We scanned 100 websites. A score was not always an answer.

mistral.ai · observed 2026-10-02T18:18:25.625Z

docs.stripe.com · observed 2026-10-02T18:19:01.034Z

The result, with the denominator intact

On October 2, 2026, we submitted a fixed roster of 100 hostnames to the public AIWebSignals scanner. Ninety-nine requests returned usable Evidence Report payloads. Fifty-five received a numeric technical-readiness score; 44 returned a report with that score withheld. One request, for hilton.com, timed out without a usable report. That last outcome describes our observation, not the availability of Hilton to every visitor.

The roster deliberately included software, commerce, beauty, publishing, travel, education and standards sites. It was selected for useful contrasts, not drawn randomly from the web. Two affiliated controls, AIWebSignals and Machine Realms, are disclosed in the methodology. Some hostnames belong to related organizations, such as Stripe and its documentation host. It would be misleading to turn these counts into an estimate of how ready the entire internet is.

A perfect technical score still leaves questions

Mistral returned 100/100 for the bounded public technical model. Its broader report contained evidence for 28 of 48 applicable checks, rounded to 58% coverage. Twenty remained unknown. Those are compatible results: the score evaluates a narrower set of public technical checks, while coverage describes how much of the larger assessment has supporting evidence.

The reverse matters too. A withheld score is not zero. It can mean the response was not safely identified as the origin page, a probe timed out, or the available evidence did not justify an aggregate. A convincing report should let the reader inspect those limits instead of hiding them behind a confident color.

What this changes for an operator

Start with one consequential question. Are public pages discoverable? Does the intended crawler policy match the business goal? Can a machine discover authorization requirements? Is there evidence of actual requests? Those questions require different evidence. A homepage scan can help with the first questions, but private execution and customer outcomes require operator-supplied records or an explicitly authorized test.

Use the report as a decision aid: preserve the exact hostname, observation date, method and missing evidence; identify a specific next check; then verify the result after a change. Do not remove an intentional restriction merely to improve a number. Do not claim recovered traffic or sales without measuring them.

A practical next test

Run the same bounded check on a hostname you operate, read the evidence behind the headline, and choose one reproducible action. A second report can then tell you whether the public signal changed. Connected first-party observation is the next step when your question becomes what actually reached the site, rather than what could be discovered publicly.

This publication is the beginning of a research loop, not its proof of commercial success. Reader questions and reproducible counterexamples can improve the next study. Views, shares and scans remain distinct from a paid activation; none of these 100 observations establishes customer demand or revenue lift.