🎓 Biology as Poetry Reference Checker

Check Your References: Catch Errors, Fill In Missing DOIs, and Spot Citations That Don't Exist

by Professor Stephen T. Abedon Steve Abedon, Ph.D., has taught well over 100 individual classes and ~10 individual courses (including introductory biology, microbiology, physiology, biochemistry, etc.) over his 30+ years as a college professor. He has over 150 publications, including multiple monographs, and is ranked in the top 1 or 2 percent among researchers of all disciplines worldwide. His research focuses on bacteriophage biology and phage therapy.

biologyaspoetry.org  |  phage.org  |  phage-therapy.org  |  abedon.phage.org  |  google scholar | publications

What is this? A tool for checking your own reference list before you hand it in, submit it, or publish it. Paste the list in the Check tab, and each reference is looked up in the Crossref and OpenAlex scholarly databases and compared field by field with the closest record found: title, first author, year, journal, volume, and first page. That catches the ordinary mistakes that creep into any reference list, such as a wrong year, volume, page range, or DOI, including in citations downloaded from PubMed, Google Scholar, or a journal's website. The DOI of each matched work can be copied with one click, so missing DOIs are easy to fill in, and Find duplicates spots the same work listed twice. The Add DOIs tab goes further: give it a reference list, including a Word document, and it returns the same document with DOIs added, adding a DOI on its own only when exactly one record matches the reference in every detail.

The Library Check tab checks a whole reference-manager library at once. Load an export from Reference Manager, EndNote, Zotero, or another reference manager (RIS, MEDLINE, comma- or tab-delimited, or XML), and every record is compared with Crossref and OpenAlex, and with PubMed for records that have a PMID. The results are sorted into categories, such as details that differ, authors’ names spelled differently, DOIs that are missing, articles once ahead of print that are now published, and preprints with a journal version, along with possible duplicates and links to free full text. You get a downloadable report, a copy of your file with DOIs added, and a progress file so that a large check can be saved and picked up again later.

The Link Check tab does something related for links: paste a list of them, or a page’s HTML, and each is looked up to show what is actually there (the page’s title, description, and opening sentences), with Wikipedia disambiguation pages, redirects, missing pages and sections, and link text that does not match the page flagged.

The same checks also flag citations that may not exist at all, such as those invented by AI chatbots. The Practice tab has citations to investigate before seeing what the checker reports, and the By Hand tab explains how to check a reference without any tool, with lookups for DOIs, journals, authors, and terms.

🔒 Privacy: the references you check are sent from your browser to public scholarly databases to be looked up: Crossref, OpenAlex, DataCite, and doi.org; Google Books and Open Library for books; and PubMed for records with a PMID in the Library Check tab. Nothing is stored, and nothing is sent anywhere else. To share a particular section, use the 🔗 link beside its heading, which copies a web address that opens the page right there.
To cite this tool: Abedon, S.T. (2026). Biology as Poetry Reference Checker. https://references.biologyaspoetry.org/  

references.biologyaspoetry.org  ·  Abedon's Books

How can I improve this page?  contact: references@biologyaspoetry.org

✅ Check Your References

Check your own references first. Reference lists easily pick up mistakes: a wrong year, volume, page range, or DOI, a misspelled author, a preprint cited in place of the published paper. Citations downloaded from PubMed, Google Scholar, or a journal's website are not immune. Running your list through the checker before you hand it in, submit it, or publish it takes a minute, can catch many of these, and fills in missing DOIs along the way. It also finds the same work listed twice, with Find duplicates.

Paste references below, one per line or separated by blank lines, and select Check references. Each reference is looked up in the Crossref and OpenAlex databases, and any DOI it contains is checked to see which paper it actually belongs to. When a reference matches a database record, that record's DOI can be copied, so references missing a DOI, or carrying the wrong one, can be completed. The list below holds examples, so you can see every kind of result before pasting your own. The same checks flag citations that may not exist at all.

Or read them from a file or a web page: or Word (.docx), PDF, web pages (.html, or an address, Wikipedia’s included), OpenDocument (.odt), RTF, and plain text, one file or several, or a zip of them. The reference lists are found and their references put in the box above.

About the examples. The list mixes real and problem citations from the bacteriophage literature:
  1. A real article, no DOI given.
  2. A real article with its correct DOI.
  3. A real article, no DOI given.
  4. The same real article as #1, with its year, journal, volume, and pages changed.
  5. The real article from #2, given the DOI of a different paper.
  6. An invented article attributed to a real author.
  7. A real book, of the kind often missing from article databases.

🔗 Add DOIs to Your Reference List

Give this tab a reference list, and it gives the list back with DOIs added. A Word document comes back as the same Word document, with its formatting left alone and each DOI added at the end of its reference. A reference manager's export (RIS or BibTeX) comes back with each DOI in its DOI field, ready to import.

A wrong DOI is worse than no DOI, so three rules decide what is added:
  1. A DOI is added on its own only when exactly one database record matches the reference in title, first author, year, volume, and first page. Corrections, replies, and similar notices are never matched to the paper they concern.
  2. Anything less certain comes to you for a decision, with the candidate records side by side and each DOI linked so you can open it and confirm. Nothing in that group changes unless you choose.
  3. A DOI already in a reference is never removed. When it belongs to a different work and exactly one record matches the reference, a replacement is offered for your approval.
The tool cannot tell whether a reference supports the statement it is cited for, so it is no substitute for reading.
Each reference takes a second or two to look up; a list of 100 takes a few minutes.

