Paperpal or statistical review: what each one covers

You have the manuscript finished, budget for one tool rather than three, and two tabs open promising the same thing in the same two words: submission-ready. Neither tells you what it leaves out, which is precisely the fact you need in order to choose. This is not about which one is better —that question has no answer, because they do not answer the same thing— but about which question each one answers. I checked it on Paperpal's own site on 4 September 2026, page by page, because the two costliest mistakes when writing about a competitor are inventing a flaw and refusing to acknowledge a strength.

What Paperpal does, according to Paperpal

Paperpal presents itself as an academic writing assistant and it runs where you actually write: there is an add-in for Microsoft Word, a Google Docs extension, a Chrome extension, an Overleaf integration, and its own web editor. Its checker is trained on papers edited by professional academic editors rather than on general prose, and it shows in what it suggests: it preserves technical terminology instead of fixing it for you. Around that sit a paraphraser that also trims length, a reference finder with a citation generator, and a plagiarism check.

For submission it has two pieces. One is the Document Health Check, four checks together in a single upload: plagiarism, AI detector, grammar correction and reference check. The other is the set of pre-submission checks, more than thirty of them, which its site groups into six areas: language and grammar; references, where it says it evaluates missing citations, reference age and count, self-citation patterns, and citation-reference mismatches; structure and counts, verifying section completeness, word counts and heading hierarchy; figures and tables, with their callouts and captions; disclosures, meaning funding, ethics, conflict of interest and author contributions; and metadata.

That list is not empty marketing. If you are the second author who has been handed the job of making the manuscript presentable in a week, four of those six areas are literally your week: nobody enjoys checking by hand that Figure 3 is cited in the text, that the abstract stays under 250 words, and that the conflict of interest statement did not stay behind in the other draft. Having a machine do it in two minutes is a genuine improvement to your life, and saying otherwise would be dishonest.

And there is one more thing worth conceding without haggling: for anyone writing in English as a second language, a tool trained on academic edits is exactly the right tool. The editorial comment the English requires revision by a native speaker is one of the most frequent and most unfair rejections there is, because it says nothing about your science, and preventing it saves you a month of calendar. I know of no honest argument for withholding credit there.

Impeccable English protects you from nothing

Here is the misunderstanding that makes people choose badly. People assume a well-written manuscript is a well-conducted one, and that if you polish it enough it will get through. The results section does not work that way: it is the one part of a paper that can read straight through, without a single typo and with perfect transitions, and still be wrong. Prose is the wrapping around the numbers, and the wrapping tells you nothing about what is inside.

Assumptions. An ANOVA with a significant Levene test and badly unequal group sizes is poorly executed, not poorly written. The style checker will see a flawless sentence in a one-way ANOVA revealed a significant effect of condition and has no way of knowing that behind it sit 12, 41 and 37 people and variances that look nothing alike. A reviewer catches that by looking at your Table 1, and it usually arrives with the most expensive request in the process: run it again.

Power. If you compared two groups of 30 looking for a medium effect, your power was .48. That is: if the effect existed, you had less than an even chance of finding it, and to reach the conventional .80 you would have needed 64 per group. That is not a writing problem, it is a design that could not work, and no amount of work on the commas changes it. If you have never run that calculation, it is set out step by step in how to report a power analysis in APA 7.

References that do not exist. Ever since people started assembling bibliographies with help from a model, plausible entries turn up —two surnames that sound right for the field, the journal that publishes exactly that, a DOI with the correct shape— that correspond to no actual work. A fabricated reference is grammatically perfect by construction. It is the error that best survives any copy-editing pass, and the one that lands worst when a reviewer finds it.

Where they genuinely do overlap: the bibliography

If you have read this far waiting for the moment I say Paperpal does not look at references, it is not coming, because it would not be true. Its reference checker runs nine checks and its site tells you explicitly not to let retracted papers, broken links, or AI-hallucinated references reach reviewers, with a broken-link detection it describes as one that catches dead URLs and invalid DOIs to prevent fake, made-up references. That is exactly the problem from the previous paragraph, and it addresses it head on.

So if somebody tells you that only one particular tool detects fabricated references, that somebody has not looked at the competition in 2026 and probably does not want you to look either. On bibliography, the overlap between a language tool with a reference module and a statistical review is high. The difference is in the detail: which databases you compare against, and what you do when nothing comes back. If a reference cannot be located it may not exist, or the database may simply not index that Ibero-American journal, and those two conclusions are nothing alike.

The criterion that seems defensible to me, whoever applies it, is two independent sources: a reference is accepted when two different databases return the same work with the same authors, and it is marked not locatable —not false— when none of them finds it. That is the difference between a check you can show a co-author and an accusation. I set it out in checking your references before submitting, and it holds whichever tool you end up using.

The real line: commenting versus recomputing

Paperpal does not stay out of methods either, though not where you would look for it. What assesses methods is not the writing assistant: it is Preflight, a document assessment platform they sell to publishers and institutions that also offers pre-submission screening for authors. It holds several modules, and one of them, PeerPilot, is what their site describes as an AI-assisted, reviewer-style assessment to improve scientific presentation and review readiness, organized across four domains —Study Quality Design, Statistical Soundness, Reference & Citation Quality and Guideline Adherence— and one that highlights major strengths, weaknesses, and areas that warrant closer scrutiny. So yes, it does comment on your methods, and it does so in the shape of a reviewer.

And that is where the line runs, which is not where most comparisons draw it. A module like that reads your text and issues a judgment: the sample looks small for the design, the choice of test is not justified, you do not cite the reporting guideline your journal requires. All of that is useful, and sometimes it is exactly what you need to hear before submitting. But it is commentary: a model's opinion about what you wrote, as good as the model and never verifiable by you.

