# The Dead Reader — web 4.0

> An objective experiment · fibernetics skunkworks · 2026
> You are reading the agent representation of https://webv4.mosaic.site — this
> file IS the site, minus the pixels.

The Dead Internet theory claims the **writers** were replaced by machines. It
got the direction wrong. It's the **readers** who left. No one is reading your
website anymore — their agents are. And every pixel bought to persuade a human
is now a tax on being understood.

## What the agent actually reads

We measured a real business website — 54 pages, fetched the way an agent
fetches them, tokenized the way a model bills them (helmwise.com, 2026-09-02,
o200k tokenizer):

- Raw HTML: 1,283,103 tokens (4.70 MB)
- The actual text: 71,473 tokens (0.36 MB)
- **94.4% of tokens are markup, not meaning — 18× waste**
- Cost to read the whole site at $5/M input tokens: $6.42 raw vs $0.36 as text
- Context windows (200k) needed to hold it: more than six raw vs a third of one as text

Every number on this page is a measurement, not an estimate.

## The inversion, in four moves

1. **The audience changed species.** Traffic is increasingly agents
   summarizing, comparing, extracting, buying. The funnel converts a reader who
   no longer arrives.
2. **Pixels flipped sign.** Design was costly to make and free to consume; for
   machine readers the economics invert — presentation is billed by the token,
   and persuasion-by-aesthetics has no target. Chrome makes a site harder to
   summarize correctly.
3. **Reading became a metered utility.** Machine attention is purchasable and
   priced; legibility is now a line item. The world is burning astonishing
   token volumes reading JavaScript that was never meant to be read.
4. **Text is the interface again.** The web began as documents. Its newest
   readers want what its first ones did: words, structure, provenance. The
   rhetoric that persuades an agent is being clear, sourced, and fresh.

## Newgle: own the library, not the crawl

Our experiment in an agent-native web: mirror the written internet — Common
Crawl's extracted text, 2.14 billion pages, 6.42 TB — into storage we own,
refine it (parse → language filter → normalize → deduplicate → Parquet → BM25
index), and serve it to our agent fleet as a CLI:

    $ newgle search "fiber backhaul peering ontario" --json
      ranked, full-text results in ~22 ms (measured, median)
    $ newgle get https://example.org/peering-policy
      the whole page as clean text — provenance kept

Benefits, measured or structural:

- 22 ms median search; zero marginal cost per read
- Deterministic: same query, same corpus, same answer — reproducible agent runs
- Provenance: every record keeps its URL and crawl date
- Whole sites fit in one context window
- No rate limits, no robots.txt standoffs, works offline

The freshness ladder: the archive answers first (milliseconds, free) → a live
fetch covers what's new or missing and writes back into the corpus → a
diff-scanner fleet keeps the pages agents actually use warm. The index becomes
a mirror of what the fleet cares about.

## The pattern: a text mirror in front of every site

We put one in front of a partner's site in an afternoon — ~200 lines of
Python, one subdomain. Same content, two representations: rendered pages for
the last human readers, raw markdown and an llms.txt for the first machine
ones. Cleaned once, not per visit — 18× cheaper for every agent, forever.

Case — Helmwise:

- The mirror: https://helmwise-com.gct.to/
- The agent index: https://helmwise-com.gct.to/llms.txt
- A page as agents read it: https://helmwise-com.gct.to/home.md
- Coverage: 100% of the text of every page, verified against the rendered site

## The eras

- Web 1.0 — Read. Documents: text and links, made for people.
- Web 2.0 — Write. Feeds, platforms, and the pixel arms race for attention.
- Web 3.0 — Own. The ownership detour; the readers were already leaving.
- **Web 4.0 — Read, again. Documents: text and links — made for machines, kept
  honest for people.**

---

This page practices its thesis: the human version lives at
https://webv4.mosaic.site, this markdown is the agent version, and /llms.txt
indexes both. Measurements from the newgle project, 2026. The archive is still
pouring.