Research question
What share of a typical public document's text is page furniture, and how much time does a listener lose to hearing the same header on every page of a linear reading?
Overview
Print pagination repeats itself: the same running header, the same footer, a page number on every page. Visual readers skip this furniture without noticing. A screen reader user reading a converted document linearly hears every repetition, on every page, unless the conversion removes it.
This census measures furniture share across public documents: what fraction of tokens are repeated headers, footers, and page numbers, how the share scales with document length, and what it converts to in listening time at typical speech rates.
The number is unpublished anywhere, easy to grasp, and directly motivates a design decision most tools never make explicitly: whether repeating page furniture belongs in the reading order at all. PDF tagging rules say it does not, and this study measures how often reality disagrees.
Why this matters
It quantifies a tax that only some readers pay
Furniture repetition costs sighted readers nothing and costs linear listeners minutes per document. Making that asymmetry measurable is the argument for furniture-aware conversion, in one number.
It exposes a benchmark design flaw
Text-retention metrics reward keeping furniture, because page numbers count as source tokens. Our own published A/B shows documents scoring lower for correctly removing repeated stamps. The census gives benchmark designers the data to fix the denominator.
What we measure
- Furniture token share
- Percent of document tokens that are repeated headers, footers, and page numbers
- Listening cost
- Furniture tokens converted to time at typical text-to-speech rates, per document and per corpus
- Scaling
- Furniture share as a function of page count and document class
- Artifact discipline
- How often source PDFs correctly mark furniture as artifacts versus leaving it in the tagged reading order
Method
Detection: repeated-content clustering by position band and digit-normalized text across pages, the same detector that runs in our production pipeline, with published precision from a manually verified sample.
Corpus: the 50-document public pilot first, then the larger readiness-index corpus as it is acquired.
Listening cost model: fixed words-per-minute assumptions, stated with the results and varied in a sensitivity table.
Evidence so far
Only recorded findings appear here. Anything not listed has not been measured yet.
The detector is in production and its effects are recorded
The August 2026 A/B report documents furniture removal working on real documents, including the scanned-archival case where removing repeated page stamps lowered the text-retention score, which is the metric artifact this census quantifies.
Questions we expect
Is removing furniture always right?
No. Furniture carries information a reader sometimes wants, like current section context. The census measures cost, not policy; the accompanying analysis discusses when furniture should become navigation aids instead of repeated inline text.