PubMed Scraper

For systematic reviews and literature monitoring, pharma and biotech competitive intelligence, academic research and RAG pipelines over medical literature: PubMed articles as flat rows, straight from NCBI's official E-utilities API — title, the full abstract with its section labels, every author with affiliations and ORCID, journal, volume, issue, pages, DOI, PMC ID, MeSH terms, keywords, article types and a retraction flag. Search the way you search PubMed, or look up a list of PMIDs.

One PubMed search stops at 9,999 results. This goes past it — without counting anything twice. E-utilities refuses to page beyond the 9,999th record ('retstart' cannot be larger than 9998), and the obvious workaround double-counts: an article with separate online and print dates falls into both slices. Measured on 2026-09-12, yearly publication-date slices of crispr added up to 77,497 against a true total of 70,772 — about 6,700 phantom articles. This Actor splits by date added instead, newest first, and never returns an article twice.

And a typo in a PubMed query becomes a bill for thousands of articles, silently. All measured on 2026-09-12, all answered HTTP 200 with a normal-looking result: crispr zzqqxxnotaword drops the unknown word and returns all 70,772 crispr articles; crispr[zzfield] ignores the unknown field tag and returns 70,772 again; "crispr gene editing zzqq" finds no such phrase, drops the quotation marks and returns 25,351 articles matching the words separately; and "Doudna JA"[Author] OR "Zzqqxx QQ"[Author] returns Doudna's 436 articles and says nothing at all about the second author.

Dates are only ever what PubMed states. They arrive as a full date, a year and month, a year and season (1976 Fall) or free text (1976 Dec 23-30), so publicationDate holds only the precision actually published (2026-08-31, 2024-03 or 1976) with PubMed's own wording kept alongside. Of 300 records from 1976, 167 are in one of those looser forms. The DOI and PMC ID come only from the article's own ID list, and the PMID only from its own field — never inferred from a neighbouring record.

No API key, no browser, no proxy. Export as JSON, CSV or Excel.

{
  "query": "semaglutide",
  "publicationTypes": ["Randomized Controlled Trial"],
  "dateFrom": "2024",
  "maxResults": 500
}

Look up PMIDs directly (numbers, PMID: 33087898 or PubMed links all work):

{ "pmids": ["33087898", "23287718", "https://pubmed.ncbi.nlm.nih.gov/22745249/"] }

What this fixes about the raw API

Everything below was measured against the live API on 2026-09-12, not read from documentation.

1. One PubMed search stops at 9,999 results - this goes past it

E-utilities refuses to page beyond the 9,999th record of a search ('retstart' cannot be larger than 9998). Above that the only way through is to cut the search into date ranges. The obvious cut, publication date, double-counts: an article with separate online and print dates falls into both. Measured on 2026-09-12, the yearly publication-date slices of crispr added up to 77,497 against a true total of 70,772 - about 6,700 articles counted twice. The date an article was added to PubMed gives each article exactly one date, and those slices added up to the total exactly. (PubMed grows every day, so these counts are larger by the time you read this; what stays true is which cut double-counts.)

So when you ask for more than 9,999 articles, this Actor splits the search by date added, newest first, and never returns an article twice. PubMed sometimes adds more than 9,999 matching records on a single day (loading days in the 1990s did this); only the first 9,999 of such a day can be read, and a row tells you how many could not be reached instead of pretending the result is complete.

2. PubMed quietly widens a search it does not understand

Counts measured on 2026-09-12:

what is sent	what PubMed does
crispr zzqqxxnotaword	drops the unknown word and returns all 70,772 crispr articles
crispr[zzfield]	ignores the unknown field tag, again 70,772
"crispr (unclosed quote)	"Unmatched double quote ignored." and returns everything
"crispr gene editing zzqq"	finds no such phrase, drops the quotation marks and returns 25,351 articles matching the words separately

In each case the answer is HTTP 200 with a normal-looking result, so a typo turns into a bill for thousands of articles you did not ask for. Here nothing is fetched or charged when PubMed says it changed the search; a row states what it would have changed and how many articles the wider search matched. Ordinary words PubMed leaves out on purpose (of, on, the) do not stop a run, and a quoted phrase that does match is searched as the phrase, exactly as asked.

3. An unknown author in a list disappears without a word

"Doudna JA"[Author] OR "Zzqqxx QQ"[Author] returns Doudna's 436 articles and says nothing about the second name. When you give two or more authors or journals, each one is also counted on its own, and a name with no articles at all comes back as its own row - so a misspelt name in a list of twenty is visible.

4. Each record carries its references' IDs

A PubMed record includes the reference list, with the PMIDs and DOIs of the papers it cites (46 records held 1,464 DOI entries). The DOI and PMC ID here come only from the article's own ID list, and the PMID only from its own citation - never from a reference or from a retraction notice.