🧪 Practice: Track Down Whether a Citation Is Real

Eight citations to classic and recent biology papers, one at a time. Some are exactly right, and some are not. Investigate each one with the same lookups the checker uses, decide what it is, and then see how you did, along with what the checker reports for the same citation.

Investigating first and checking second is the point. A tool that answers for you teaches you to trust the tool, while tracking a citation down yourself teaches you how to tell whether any reference is real.

The five kinds of citation

  • Real and correct: the work exists, and the citation describes it accurately.
  • Real, details wrong: the work exists, but the year, journal, volume, or pages are off.
  • Real, wrong DOI: the work exists, but the DOI given belongs to a different work.
  • Invented: no such work exists, even if the authors and journal do.
  • Real, hard to find: the work exists but is missing from the article databases, as many books are.

Investigate

  1. If the citation has a DOI, look it up and compare the record with the citation.
  2. Search for the title, then compare the closest record with the citation, field by field.
  3. For books and chapters, try WorldCat or Google Books, since the article databases often lack them.
DOI Open at doi.org
or open a search in
What is this citation?

🔎 Checking a Reference by Hand

Every check the tool makes can be done by hand in a few minutes, and doing it by hand catches things no tool can. Work through these steps in order, stopping once the reference is confirmed or clearly fails. The lookup boxes on this page do some of the legwork for you.

  1. Start with the DOI, if there is one. A DOI (digital object identifier) is a permanent address for a published work, and doi.org is the official service that looks DOIs up. Type https://doi.org/ followed by the DOI into a browser. The page that opens should be the work cited, with the same title and authors. A DOI that goes nowhere, or that opens a different paper, is a red flag.

    Asks doi.org whether the DOI is registered and where it leads, asks Crossref which publisher owns its prefix (the part before the slash), and looks up the record behind it.

  2. Search the exact title. Put the title in quotation marks in Google Scholar, or search it in PubMed for biomedical work. Real journal articles almost always turn up. If nothing does, try a few distinctive words from the title without the quotation marks, since citations are sometimes copied with small changes to the wording.
  3. Compare the details. Once you find the work, check the citation against it: author names and order, year, journal, volume, and pages. A mismatch can be an innocent typo, but several mismatches together suggest the citation was assembled from memory, copied carelessly, or invented around a real title. Do not assume a citation is right because it was downloaded: citations exported from PubMed, Google Scholar (its "Cite" button especially), journal websites, and reference-management software can carry errors too. Common ones are the online-first year given in place of the print year, an article number given as a page number, missing or misspelled authors, and the preprint cited in place of the published version. While you are there, check whether the work has been retracted: PubMed and the journal's own page show retraction notices, and the Retraction Watch Database lists them.
  4. Check that the journal and volume make sense. Does the journal exist? Was it publishing in that year? Does that volume number belong to that year? The NLM Catalog lists biomedical journals by full title and standard abbreviation, with the years each was published, and a journal's own website lists its back issues by year and volume.

    Searches Crossref's list of journals for the publisher, ISSN, and the years in which its articles have DOIs, with links to the NLM Catalog for abbreviated names.

  5. Check the authors. Do the named authors work in this field? Their publication lists on Google Scholar, ORCID, or PubMed show what they have actually written. Fabricated citations often attach a plausible title to real people who never wrote it.
  6. For books and chapters, use a library catalog. Books are thinly covered by article databases. Search WorldCat or your own library's catalog instead. For a chapter, search for the containing book as well as the chapter title, since catalogs often list only the book. Google Books can often show a book's table of contents, which confirms that a chapter, its authors, and its page range are real. For theses, try WorldCat, OATD, or the university's own repository. For older books and journals, HathiTrust and the Internet Archive have scanned many of them.
  7. Read the source. A citation can point to a real paper that does not say what the citation claims. Only reading the cited source catches this, and no reference checker can do it for you. You rarely need to read every word. Work from the outside in:
    1. The title. Is the paper even about the topic it is cited for?
    2. The abstract, if there is one. It states the main question, methods, and findings in a paragraph, and it is free on PubMed and on the journal's page even when the article is not.
    3. A summary. Many articles end the introduction by stating what was done, and end the discussion with a conclusion. Some journals add a highlights list, a plain-language summary, or a "significance" statement. In a review article, the section headings show what it covers.
    4. The figures and their legends. These often show the key results more directly than the text.
    5. Search inside the article. In a PDF or web page, use Find (Ctrl+F, or ⌘F on a Mac) to search for words from the claim: the organism, the gene, the method, or a number. Try a few different words, since the paper may phrase things differently from the citation.
    6. Check whose finding it is. If the passage you find is itself citing another paper, the claim belongs to that other paper, which is the one to cite.
    If the article is behind a paywall, look for a free copy through PubMed Central, Unpaywall, or the authors' own websites, or ask your library, which can usually get any article through interlibrary loan.
  8. Ask a librarian. Reference librarians track down obscure, old, and foreign-language sources every day, and they are usually glad to help.

