Ask a researcher how they read the literature and the answer, more often than not, still involves a PDF. Papers are downloaded, annotated, filed in reference managers, emailed to colleagues and printed for journal clubs. Three decades after journals moved online, the fixed-layout page remains the form in which much of science is actually consumed and passed around.
Yet behind the scenes the PDF has already lost its place as the format of record at many publishers. The version that gets archived, indexed and exchanged between systems is usually structured XML, and the version most readers first meet through a search engine is an HTML page. Accessibility law, machine reading and computational research are all adding pressure from different directions.
The more useful question, then, is not whether the PDF will disappear. It is what the PDF is still for, and what publishers lose when they treat it as the article rather than as one rendering of it.
Why the PDF has lasted
The PDF's strengths are real, and they explain its staying power better than inertia alone.
The first is fidelity. A PDF looks the same on any device and in any viewer, which matters for disciplines built on complex equations, chemical structures, dense tables and carefully composed figures. Authors and typesetters know that what they approve at proof stage is what readers will see.
The second is portability. A PDF is a single, self-contained file. It can be saved locally, read offline, shared outside a publisher's platform and attached to a grant report or a promotion dossier. For researchers working with unreliable connectivity, or behind institutional access barriers, a downloaded file is simply more dependable than a web page.
The third is citation habit. Page numbers, fixed pagination and the visual identity of the typeset article still anchor how many fields quote and reference work, particularly in the humanities and in legal and policy writing.
The fourth is preservation. PDF is an open ISO standard (ISO 32000), and the PDF/A family (ISO 19005) was designed specifically to keep the visual appearance of page-based documents stable over the long term, independent of the software used to open them. The Library of Congress describes PDF/A as a set of constrained forms of PDF intended to be self-contained and device-independent. For a format whose job is to look the same in fifty years, that is a strong credential.
The record behind the page is already XML
What many readers do not see is that, for a large share of journal content, the authoritative machine-readable version of an article is not the PDF at all. It is XML tagged to the Journal Article Tag Suite, or JATS.
JATS is a NISO standard (currently ANSI/NISO Z39.96-2024, JATS 1.4) that defines elements and attributes for describing journal articles: titles, contributors, affiliations, abstracts, sections, references, figures, tables, funding and more. It grew out of the archiving and interchange DTD work that the US National Center for Biotechnology Information began in 2002, and it comes in three tag sets ranging from the permissive Archiving and Interchange model to the more prescriptive Article Authoring model.
PubMed Central's file specifications show where the weight sits. Publishers depositing to PMC must supply full-text XML conforming to an accepted journal article DTD, with JATS preferred. A PDF matching the XML is expected for each article as well, and PMC states a preference for PDF/A. In other words, one of the world's largest full-text archives treats structured XML as the substance of the record and the PDF as an accompanying rendition.
That distinction matters for publishers. An article that exists only as a well-designed PDF, with no clean XML behind it, is harder to deposit, harder to index, harder to convert into accessible formats and harder to correct consistently across platforms.
HTML and the reader on a phone
On the reading side, the most visible challenge comes from HTML. A responsive article page reflows to fit a phone screen, links references directly to their sources, exposes figures at full resolution, and can carry information that a static file cannot: version histories, correction notices, linked datasets, ORCID-identified authors and published peer-review reports.
The preprint server arXiv offers a clear example of the shift. It now generates HTML versions of papers alongside the PDF, using a LaTeX-to-HTML converter developed with the LaTeXML team at the US National Institute of Standards and Technology, and it is gradually backfilling its corpus of more than two million papers. arXiv says the main motivation was accessibility: HTML works better than PDF with screen readers, text-to-speech, magnification and mobile devices. It is also candid about the limits. The HTML service is labelled experimental, some papers do not convert, unsupported LaTeX packages cause errors, and the layout differs from the typeset PDF.
That candour is instructive. Converting legacy PDF-first workflows into good HTML is difficult, especially for mathematics-heavy content. Publishers that already produce structured XML upstream are in a far better position to generate high-quality HTML than those trying to extract structure from finished pages.
Accessibility is now a compliance question
Accessibility has moved from good practice towards legal obligation, at least in part. The European Accessibility Act (Directive (EU) 2019/882) has applied since 28 June 2025 to a defined list of consumer products and services in the EU market. The European Commission's summary lists e-books and e-commerce among them, alongside banking, transport and telecommunications. Microenterprises providing services can be exempt, and there are transitional arrangements running to 2030.
For scholarly publishers the scope needs care. E-books, the dedicated software used to read them and online shops selling them are clearly covered. Whether journal articles delivered through a publisher's website fall within the Act is less settled, and trade commentary has been divided. Elsevier's public statement, for example, describes the Act as applying to e-books, e-commerce sites and mobile apps, while noting separately that its books and journals published from July 2025 are tagged for assistive technology. Publishers should take legal advice on their own products rather than assume either that journals are exempt or that they are fully in scope.
Whatever the legal position, the technical direction is clear. An accessible PDF has to be tagged so that assistive technology can follow reading order, headings, tables and alternative text. The relevant standard is PDF/UA (ISO 14289); its second part, PDF/UA-2, published in 2024, is based on PDF 2.0 and adds detailed requirements for mathematics, including MathML support. The PDF Association itself notes that conformance does not on its own guarantee that content is accessible. EPUB and well-structured HTML generally make accessibility easier to achieve because structure is native to the format rather than layered on afterwards.
Machines are readers too
A growing share of the audience for research articles is software: indexing services, text- and data-mining tools, citation-checking systems and, increasingly, AI models. For these readers the PDF is an awkward container. Extracting clean text, reference lists, tables and equations from typeset pages is error-prone, and two-column layouts, ligatures and running headers all introduce noise.
Structured XML and clean HTML solve much of this by design. They also make it easier to attach machine-readable licensing and rights information to the content itself, which has become commercially significant as publishers negotiate the terms under which their content may be mined or used to train AI systems. The economics of those negotiations are a separate story; for the format question, the point is that structured content is the version machines can use reliably.
Articles that run
At the far end of the spectrum are articles that are not documents at all in the traditional sense. In August 2020 eLife launched Executable Research Articles, built with Stencila, which let authors of published eLife papers add live code and computed outputs. Readers can inspect and re-run analyses in the browser without altering the published article, and download the code and data. eLife positioned the format explicitly as a complement to the original paper, not a replacement.
That framing captures where interactive formats stand. They are valuable for computational work, but they depend on sustained platform investment, and they raise their own preservation questions: a static PDF can be archived far more simply than a live computational environment.
What this means for publishers
The PDF is not going away. Researchers want it, archives accept it, and for many readers it remains the most practical way to keep and annotate a paper. But it is no longer a sound foundation on which to build a publishing operation.
The practical lesson is to treat the article as structured content first and to generate the PDF, the HTML and any other formats from that single source. Publishers that do so can deposit cleanly to archives such as PMC, produce accessible outputs with less remediation, keep corrections and versions consistent, and serve both human and machine readers. Those that still begin with a typeset page and work backwards will find each of those tasks harder and more expensive.
The PDF, in short, is still at the centre of how many people read science. It is steadily moving away from the centre of how science is published. For a wider view of how the journal itself may change, see our analysis of what academic journals could look like in 2035, and for the cost side of production technology, whether AI can really cut the cost of running a journal.