5. Dates are not made up

Publication dates come as a full date, a year and month, a year and season (1976 Fall) or free text (1976 Dec 23-30). publicationDate holds only what PubMed actually states (2026-08-31, 2024-03 or 1976), publicationDatePrecision says which (day, month, year), and publicationDateText keeps PubMed's own wording, written the way PubMed writes it (2023 Feb 1, not 2023 Feb 01). Of 300 records from 1976, 167 have no abstract; those rows have abstract: null and hasAbstract: false, not an empty string.

dateAddedToPubmed is the same day PubMed's own [edat] index uses - checked record by record against the EDAT field of PubMed's MEDLINE output on 2026-09-12 - so the newest value in a run can be fed straight back in as dateFrom with dateType: added without re-paying for the boundary day.

6. Missing PMIDs are reported, not dropped

efetch silently leaves out a PMID it has no record for. In a lookup, every PMID you gave comes back: as an article, or as a row saying it was not found, not read, or not looked up because the run hit its charge limit or time limit. Only article rows are charged.

What you get

pmid, url, title, abstract, hasAbstract, authors (as Last, First), firstAuthor, lastAuthor, authorCount, affiliations, authorDetails (last name, fore name, initials, collective name, ORCID and affiliations for each author), editors and editorCount (the editors of a book, who are not its authors), journal, journalAbbreviation, issn, volume, issue, pages, publicationDate, publicationDatePrecision, publicationDateText, publicationYear, electronicPublicationDate, dateAddedToPubmed, doi, pmcId, pmcUrl, meshTerms, majorMeshTerms, keywords, publicationTypes, languages, isRetracted, publicationStatus, citationStatus, articleKind (journal-article or book), bookTitle, publisher, copyrightInformation, query, checkedAt, plus source, status and note on every row.

Input

Field	What it does
query	Anything you would type into PubMed, with its syntax: AND/OR/NOT, "exact phrases", [Title/Abstract], [MeSH Terms], [Affiliation].
authors	Author names as PubMed writes them (Doudna JA or Doudna, Jennifer A), one per line. Several mean any of them. A name is matched exactly as written: Smith J matched 8,016 articles, while PubMed's unquoted Smith J[au] also takes every Smith JA, JB and so on (34,950) and splits O'Brien S into o brien s. For that wider form, write it in query instead.
journals	Journal titles or abbreviations (N Engl J Med), one per line. Several mean any of them.
publicationTypes	Article types such as Review, Systematic Review, Meta-Analysis, Randomized Controlled Trial, Case Reports, Preprint.
dateFrom / dateTo	2024, 2024-03 or 2024-03-15. A date that does not exist is refused, because PubMed answers a malformed date with zero results.
dateType	publication (publication date) or added (date added to PubMed - use this to collect new articles since a previous run).
pmids	PMIDs to fetch directly, up to 10,000 per run. Can be combined with a search; an article found both ways is returned and charged once.
sort	relevance (Best Match), pub_date (newest first), Author, JournalName.
maxResults	How many search results to return, up to 100,000. You are charged for the articles you receive. Measured rate is about 19 articles a second, so the default one-hour run timeout reaches roughly 69,000; for more than that, raise the run's timeout. A run that runs out of time stops and says so in a row rather than presenting a part as the whole.
maxRetries	Attempts per request when PubMed does not answer or says too many requests arrived.

Fields are combined with AND. If you give no input at all, the Actor runs the example search CRISPR gene editing and every row says so in its query column; if you give anything, only what you gave is searched.

Speed and limits

NCBI allows three requests per second without an API key, and this Actor keeps to that (a fourth request in the same second is answered with HTTP 429; the Actor waits and retries). Articles are fetched 200 at a time, so a search of 1,000 articles takes about six requests.

That limit is per IP address, not per run: two runs of this Actor going at the same time from the same machine share the same three requests a second and can exceed it together. Each run then spends time waiting out 429s and retrying, so it gets slower rather than failing - but if you want full speed, run them one after another.

Measured throughput on 2026-09-11: 14,753 articles in 12 minutes 50 seconds (about 19 a second), on a search large enough to be split by date added. The default run timeout of one hour therefore reaches roughly 69,000 articles; ask for more only with a longer timeout.

Source, attribution and disclaimer

Data: PubMed, courtesy of the U.S. National Library of Medicine, retrieved live through the NCBI E-utilities (eutils.ncbi.nlm.nih.gov) at the time of each run - the checkedAt column records when. Records in PubMed are corrected and updated after publication, so a row reflects PubMed on that date, not necessarily today. This Actor identifies itself to NCBI with its own tool and email values on every request, as the E-utilities usage guidelines require.

This Actor is not affiliated with, endorsed by or sponsored by the National Library of Medicine, NCBI or the U.S. government. NCBI disclaimer: the U.S. Government does not warrant or assume any legal liability or responsibility for the accuracy, completeness or usefulness of any information, apparatus, product or process disclosed. NCBI does not endorse or recommend any commercial products, processes or services. See https://www.ncbi.nlm.nih.gov/home/about/policies/.

Copyright: NLM does not claim copyright on PubMed abstracts, but journal publishers or authors may. Anyone reproducing, redistributing or making commercial use of abstracts is expected to follow the terms set by the copyright holder; the copyrightInformation column carries the publisher's notice where PubMed has one. PubMed data is not medical advice.