Signs that deserve a second look, and how to tell

  • A title that matches the argument being made a little too perfectly.
    How to tell: search for the exact title in Google Scholar. Real titles turn up; invented ones usually do not.
  • Methods or ideas that did not exist in the year given, such as CRISPR gene editing in a paper from the 1980s.
    How to tell: find out when the method or term first appears in the literature. A PubMed search shows a results-by-year chart, the Wikipedia article on a method usually has a history section, and the box below counts papers by year.
  • A DOI whose prefix belongs to a different publisher or journal than the one named.
    How to tell: put the DOI in the DOI lookup in step 1, which names the publisher that owns the prefix. For example, 10.1084 belongs to Rockefeller University Press, which publishes the Journal of Experimental Medicine, so that prefix on a citation to PNAS would not fit. Journals do sometimes change publishers and keep their old DOIs, so a mismatch calls for a closer look rather than a conclusion.
  • Page numbers that are implausible for the journal, or an article number in a journal that uses page numbers.
    How to tell: look at a few other articles from the same journal and year, on the journal's website or in PubMed, to see what its page numbers or article numbers look like.
  • A list in which no citation includes a DOI or link, for recent work that would normally have one.
    How to tell: run the list through the Check tab, which finds the DOI of each work it can match.
  • A reference you were given by an AI tool that did not show you where it found the reference.
    How to tell: treat it as unconfirmed until you have found the work itself, by the steps above or with the Check tab.

🤖 Why Fake Citations Happen

How AI tools produce them

Chatbots built on large language models generate text by predicting what plausibly comes next. A citation has a very regular pattern: author names, a year, a title, a journal, and a set of numbers. A language model can produce text with that pattern fluently, drawing on the names, topics, and journals it has seen, without ever looking up whether the combination exists. Unless an AI tool actually searches for a source and links you to it, a citation it gives you is expected to be a reconstruction rather than a record, and it can be partly or entirely made up while looking completely ordinary.

How often this happens has been measured. Walters and Wilder (2023) asked two versions of ChatGPT to write short literature reviews and then checked the 636 citations produced. With GPT-3.5, 55% of the citations were to works that do not exist, compared with 18% for GPT-4. Among the citations to real works, 43% (GPT-3.5) and 24% (GPT-4) contained substantive errors, such as a wrong year, journal, or page range. Newer tools, especially ones that search the web, are expected to do better, but none should be assumed to be error-free.

Why it matters

A citation is a promise that the source exists and supports what is being said, so that a reader can go and check. A fabricated citation breaks that promise even when the statement it supports happens to be true. The consequences can be serious. In 2023, lawyers in the case Mata v. Avianca filed a brief citing court decisions that did not exist. The decisions had been generated by ChatGPT, and the court fined the lawyers.

People make citation errors too

Not every bad citation comes from AI, and not every bad citation is dishonest. People copy references from other reference lists without reading the source, mistype volume numbers, and misremember years. Even citations downloaded from PubMed, Google Scholar, or a journal's website, or produced by reference-management software, are not always right: an online-first year may stand in for the print year, an article number for a page range, or a preprint for the published paper. Those are errors, and they are worth fixing, but they are different from citing something that does not exist. When the checker flags a reference, the right first question is what went wrong, not who to blame.

Catching ordinary errors like these, in a reference list that is otherwise real, is likely to be the most common use of this tool. Run your own references through the Check tab before you hand them in, and use it to fill in missing DOIs while you are there.

What this means for you

  • List only sources you have actually seen. A reference list is a claim that you have looked at each work in it.
  • AI tools can be a reasonable place to start looking for sources, but treat every reference one gives you as unconfirmed until you have found the work itself.
  • When you copy a citation from another paper, check it against the original. Errors spread by copying.
  • Assume someone will check. Instructors, reviewers, editors, and librarians increasingly do, often with tools much like this one.
  • Check your own reference list with this tool, in the Check tab, before anyone else does.

Reference

Walters WH, Wilder EI. 2023. Fabrication and errors in the bibliographic citations generated by ChatGPT. Scientific Reports 13:14045. doi:10.1038/s41598-023-41032-5

📖 How the Checker Works

Steps

  1. The pasted list is split into individual references, and any DOI in each reference is pulled out.
  2. A cited DOI is looked up directly, first in Crossref, then in DataCite, then in OpenAlex. Crossref registers most journal DOIs, and DataCite registers DOIs for software, datasets, and archives such as Zenodo. DataCite is asked before OpenAlex because OpenAlex can list a particular version of an archived work under the work's all-versions DOI.
  3. The full reference is sent to Crossref's bibliographic search, which returns the closest matching records. If no record's title matches, a probable title is pulled from the reference and searched in OpenAlex.
  4. For a reference that looks like a book or a book chapter and is not found in the article databases, the book catalogs Google Books and Open Library are searched as well. A book counts as verified when its title, first author, and year agree. Book records do not list chapters, so for a chapter the book it appears in is looked for instead; finding it gives a possible match, not a verified one.
  5. Each candidate record is compared with the reference. The title check counts how many of the record's title words appear, in order, in the reference. Because references often give only a main title, leaving off a subtitle after a colon or dash, the record's main title alone also counts as a match when nearly all of the reference's own title words appear in the record's title. A main title of only one or two words must also have the right first author. The first author's surname, the year, the journal (in full or abbreviated), the volume, and the first page are each checked for presence in the reference.

