Of the 1,693 bookmarks I'd accumulated in Chrome since 2014, 212 were dead. That's 12.5% — one in eight links pointing at nothing. I found this out because I built a tool that crawls every bookmark, extracts its content, classifies it with Claude, and assembles the whole thing into a structured markdown book. The crawl step is where things got morbid.

The tool tries each URL, and when that fails, it asks the Wayback Machine for a cached copy. Of those 212 dead links, exactly one was recoverable. It was a page called "AI Agents" on commands.com — an AI workflow automation platform. The Wayback Machine had captured it in November 2025. What it preserved was a loading spinner, empty containers, and the text "No items found." The archive saved the skeleton but not the body.

The other 211 are just gone.

The act of saving

Bookmarking feels productive without being productive. You hit Ctrl+D and your brain gets a tiny hit of "I'll get to that." The problem is that future-you has the same habit.

My collection tells the story clearly. 2014: one bookmark (a cPanel webmail login — purely utilitarian). 2016: three. 2017: nine. Then the exponential curve kicks in. 2018 brought 33. 2024 exploded to 390. 2025 hit 808. By the time I ran the pipeline in early 2026, I'd added another 306. The rate of saving accelerated every year. The rate of reading did not.

The interest graph

The pipeline classifies every bookmark by topic, type, quality, and reading depth using a two-pass approach — first free-form labels, then a normalization step that merges synonyms into canonical categories. Here's what it found:

(Percentages exceed 100% because bookmarks can have multiple topic labels.)

The top domain was github.com with 278 bookmarks — more than 10x the second-place domain. I was archiving an entire professional education that I was never going to complete.

The categories weren't surprising. The shape was. The interest graph has two nearly disconnected clusters. There's the tech cluster — backend, AI, web, DevOps, business — all densely interconnected. Backend and Web Development share 173 bookmarks. Backend and AI share 163. It's one big blob of overlapping professional interests.

Then there's the music cluster. Music Production and Electronic Music share 102 bookmarks. But music connects to almost nothing else. 508 music bookmarks existing in their own universe, with minimal crossover into the technical categories. No "audio programming" bridge. No "music + code" cluster linking the two. Just two islands. When I'm bookmarking music stuff, I'm apparently a completely different person.

The graveyard

The dead links form their own kind of archaeology. The oldest casualties are from 2017 and 2018: an Excel tutorial site called Excel Exposure, an IKEA product page for a side table I apparently considered buying, a Codota page from when AI code completion was a novelty. A page for "The Hemingway" watch from Hawthorn Watch Co. — a company that no longer exists.

The dead links cluster around certain categories. Music production resources vanish at a high rate — sample pack sites, synthesizer preset pages, trance production tutorials. ACID PATTERN GROUP's acidboxblues archives: gone. Euphoric Trance Vocals from Producer Space: gone. A KVR interview with Christoph Kemper, creator of the Kemper Profiler: gone. Underground electronic music culture lives on sites maintained by one person with one server, and eventually, one reason to stop paying for hosting.

Business tools and SaaS products have the highest mortality. Scrollbar.app, Mapster.live, Scrumpy.io — all 404s now. Even the big players aren't immune. My collection includes dead links to AWS console pages, Auth0 dashboards, and Adobe XD prototypes. Nothing is immune.

Then there are the personal artifacts. A Ving travel booking for a hotel in Mallorca. A recipe for mozzarella bread with herbs and garlic from a Norwegian food blog. A Stressless recliner I was shopping for. A Seiko watch listing. Not knowledge — just life. Decisions being considered. Dinners being planned.

Time capsule

The classification also tags each bookmark with an "era context." Year by year, the bookmarks become a timeline.

2016 me was doing C# socket programming and looking up FullCalendar documentation. 2018 me discovered synthesizers and Laravel in the same year — two obsessions that would persist for the next eight years. 2020 me bookmarked 43 things during the pandemic, which feels low, but pandemic-year bookmarks tend to be heavier. 2023 me started saving AI links in earnest. By 2024, AI and LLM bookmarks had become the third-largest category in the entire collection, with 504 entries.

The Norwegian recipes appear throughout. Grilled salmon with garlic butter. Chicken pakora. Bread. They're scattered between GitHub repos and Claude conversations like small reminders that the person doing the bookmarking also eats food.

What I didn't expect was how clearly the bookmarks showed phase transitions. There's a visible moment where backend development stops being my primary learning focus and AI takes over. There's a moment where music production bookmarks shift from "how to use Ableton" to "Waldorf Blofeld SysEx protocols" — from consumer to practitioner. The bookmarks track it all whether you're paying attention or not.

Aspirational vs. actual

Claude rated each bookmark from 1 (stub/broken) to 5 (exceptional), and the depth from "quick-read" to "rabbit-hole." The collection is heavy on rabbit holes I clearly never went down. Deep-dive articles on system design patterns, 40-page PDF specifications, comprehensive guides to things I bookmarked with obvious intent to study later.

Every bookmark is a small act of self-mythology. I'm the kind of person who reads about DSP theory. Who studies system design patterns. Who cooks Norwegian salmon recipes from scratch. The bookmark says so. The unvisited link says otherwise.

Link rot

A 2024 study by the Pew Research Center found that 38% of web pages from 2013 were no longer accessible. My own collection — spanning 2014 to 2026 — shows 12.5% link rot. The discrepancy makes sense: I'm biased toward large, stable domains. GitHub isn't going anywhere. My recipe sites and niche synth forums, on the other hand, are on borrowed time.

The Wayback Machine had snapshots for some of the dead URLs, but most were JavaScript-rendered — the archive captured HTML shells, navigation bars, footers, loading states. The web moved to client-side rendering and left the archivers behind.

URLs feel like citations — stable references to stable knowledge. They're pointers to servers that someone is choosing to keep running. When that choice changes, whatever was behind the link might as well have never existed.

The triage

The pipeline doesn't just catalog the dead. It clusters everything into chapters and generates a Rabbit Hole Index — 1,481 surviving bookmarks ranked by time investment: quick reads, weekend reads, deep dives. A reading list that made my own collection legible to me for the first time. It surfaced things I'd forgotten saving. A two-volume book on the mathematical foundations of music. Brendan Gregg's USE Method for performance troubleshooting. An interpreter-building tutorial in Go.

I saved all of these with genuine intention. Some of them years ago. It took an AI to sort through my own collection and tell me what was worth reading.

I have not read the reading list.

I did, however, immediately bookmark the generated book.

The project is at github.com/HelgeSverre/bookmarking-book.