Recomputing is a different operation. If you write F(2, 87) = 3.42, p = .04, there is exactly one correct answer and it depends on nobody's opinion: the F distribution with 2 and 87 degrees of freedom returns p = .037, and your .04 was correctly rounded. If you write t(58) = 1.72, p = .04, the answer is equally unique: p = .091, and there you have a problem no amount of careful reading is going to settle by discussion.

That is the whole difference. On the Paperpal pages that describe its checks —consulted on 4 September 2026— there is nowhere any claim that it recomputes a p value from the degrees of freedom, or that it checks an effect size against the F or the t you yourself wrote. I do not say that as a reproach: it is not its job, just as mine is not deciding whether your English reads like a first-quartile journal article.

Three errors that only appear when somebody redoes the sum

The p value that does not match its degrees of freedom. The most common and the most trivial: you paste in the output of one run, you change the model, you update the F and you forget the p. In a manuscript with twenty tests there are twenty chances for it to happen once. When a reviewer catches it —and some reviewers run every manuscript through an automated checker— the doubt stops being about that line and becomes a doubt about all the others.

The effect size that does not follow from your own test statistic. With F(2, 87) = 3.42, the partial eta squared that corresponds to it is η²_p = .073. If your table says .21, then either the test statistic is not what you say it is or the effect was computed on a different subsample, and both explanations need telling before somebody else asks. The same goes for d: a t(58) = 2.11 with equal groups gives d = 0.55 and p = .039, and nothing else.

The reference that points to a different paper. Not the fabricated one, which is easier to catch, but the one carrying a real DOI belonging to another article because the reference manager merged two entries. It opens, it resolves, it throws no error, and it credits somebody with something they did not say. It only surfaces if somebody compares the title and authorship of your entry against what the database returns, one by one.

All three have something in common: they are found by comparing numbers with numbers, not by reading. And all three are still there after the manuscript has been through copy-editing, because copy-editing does not look at them. The full order of checks before you hit submit is in checking your analyses before submitting.

How to decide, and why it is often both

Look at your previous rejections, which is the evidence you already have and nobody uses. If the letters talk about English, about clarity, about an abstract that does not read, or about the manuscript not meeting the journal's format, your bottleneck is language and compliance: a tool like Paperpal solves that and you need nothing else. Paying for a statistical review to fix it is buying the wrong key.

If the letters talk about design, assumptions, sample size, or the analysis not answering the question, or if a reviewer has asked for a test you do not command, you do not have a writing problem however much it shows up in the writing. What is needed there is somebody redoing the sums. The specific reasons manuscripts fall at the first screen are in why Q1 journals reject you, and almost none of them are grammatical.

And if you have budget for both, the order matters more than it looks: numbers first, prose afterwards. Polishing the wording of a results section you will have to rewrite because the model changed is work thrown away twice, once when you pay for it and once when you delete it. When the statistics are final, then yes, have somebody make the English impeccable.

What makes no sense is treating them as rivals. They are two different trades that land on the same desk in the same week, and the industry has filed them under the same heading of AI tools for researchers because it is convenient for selling, not because they do the same thing.

Frequently asked questions

Does Paperpal check my statistics?

Partly, and not quite where you would look. The writing assistant does not; what goes into methods is Preflight, their document assessment platform for publishers and institutions, which also offers pre-submission screening to authors. Its PeerPilot module is, in their site's words, an AI-assisted, reviewer-style assessment, with four domains, one of them Statistical Soundness. That is reviewer-style commentary on your methods. What it does not say on any of its product pages, consulted on 4 September 2026, is that it recomputes a p value from the degrees of freedom or checks your effect size against the F or the t you reported. Those are two different things: one gives an opinion on your analysis, the other redoes its arithmetic.

How much does Paperpal cost?

I deliberately do not reproduce their prices here. They change, they vary by country —their site says it offers billing in local currencies— and an out-of-date price in an article like this is worse than no price at all, because it looks like a fact and is not one. Check their plans page on the day you decide, and compare what each tier includes rather than the figure alone: with these tools the difference between plans is usually in the monthly allowances more than in the features.

Will Previum fix my English?

No, and it should not try. It recomputes p values from the degrees of freedom with code, checks effect sizes against the F or the t in your own manuscript, and verifies references one by one against eight databases (Crossref, OpenAlex, Semantic Scholar, PubMed, DOAJ, arXiv and OpenLibrary). The prose of the report is written by a model, but around numbers it cannot change. If what you need is for your English to read like a published paper, that is a different tool and there is nothing wrong with using both.

Can I use both on the same manuscript?

Yes, and it is the sensible thing if the budget stretches. The order I recommend is numbers first, language second: close the analyses, check that every test statistic agrees with its degrees of freedom and that the bibliography exists, and only then pay to have the writing finished properly. The other way round you polish prose that may go in the bin if one test changes.

Does a report like this count as peer review?

No. Neither theirs nor mine. It is author self-assessment of your own manuscript before submission: a simulation of what you are going to be asked, commissioned by you and with no independence from you whatsoever. Peer review is organized by the journal, with reviewers you did not choose. Confusing the two in a cover letter is a quick way to start badly with an editor.

What exactly does the statistical report include?

There is a full one published so you do not have to take the description on trust: the sample report, running through a manuscript from start to finish, including the lines where the p value did not match its degrees of freedom and the references that could not be located. Reading it before deciding is more informative than any feature list, mine included.

The honest comparison between these tools is not about quality, it is about category: Paperpal answers "is this ready to be sent?" and answers it well, and a statistical review answers "do the numbers you wrote add up?", which is a question with a single, checkable answer. Choose by whichever rejection is currently hurting, not by whichever site argues better, and if one day either of them starts promising the other one's job as well, ask to see the example before you pay.