What each result means

Verified
A record's title matches, and its first author, year, volume, and first page are consistent with the reference.
Details disagree
The title matches a real record, but other details do not. A real title attached to the wrong year, journal, or authors is a common signature of fabricated citations, and also of ordinary copying errors.
Author issues
Everything matches except the spelling of one or more authors' names, and the two spellings differ only in accents or letters outside the English alphabet (Genç and Genc, Müller and Mueller, Sørensen and Sorensen), or in a character lost from one of them (a “?” or “�”). It is very likely the same person, but one spelling is probably wrong. Every author in the reference is compared, not only the first. When the reference has accents that the record lacks (Delbrück where the record has Delbrueck), the reference is not at fault, and it stays Verified with a note. A first author whose surname is actually different counts as Details disagree.
Possible match
A record's title partly matches. Check by hand.
DOI points elsewhere
The cited DOI exists but belongs to a work with a different title.
DOI has no record
doi.org confirms that the DOI is registered, but none of the three databases has a record for it, so its title could not be compared. This happens with DOIs from smaller registration agencies, and with DOIs registered without details. Open the DOI to see what it belongs to.
DOI not found
The cited DOI was not found in Crossref, OpenAlex, or DataCite, and doi.org did not confirm that it is registered. A DOI that does not exist is a strong warning sign, though it is often just a typing error.
Not found
No record with a matching title was found. This is a reason to look further, not proof that the work does not exist. The closest record the search returned is kept out of sight, behind a Show link, because it is a different work.

Each result badge also works as a filter: select one to show only the references with that result.

Limits

  • Books, chapters, theses, conference abstracts, reports, and much of the older and non-English literature are thinly covered by these databases, so real works of these kinds may be reported as not found. The checker tags references that look like these kinds of work and says why they may be missing, but it can only guess from the wording.
  • A real paper cited for a claim it does not make passes every check here. Only reading the source catches that.
  • Matching is approximate. Unusual citation styles, typos, and errors in the databases themselves can all produce flags for references that are fine.
  • Every flag is expected to need a human look before any conclusion is drawn about a reference, or about the person who cited it.
  • The Check tab needs an internet connection. The other tabs work offline.

Books, chapters, and other hard-to-find works

Each reference is read for signs that it is a kind of work the article databases cover thinly, and tagged accordingly. The tag is a guess from the reference's wording, made to explain why a real work may be missing and where else to look.

  • Book chapter or Book: named editors (Ed., Eds., editors), a containing book introduced by "In", the name of a book publisher such as Caister Academic Press, Springer, or CRC Press, an ISBN, or an edition. Search links lead with WorldCat and Google Books.
  • Thesis: "thesis", "dissertation", "PhD", and the like. Links lead with WorldCat and OATD.
  • Conference: "abstract", "poster", "proceedings", "conference", "congress", "symposium", or "meeting".
  • Report: "technical report", "white paper", or an agency such as the WHO, CDC, or FDA.
  • Preprint: bioRxiv, medRxiv, arXiv, and other preprint servers.
  • Web page: a web address (other than a DOI link) or an access date. Links lead with the Wayback Machine.
  • Patent: a patent number. Links lead with Google Patents.
  • Software or data: an archive such as Zenodo, figshare, Dryad, or GitHub, or a label such as [software] or [dataset].

When a reference is not verified, two further notes may appear: Older work, for references from before 1970, since older literature is less completely covered, with links to HathiTrust and the Internet Archive for older books; and Possibly not in English, for references marked as being in another language, written in a non-Latin alphabet, or with several non-English words in the title.

Search links

For any reference that is not verified, links to Google Scholar, Google, PubMed, and WorldCat (or WorldCat and Google Books first, for books and chapters) appear right under the explanation, each searching for the reference's title. Verified references carry the same links at the bottom of their card. Google Scholar is useful even for verified references, since its "Cited by" counts show who has cited a work. These searches cannot be run automatically, because Google Scholar offers no public way for other sites to query it, so the links open the search in a new tab.

Checking a whole library (Library Check tab)

The Library Check tab reads exports from reference managers rather than formatted citations: RIS, MEDLINE/MEDLARS tagged, comma- and tab-delimited, and XML (PubMed, EndNote-style, or other). The format is worked out from the file's contents, not its name. Each record's own fields are used directly (title, authors, year, journal, volume, issue, pages, DOI, and PMID), and each record's own ID is shown with every result. For delimited and unfamiliar XML files, the page guesses which column holds which field and shows its guesses, which can be corrected.

