Research question
When a PDF is converted to HTML, do its links still go where the author pointed them, and does the answer differ for printed URLs, link annotations, and internal references?
Overview
A PDF carries three different kinds of link: text that merely looks like a URL, invisible link annotations whose destination can differ entirely from the visible words, and internal references that jump to another page of the same document. Conversion tools routinely conflate the three, keeping link-looking text while silently discarding authored destinations and internal navigation.
No existing benchmark scores link destinations. Text-similarity metrics count a dead link and a working link as identical output. This study measures anchor and destination survival per link class, per tool, on a controlled corpus where every authored link is known.
The stakes are concrete: a benefits notice whose Apply link goes nowhere, a table of contents that no longer navigates, a printed sentence fragment autolinked into a fabricated URL. We found and fixed an instance of that last failure in our own pipeline during the August 2026 corpus run, which is exactly why the measurement deserves publishing.
Why this matters
Link purpose is a WCAG requirement, link truth is more basic
WCAG 2.4.4 asks whether link text conveys purpose. Conversion adds a prior question nobody audits: does the link still point where the author pointed it? A confidently wrong link is worse than no link.
Internal navigation is what makes long documents usable
For a 200-page report, the table of contents is the accessibility feature. Converters that flatten internal links to plain text quietly remove the primary navigation aid for every reader, and disproportionately for assistive technology users.
What we measure
- Annotation survival
- Share of authored URI link annotations that reach the HTML with correct destination and sensible anchor text
- Internal reference survival
- Share of GoTo and named destinations that become working in-document anchors
- Printed URL handling
- Correct linking of real printed URLs and, separately, the false-positive rate of fabricated links
- Anchor text integrity
- Whether visible anchor text survives conversion intact
Method
Corpus: a controlled set with fully enumerated authored links, plus the 50-document public pilot for in-the-wild rates.
Ground truth: every link annotation, destination, and printed URL extracted and recorded from source PDFs before any conversion runs.
Subjects: common conversion paths including office-suite exports, browser and command-line converters, document AI parsers, and our pipeline, with versions recorded.
Metrics: per-class survival and false-fabrication rates, published per tool with raw counts.
Evidence so far
Only recorded findings appear here. Anything not listed has not been measured yet.
The measurement machinery is in production
Our pipeline extracts printed URLs, URI annotations, and internal GoTo destinations as distinct classes, resolves internal links to heading anchors, and records provenance for each. This is the instrumentation the study runs on.
A registered false-fabrication case
The August 2026 corpus run caught a text autolinker fabricating URLs from missing-space sentence fragments such as 1.Less, in output from the earlier pipeline. The defect class is real, shipping in tools today, and measurable.
Questions we expect
Why measure fabricated links, not just lost ones?
Because fabrication is the silent failure. A missing link is visible on inspection. A fabricated one looks intentional, passes casual review, and sends readers to domains the author never referenced.