sits, and AJE makes the same case for the discussion in its guidance on writing limitations. The conclusion carries a one-sentence version of that honesty.
The failure mode is the apology tour, a list of confessions where one boundary sentence belongs. A conclusion that ends in apology leaves the reviewer’s final impression of the paper as its weakest reading.
How do you start a conclusion paragraph?
Strong conclusion paragraphs open on the answer itself. Announcing that a conclusion is arriving spends the first sentence telling the reader what the heading already said. The reader reaching that sentence has read the paper. The sentence owes them content.
Writers working in a second language often lean on stock openers, and formulaic starters can help while the content is still forming. The scaffold is for drafting.
A final version assembled entirely from those sentences reads as assembled rather than argued, and reviewers can tell.
There’s a mechanical reason the first sentence should carry substance. Conclusion sentences travel. They get quoted in citing papers and read out by AI assistants summarizing the study for someone who’ll never open the PDF. A conclusion that opens with its claim hands every one of those channels the right sentence to take.
Why the final paragraph is the one writers defend rather than report, we don’t fully know. Our guess is that it’s the sentence most likely to be read by strangers.
How long should a conclusion be?
There’s no universal limit, and standalone conclusions in articles commonly run one to two paragraphs. The length follows from the moves inside, each taking a sentence or two, which lands most conclusions between 150 and 250 words.
Thesis conclusions run longer, a chapter rather than a paragraph, because committees expect restated contributions and reflected methods. Elsevier’s guidance on abstracts versus conclusions covers both scales under one heading, which trips up writers moving between formats.
A thesis chapter pasted into an article conclusion is the clearest case of a conclusion doing another document’s job.
A conclusion running past half a page in an article usually contains a second discussion wearing the wrong heading.
The symptoms are citations appearing for the first time and sentences that begin “however, some studies.” Everything that argues belongs in the discussion proper.
What never goes in a conclusion?
New evidence, new arguments, new citations, statistics absent from the results section, and any sentence that improves the study’s outcome. That last category has a name and a measured prevalence.
Boutron called it spin, the use of reporting strategies that emphasize a benefit the primary outcome did not support, and her team found it in 36 of the 72 main-text conclusion sections they read. The strategies weren’t subtle, a secondary endpoint promoted to the headline or a subgroup that happened to win. Reviewers are trained on this literature now.
A conclusion sentence that improves the outcome is the sentence most likely to be checked against the results, and the mismatch costs more than the null result ever did.
New material at the end of a paper fails structurally rather than rhetorically. A reviewer meeting an argument in the final section reads it as an argument kept in reserve, then re-reads the results wondering what else was held back. The same mechanism makes pasted abstracts costly. Angel Borja’s step-by-step guide for Elsevier Connect names “repeating the abstract, or just listing experimental results” as the standard conclusion mistakes, because a conclusion facing backward duplicates a section the reader already passed while the actual job goes undone. A reader who has seen the numbers once does not need them twice.
One softer failure fills the same space. A conclusion that ends on more research being needed has hedged into nothing, an observation that is always true and never informative. The sentence reads as caution. Reviewers read it as having nothing to land.
What does a good conclusion look like on the page?
A constructed example, four sentences, each labeled:
Taken together, the two experiments indicate that retrieval practice, and not simple re-exposure, drove the retention gain. [answer] Students who tested themselves recalled 18 percent more drug-interaction rules after one week than students who reread the same passages, and the gap held at one month. [the number the answer stands on] For a course built around rereading, that points to a specific substitution in one course rather than a redesign of instruction. [consequence, scaled] What the study does not show is whether the advantage survives a semester of cumulative testing, which is the open question. [limit]
The margin test from our post on abstracts works here with three labels. A printed conclusion gets answer, consequence, or limit written beside each sentence, and the mismatches tell the story. Sentences with no label are cut, or moved into the discussion if they argue. Labels with no sentence are debts still owed.
The test takes five minutes, on paper.
Part of the check can be automated. Fukuro, the review tool this blog belongs to, flags conclusion sentences whose claims no reported result supports, which is the spin pattern in mechanical form. It can’t judge whether a consequence is proportionate to its sample, it doesn’t replace a reader who has followed the argument, and it says nothing about acceptance. The sentences it flags are the ones a reviewer will check first.
The conclusion is where a paper’s claims finally meet their budget. A good one spends exactly what the results earned, and stops.
References and useful policies
- Boutron I, Dutton S, Ravaud P, Altman DG. Reporting and interpretation of randomized controlled trials with statistically nonsignificant results for primary outcomes. JAMA. 2010;303(20):2058-2064.
- Elsevier Author Services. The difference between abstract and conclusion.
- AJE. How to write the discussion section of a research paper.
- Elsevier Author Services. 6 steps to write an excellent discussion in your manuscript.
- AJE. Implications or recommendations in research: what is the difference?
- AJE. How to write limitations of the study.
- Borja A. 11 steps to structuring a science paper editors will take seriously. Elsevier Connect; 2021.