Files are read byte by byte, and the text encoding (UTF-8, DOS cp437, or Windows cp1252) is worked out from the contents, with a sample of accented words shown for checking. Reference Manager exports can be in UTF-8 or DOS (cp437) text, and its RIS files mark italics with control characters; all of these are handled, and a DOI-added copy is written in the original encoding, so nothing else in the file changes. Characters that Reference Manager could not store appear in its exports as “�” (often a narrow space before a unit, as in “38–74 nm”); they were lost before the export, and the page leaves them as they are. Reference Manager keeps DOIs in its M3 field as “10.xxxx/yyyy [doi]”, so in its files an added DOI goes there, in that form. Its comma- and tab-delimited exports have no header row and are recognized by their fixed column order, and in its MEDLARS exports the journal, date, volume, issue and pages are read from the single SO line. RIS tags are recognized only in their exact form (two characters, two spaces, a hyphen, and a space), so lines inside notes are never mistaken for tags. "Online ahead of print" in the issue field marks a record as ahead of print, and "PM:" in a link field gives a record's PMID.

Each record is looked up first through its own DOI, then through its PMID (PubMed gives a DOI and current volume and pages), and then by a search on its title and first author. A match must agree on title and first author. Records tagged ahead of print are reported as now published when the database lists a volume and pages. For preprints (bioRxiv, medRxiv, Research Square, Preprints.org, and others), a journal version is looked for by title and first author with the server name left out. It counts as confirmed only when Crossref links the preprint to the article or PubMed lists the article under the preprint's “Update in”; otherwise it is reported as a lead to check. PubMed retraction notices and expressions of concern are reported first of all. Every author in a record is compared with the database’s authors (each with the one at the nearest position whose name matches, so two authors named Wang are not crossed). A surname that differs from the database’s only in accents or letters outside the English alphabet, or in a character lost from one of them, is marked as spelled differently; when that is a record’s only problem, it is listed under Author issues rather than Details differ, which is kept for first authors that are actually different. Each author spelled differently gets the database’s spelling in ordinary capitalization, written the way your record writes that author (“Genç,Y.”), with a button to copy it. Choosing Author issues shows checkboxes that sort them by kind. Spellings with accents that the database’s spelling lacks (yours has Delbrück where the database has Delbrueck or Delbruck) are just a discrepancy, most likely an error in the online source, so most likely nothing needs changing. The others are spellings whose accents or letters yours lacks. For a file from Reference Manager (including any RIS file), these are split again by whether Reference Manager can store the spelling: its Insert Symbol list holds the Windows-1252 characters, which include ç, ü, and ø but not, for example, the Turkish dotless ı or the Polish Ł. For a spelling it cannot store, the closest one it can is offered as well (Kılıç becomes Kiliç). For files from other reference managers, which store any letter, that split is left out.

Choosing Details differ lists each kind of difference with a checkbox (journal name written differently, issue missing from yours, pages that differ, and so on), so the list can be narrowed to the differences that matter most to you. One kind deserves a note: ASM journals number their articles like “e02263-25”, the last two digits being the year, but PubMed writes the same number without its hyphen (“e0226325”). References copied from PubMed therefore often lack the hyphen. The checker recognizes the two forms as the same article number and says which has the hyphen, rather than reporting the pages as different. The Check tab does the same.

Your collection. Possible duplicates within the library are found as soon as a file is read (same DOI, same PMID, same title, the same first author, year, volume, and first page, or nearly the same title) and again after the check, when records matched to the same database record are grouped too. A preprint grouped with an article is flagged as possibly a preprint and its journal version. Choosing Possible duplicates shows checkboxes for the kinds of duplicate, since they call for different things: a correction, erratum, or retraction notice with its article; a preprint (bioRxiv, Preprints.org, and the like) with its journal version; preprints only; published records with the same DOI, which are the same article entered twice; the same work published in two places, such as a journal article and a book chapter, or articles in two different journals; and published records in the same place without the same DOI, split by whether their volumes and first pages agree (where both have them). Those that agree are probably true duplicates; those whose volumes or pages differ are almost always different articles, often by the same authors with similar titles. Each possible duplicate has two buttons that copy the record IDs joined by OR, one plain (“46527 OR 46528”) and one with each ID in curly brackets (“{46527} OR {46528}”), the form Reference Manager’s search needs to bring the records up together for comparison. The report has both, and the kind of duplicate, in their own columns. Reference Manager’s Reprint field (“In File”, “Not in File”, “On Request”), or a keyword such as “received: PDF”, says whether the article itself is held. For records not in file, the check then asks OpenAlex, which carries Unpaywall’s open-access data, whether a free, legal copy exists (on the publisher’s site, in PubMed Central, or in a repository) and links to its PDF where there is one. Buttons under “Your collection” list the possible duplicates, the records not in file, those of them with free full text, and the records in file; the report has columns for all of these.

Titles in Title Case. A journal article’s title is expected in sentence case, so a title written in Title Case (or all capitals) is noted, with a sentence-case version to copy. Abbreviations (any word with two capitals, a digit, or a capital inside it: DNA, T4, mRNA, CRISPR-Cas), genus names and names of higher taxa (Escherichia, Siphoviridae, Enterobacteriaceae), Salmonella serovars, and common proper nouns (Gram, Poisson, countries) keep their capitals, and species epithets are lowercased (Escherichia coli). The first word after a colon keeps its capital too. After a check, when the matched database record writes the same title in sentence case, its capitalization is used instead, which also gets any other proper nouns right. Books and book chapters are left alone. The “Titles in Title Case” button under “Your collection” lists them, and the report has the suggestion in its own column.

