Every capture hashed SHA-256 Bound to an RFC-3161 timestamp 98,721 domains changed their robots.txt in 30 days Sealed nightly into a Merkle digest You cannot collect this later Every capture hashed SHA-256 Bound to an RFC-3161 timestamp 98,721 domains changed their robots.txt in 30 days Sealed nightly into a Merkle digest You cannot collect this later

For publishers & rights holders · Opt-Out Efficacy Report

You opted out of AI training. Did it work?

Blocking GPTBot was step one. This report is step two: we join our timestamped record of when your domain opted out with live probes of the models released since, and put both in one document. "You blocked the crawler on date D, proof attached. Here is what models released after D still reproduce and recognize of your content."

Run the report · $499 How the join works

Why only this archive can say it

The claim needs both halves, and nobody else holds both

Membership-inference tools can probe a model but can't prove when you opted out: your robots.txt only shows its current state, and the past isn't reconstructable. Monitoring tools can watch your robots.txt but can't ask the models anything. BotWitness has been independently capturing and RFC-3161-timestamping robots.txt across the web and runs the honest model-probe suite from our Exposure Report. The join (dated opt-out, then dated model behavior) is the product.

How it works

What the report contains

1 · Archive facts

Your opt-out record

Per crawler: when your domain first blocked it at the root, every recorded state change since, and whether it's blocked today, each verdict citing the exact capture (SHA-256, RFC-3161) that proves it.

2 · Timing

Models vs your date

Each probed model's public release date is set against your earliest recorded block. Models released after your opt-out are the ones your opt-out should have reached. Where a release date isn't established, the report says so instead of guessing.

3 · Probes

What newer models still know

The same tiered evidence as our Exposure Report (verbatim reproduction, refusal-resistant recognition (DE-COP), and a relative membership signal) run on your text, with reproduced passages as exhibits.

Read this first: what the finding means

The report measures outcome efficacy: whether models released since your opt-out still show signal on your text. That is the question an opt-out is supposed to answer.
Signal in a post-opt-out model does not by itself prove post-opt-out crawling: the model may have trained on older crawls, licensed data, or copies hosted elsewhere. The report says this plainly, on the page.
Release dates are public facts; training cutoffs usually aren't. We compare against what is provable and label what isn't.
Probe results follow the Exposure Report's honesty contract: facts and clearly-labelled relative signals, never a fabricated probability. Not legal advice.

Who it's for

When your opt-out is the exhibit

Publishers

You blocked the crawlers a year ago. This is the document that says whether the industry's newest models respected the fence, with your opt-out date proved, not asserted.

Counsel

An opt-out that demonstrably didn't change outcomes is the beginning of a licensing demand or a complaint. Every date in the report is backed by a hash and a timestamp a court can check.

Rights organizations

Run it across members' domains to quantify, honestly, what opting out has and hasn't achieved, and pair it with copyright traps for the prospective half.

Get started

One report that proves the date and probes the models

Self-serve reports are $499. Pay, then run from your dashboard: your domain, your text, every configured model. Monitoring subscribers run reports from the dashboard with no per-report charge. For a portfolio of domains or filing-ready delivery, talk to us.

Related: a Compliance Audit checks your server logs against the same archived rules, and the Exposure Report is the probe half on its own.

Run the report · $499 [email protected]