The web stopped being read. It started being taken.
For thirty years a visit meant a reader: attention arrived, traffic came back, everyone could see the exchange. Now the dominant visitor is a machine that takes content and answers with it somewhere else. The exchange broke, and the record of what happened is held entirely by the party that did the taking.
A dispute was inevitable. The evidence for it was not.
The reader became a taker.
| The web as read | The web as taken | |
|---|---|---|
| Unit | A page view | A corpus |
| What moves | Attention | Content |
| What comes back | Traffic | An answer, somewhere else |
| Who notices | Analytics | Almost nobody |
| What remains | A log line | A dispute |
Grey nodes hold content. Cobalt nodes take it.
Each cobalt node is a crawler fleet, an answer engine or an agent platform that reads the field around it and turns what it finds into products elsewhere. Each edge is a taking: content moving in one direction, with nothing signed at the boundary.
Schematic, not a measurement.
How much faster machine traffic to content grew than human traffic over the last three years. On many sites, bots stopped being a rounding error and became the audience.
Order-of-magnitude estimate from published infrastructure-provider reporting. The exact figure depends on methodology; the order does not.
A taking leaves no record.
On the reader's web, value came back as traffic and everyone's analytics agreed roughly what had happened. On the taker's web, value leaves silently, and the only detailed trail of the visit is held by the party with the least interest in producing it later.
Blocking matured. Metering matured. Evidence did not.
- 2023 · Taking is normal, deals begin
- Industrial crawling for training is the default, and the first content-licensing deals are signed by the largest publishers. Everyone else supplies the corpus for free.
- 2024-2025 · The lawsuits arrive
- Rightsholders sue frontier labs and answer engines. A book-piracy class action settles for $1.5 billion. Complaints begin describing rotated agents, rented residential proxies and “data laundering”, because direct attribution keeps failing.
- 2025 · The web fights back with blocking
- CDNs flip to blocking AI crawlers by default across millions of sites; pay-per-crawl rails and machine-readable licensing standards launch. Blocking and metering mature. Evidence does not.
- 2026 · The regulator and the courtroom converge
- EU obligations on general-purpose AI become enforceable in August, with training-data summaries and copyright policies now examinable. Meanwhile courts discount plaintiff-engineered exhibits, and stealth crawling is exposed by honeypot forensics. Both sides discover the same gap: nobody kept signed records.
Summarised from public filings, judgments and published reporting through August 2026.
Every generation of the web sold its own unit. The fourth column is empty.
Attention
- Sold
- An impression and a click
- Paid by
- The advertiser
- Currency
- Seconds of a human gaze
Access
- Sold
- Traffic and addresses
- Paid by
- The developer
- Currency
- Gigabytes and requests
Result
- Sold
- A completed machine task
- Paid by
- The business
- Currency
- A successful answer
Proof
- Sold
- The standing to prove what machines did
- Paid by
- Whoever faces a judge or a regulator
- Currency
- A signed record
The first three columns are occupied: blocking, metering and execution all have their giants. The fourth is not, because it is not sold to an engineer. It is sold to whoever signs the pleading, the filing or the compliance statement. That is the column this project is about.
TCP/IP, SSL and OAuth each closed its own question of trust and then became invisible. The question "what did the machines actually do, and who can prove it" is open right now.