Requests go out three at a time with a pause between records, and every lookup waits when a service asks it to slow down. Progress is saved in the browser after each record, so a long check can be paused, or the tab closed, and resumed by choosing the same file again. Browser storage can be cleared, though, and does not travel between computers, so a progress file (JSON) can also be saved at any point. It holds the export itself, every result so far, the DOI ticks, the PubMed records, and any column choices; choosing it as the file, in any browser, restores the check without looking anything up again, and the remaining records can then be checked. The report and the copy with DOIs can be made from a restored check just as from the original. While a check is running, or once it has results, reloading or closing the page brings up the browser’s own “Leave site?” question first.

Stopping a check

Each tab that looks things up can be stopped part way: Stop beside Check references in the Check tab, Stop in the Add DOIs and Link Check tabs, and Pause in the Library Check (which keeps what it has found, to go on later). Lookups already under way finish first, so a check stops within a few seconds. In the Check tab, references not yet looked up are listed as “Not checked (stopped),” and the results so far can be read, copied, and downloaded as usual; choosing Check references again checks the whole list from the start.

Reading references from a file or a web page

Rather than pasting a reference list, you can choose a file or give a web page’s address under the box in the Check tab. The page reads Word (.docx), PDF, OpenDocument (.odt), RTF, web pages (.html files, or an address), and plain text, finds the reference list, and puts its references in the box, one per paragraph, for you to look over before checking. The reference list is found by its heading (References, Literature Cited, Bibliography, Works Cited, Sources, Further Reading, and the like), ending at the next heading or at figure legends, tables, or acknowledgments; a web page with several such sections gives all of them. Without such a heading, the paragraphs that look like references (with a year and an author list, a DOI, or a volume and pages) are taken instead. Wikipedia articles are read through Wikipedia’s own interface, and their reference lists are read item by item, without the “^ a b” back-links. Several files can be chosen at once, or a zip of them, and their references are put in the box together, with the status saying how many came from each file.

PDFs keep no paragraphs, only lines, so in a PDF a new reference is taken to start at a number ([1], 1.) or at a line beginning with an author’s name after a line that ended a sentence; check that each reference in the box is whole. Older Word files (.doc) and Pages files cannot be read; save them as .docx first. A web page on a site that does not let other pages read it can be saved (Ctrl+S) and the saved file chosen instead. To have DOIs written into a Word document itself, use the Add DOIs tab.

Copying a single report

Every result in the Check tab, every checked record in the Library Check tab, and every link in the Link Check tab has a Copy report button. It copies that one result as plain text, ready to paste into an email, a note, a spreadsheet cell, or another program: for a reference, the reference as cited, the result and its explanation, and the record found, field by field; for a library record, every column of the report that has a value; for a link, its text, the page found, and what was flagged. A group of possible duplicates has a Copy reports button for all of its records together, and Copy these reports above a list copies the reports of everything the list is showing. If the browser blocks copying, the text is shown in a box, selected, to copy by hand.

Checking links (Link Check tab)

The Link Check tab reads a list of links and looks up what each one leads to, so that links can be checked against their pages without opening them one by one. A line can hold a bare link, or the text the link should match followed by the link (separated by a tab, as in a term-and-link table, or by a vertical bar), or a Markdown link. Pasting a page’s HTML, or choosing an HTML file, reads every link on the page along with its text. Each address is looked up once, however often it appears.

Wikipedia. Wikipedia links, and links to the other Wikimedia projects, are read through Wikipedia’s own public interface, twenty titles at a time. It reports whether the page exists, whether the title is a redirect (and to what), whether the page is a disambiguation page (a list of pages with similar titles, which a link rarely means to land on), the page’s short description, and its opening sentences. A link to a section (the part after #) is checked against the page’s section headings; section links are case-sensitive, so a heading that differs only in capitalization is flagged too. An anchor set by a template rather than a heading will not be found this way, so treat a missing section as worth a look, not as certain.

Does the link match? When a link comes with its text, the text is compared with the page’s title, and with the title in the link itself when that redirects. Differences in capitalization, accents, punctuation, plural endings, and a qualifier in parentheses, as in “Transduction (physiology),” are ignored. Text that matches no title is flagged, and text that shares words with the title without matching it is listed among the links that need a look. A qualifier is shown with each result, because it says which sense of a word the page is about: a link on “transduction” that lands on Transduction (physiology) rather than Transduction (genetics) matches by text and is wrong by sense, which only the qualifier, the description, and the opening sentences reveal.

Other links. DOI links are read from Crossref or DataCite, PubMed and PubMed Central links from PubMed, Wikidata links from Wikidata, and YouTube and Vimeo links from those sites’ video-description services. For any other site the page tries to read the page’s title, description, and canonical address directly, but most sites do not allow one web page to read another, and those are listed as “Could not be read here,” to open by hand. For those, the link’s text is still compared with the end of its address (“…/terms/selection.html” reads as “selection”), and text that does not match the address is flagged, which catches many links to the wrong term page without reading the page at all.

