Longitudinal Sampling of URLs From the Wayback Machine

TL;DR

This study presents a large-scale longitudinal sampling of 27.3 million URLs from the Internet Archive's Wayback Machine, analyzing web page longevity and archiving patterns over 26 years to inform future web preservation research.

Contribution

It introduces a comprehensive sampling methodology for archived web pages, addressing biases and limitations in existing datasets, and provides insights into web archiving trends over time.

Findings

More URLs archived in later years due to increased archiving capacity.

Sampling biases towards popular domains like Yahoo and Twitter.

Lessons learned to improve future web sampling strategies.

Abstract

We document strategies and lessons learned from sampling the web by collecting 27.3 million URLs with 3.8 billion archived pages spanning 26 years (1996-2021) from the Internet Archive's (IA) Wayback Machine. Our goal is to revisit fundamental questions regarding the size, nature, and prevalence of the publicly archivable web, in particular, to reconsider the question: "How long does a web page last?" Addressing this question requires obtaining a sample of the web. We proposed several dimensions to sample URLs from the Wayback Machine's holdings: time of first archive, HTML vs. other MIME types, URL depth (top-level pages vs. deep links), and top-level domain (TLD). We sampled 285 million URLs from IA's ZipNum index file, which contains every 6000th line of the CDX index. These indexes also include URLs of embedded resources such as images, CSS, and JavaScript. To limit our sample to "web pages" (i.e., pages intended for human interaction), we filtered for likely HTML pages based on filename extension. We then queried IA's CDX API to determine the time of first capture and MIME type of each URL. We grouped 92 million text/html URLs based on year of first capture. Archiving speed and capacity have increased over time, so we found more URLs archived in later years. To counter this, we extracted top-level URLs from deep links to upsample earlier years. Our target was 1 million URLs per year, but due to sparseness during 1996-2021, we clustered those years, collecting 1.2 million URLs for that range. Popular domains like Yahoo and Twitter were over-represented, so we performed logarithmic-scale downsampling. Our final dataset contains TimeMaps of 27.3 million URLs, comprising 3.8 billion archived pages. We convey lessons learned from sampling the archived web to inform future studies.