Your own sites. Whether another site’s pages can be read is the other site’s choice: it has to answer with the header Access-Control-Allow-Origin: *. For a site you run on an Apache server, one line in the site’s .htaccess file, Header set Access-Control-Allow-Origin "*", does it (other servers and hosting services have their own setting for it). For public pages, this exposes nothing new: it lets other pages read what anyone can already open in a browser. Once it is set, the Link Check shows those pages’ titles and descriptions like any others.

The links on a page. Give a page’s address and choose “Check the links on this page,” and the page is read, its links (with their text) are listed in the box, and they are checked. Wikipedia pages are read through Wikipedia’s own interface, leaving out its edit links and its pages about itself; other pages are read directly, which works where the site allows it (your own, once the setting below is made). Where a site does not, open the page, view its source (Ctrl+U, or ⌘+Option+U on a Mac), select it all, copy it, and paste it into the box instead; “How to copy a page’s HTML” under the box gives the steps. Giving a single address in the box checks that page itself, and the page then offers to check the links on it instead.

Search links and long lists. A link to a search (Google, Google Scholar, Google Books, Bing, DuckDuckGo, PubMed, NCBI’s other databases such as Nucleotide and Taxonomy, Europe PMC, bioRxiv and medRxiv, ICTV’s taxonomy, and PhagesDB) or to a dictionary entry (Dictionary.com, Merriam-Webster) is not looked up; instead, the words searched for are compared with the link’s text, and a search for different words is flagged.

How a search is written. Each search link is also checked for mistakes that change what it finds, and those are listed as “Searches written incorrectly”: an empty search; an unmatched quotation mark or unbalanced parentheses; a lowercase “or” (or, for PubMed and Europe PMC, “not”) that is searched for as a word instead of combining terms; curly quotation marks, which PubMed and Europe PMC do not take as marking a phrase; a Google operator such as site: followed by a space; a PubMed or NCBI field tag that does not exist (such as [tiabs] for [tiab]); AND, OR, or NOT with nothing on one side; HTML codes such as & left in an address copied from a page’s source; and searches encoded twice, except for bioRxiv and medRxiv, whose search addresses must be encoded twice (%2522 for a quotation mark), and are flagged when they are not. Smaller matters, such as PubMed’s old address or a plus sign written as %2B, are noted without being flagged.

Kinds of links. “Show kinds of links” above the list shows one kind at a time, whatever site it is on: Wikipedia and the other Wikimedia sites; videos (YouTube, including youtu.be links, and Vimeo); articles (DOI, PubMed, and PubMed Central links); searches (Google, Google Scholar, PubMed searches, and the others listed below); dictionaries (Dictionary.com, Merriam-Webster, and the like); the Biology as Poetry and phage.org sites (biologyaspoetry, phage.org, phage-therapy.org, archaealviruses.org, and asmallerflea.org, with their subdomains); and all other sites. It works together with “Show links to,” “Show links on,” and “Only broken links,” and the counts, the list, and “Copy these reports” follow it.

Several pages at once. Several files can be chosen together, or a zip of them (a folder of a site’s pages, say). Each page’s links are read, the same link on several pages is looked up once, and each result says which pages it is on; “Show links on” above the list then shows one page’s links at a time, and the report has a “Found on” column. A page’s relative links are placed by the address the page gives for itself; for pages that give none, put the site’s address in the box for a page’s address before choosing the files, and it is taken as the address of the zip’s top folder (so a file at laboratories/hgt/index.html in the zip is read as being at that address plus laboratories/hgt/).

Long lists. A whole term-to-link table (thousands of lines, “#” comment lines included) can be chosen as a file and checked in one go; each address is looked up once, Wikipedia’s twenty at a time, and a list that size takes several minutes. A list is read as a page’s HTML only when it is mostly HTML, so an HTML example in a comment line does not change how the list is read.

Capitals in addresses. Most web servers, yours included, treat “/laboratories/Studying/” and “/laboratories/studying/” as different addresses. So for a link with capitals in its address, the lowercase address is tried too. A link that is broken as written but works in lowercase is listed as “Broken, but works in lowercase,” with the working address; one that works both ways is listed as “Works in lowercase too.” For Wikipedia, where only a title’s first letter is free, a missing title is tried with the rest in lowercase (“Lysogenic Cycle” as “Lysogenic cycle”).

When a page cannot be read. The page then asks the site once more, without asking to read it, to tell two cases apart. If the site answers, it is there but does not share its pages with other sites, so it lacks the setting described above, or a redirect on the way lacks it: a domain that forwards to another (one address redirecting to another, or a registrar’s domain forwarding) needs the setting where the forwarding happens too. If the site does not answer at all, it could not be reached. A link that starts with http:// cannot be read by a secure (https://) page at all, so its https:// address is read instead, with a note saying so.

Pasted HTML. Links in a page’s HTML are often relative (“/laboratories/”). They are read as on the page’s own address when the HTML gives it (in a <base> or canonical link), or as on the address typed in the box for it; otherwise they are skipped. To see only some of the links, use Show links to above the list: each site is listed with its other endings and subdomains together (biologyaspoetry.com, .org, and .net, say, or phage.org and all its subdomains), Wikipedia and the other Wikimedia projects as one, and “only” beside each shows that one alone. “Only broken links” narrows the list to pages that do not exist. These change what is shown, not what is looked up: the counts, the list, and “Copy these reports” follow them. To look up only Wikipedia and other Wikimedia links in the first place, tick the box for it before checking. The report (CSV) has a row for every link, and “Copy the links with redirects resolved” copies the whole list back, one link per line with its text, with each redirect replaced by the page it leads to.

Reporting a result that looks wrong

Every result has a 💬 Feedback button. It opens a box for a comment and puts together a report of what the checker did with that reference: the reference as checked, the result and its explanation, the title the checker picked out, the matched record and its field-by-field comparison, any other candidate records, and, on the Add DOIs tab, the DOI decision. Copy feedback copies the report, and Email feedback opens your email program with the report filled in, addressed to references@biologyaspoetry.org. Nothing is sent unless you send it, and the report holds only that one reference and what was found for it. Reports like these are how the matching rules get improved.

Adding DOIs to a list or document

The Add DOIs tab checks each reference exactly as the Check tab does, then applies a stricter test before writing any DOI into your list. A DOI is added automatically only when every one of the following holds: the record's title words appear, in order, in the reference (at least 90 percent of them); the first author's surname, the year, the volume, and the first page all agree; the record is the paper itself, not a correction, reply, or similar notice; and no other record passes the same test with a different DOI. When a preprint and its published version both pass, the published version is chosen, unless the reference cites a preprint server. Everything else comes to you for a decision.

Four kinds of input are read. A Word document (.docx) is read paragraph by paragraph: if it has a References, Literature Cited, or Bibliography heading, only the paragraphs after it are used (stopping at Figure Legends, Tables, or a similar heading), and each paragraph is taken to be one reference. Each DOI is added as a new piece of text at the end of its paragraph, in the paragraph's own font, so italics, numbering, and the rest of the document are untouched. A replaced DOI is changed where it stands. A plain-text list (.txt, or pasted) is split into references as in the Check tab. RIS and BibTeX files, the formats reference managers export, get the DOI in each record's DOI field.

If a Word bibliography was built by a reference manager (Reference Manager, EndNote, Zotero, or Mendeley), the page says so: the next time the reference manager updates the bibliography, DOIs typed into the document may be lost. In that case, add the DOIs to the reference library instead, for example by exporting the library as RIS, adding DOIs here, and importing the result.

Crossref offers a similar service of its own, Simple Text Query, which returns Crossref DOIs for a pasted reference list.

Finding duplicates

Find duplicates compares every reference in the list with every other, straight away and without looking anything up, so it works offline and on lists of up to 2,000 references. Two references are grouped as possible duplicates when they have the same DOI, the same or nearly the same title (for short titles, the same title and the same first author), the same first author, year, volume, and first page, or nearly identical wording. After a full check, two references matched to the same database record are grouped too, which catches the same work written in two different styles. Each group says why it was grouped and points out any years, volumes, pages, or DOIs that differ, which usually shows which version to keep. A paper and its correction, or parts I and II of a series, can look alike, so each group needs a look before anything is deleted.

Finding DOIs

Any reference matched to a database record gets a Copy DOI button: references marked Verified or Details disagree, and references whose cited DOI points elsewhere but whose title matched a record separately. For a Possible match, the button is labeled as unconfirmed, because the record may be a different work. Above the results, Copy DOIs copies one line per reference, in order, and Copy references with DOIs added copies the list with DOIs added under the same strict rules as the Add DOIs tab: only where exactly one record matches the reference in every detail. References with a likely but less certain DOI, or with a wrong one, are left as they are; the Add DOIs tab shows them for a decision. Adding a DOI does not correct a reference's other details, so references marked Details disagree still need fixing by hand.

Data sources

Bibliographic records come from the free public interfaces of Crossref, OpenAlex, and DataCite.

🗃 Library Check: Bulk Checking for a Reference Manager

For curating a reference library rather than one paper's reference list. Export records from your reference manager, choose the file here, and every record is checked against Crossref, OpenAlex, and PubMed. The results show problems first: records not found, details that differ from the database, articles once ahead of print that now have a volume and pages, preprints whose journal versions have appeared, and DOIs your records lack. A report lists every record by its own ID, so you can find it in your library, and a copy of your file can be made with the DOIs added.

Corrections stay in your hands. The page never writes a corrected file for re-import, because a reference manager is likely to add the corrected records as duplicates rather than update the originals. Use the report to make changes in your library itself.

Or paste the contents of an export
Settings
Sent with each request to Crossref, OpenAlex, and PubMed, as they ask of anyone making many requests; Crossref answers such requests faster. It is remembered in this browser only.
About one to two seconds per record; a library of 1,000 records takes roughly half an hour. Progress is kept in this browser as it goes. For a long check, also save a progress file now and then: it holds your export and every result so far, and choosing it here, in this or any other browser, picks the check up where it was. While a check is running, or once it has results, the browser asks before this page is reloaded or closed.

🧮 Phage Biology and Phage Therapy Calculators

A suite of free, browser-based phage biology (🔬) and phage therapy (💊) calculators, plus teaching and accessibility tools (🎓), by Stephen T. Abedon. All open in a new browser tab.

Loading calculators…