How the numbers are made

The instrument

What every respondent answers

Every assessment scores five dimensions. Each dimension is answered on a one to five scale, so a completed assessment carries a total between 5 and 25. The total maps to one of four verdict tiers:

Not ready

5 to 10

Premature, close the gaps first

11 to 16

Ready to pilot

17 to 21

Ready to scale with governance

22 to 25

The core readiness assessment adds one architecture question that classifies the kind of AI system that fits the respondent's business. Beyond the scored dimensions, each assessment also asks three short context questions. Those context questions never change anyone's verdict; they exist purely so the published statistics can be read by industry, by company size, and by situation. The library currently holds 70 diagnostics across 3 levels, for a company, for a team, and for an individual, and every one of them scores on the same 5 to 25 scale and the same four tiers.

The population

Who is counted

A submission enters the statistics only when all of the following hold. The respondent confirmed their email through the one-click link, so every counted answer is a double opt-in. The address was a corporate one; free and disposable mailbox providers are turned away at submission. The respondent did not opt out of having their anonymised answers used for statistics; anyone who opts out still receives their full report but is never counted. And the submission passed the data-quality guardrails.

Those guardrails are deliberately conservative, because a false positive must never cost a real respondent their report. A run completed in under 40 seconds is treated as implausibly fast and flagged. A run where every one of the five dimensions carries the same answer, a straight line, is flagged when it also finished in under 90 seconds. A flagged submission is excluded from every published statistic and every benchmark cohort, yet the respondent still receives their report exactly as anyone else would. We keep obvious junk out of numbers we publish under our own name; we do not punish a person for a fast honest read.

Anonymity

What is kept, and what can never be recovered

The email address exists only long enough to deliver the confirmation link. At the moment of confirmation the address is deleted from the submission; what remains is an irreversible salted hash, kept solely so the system can recognise a duplicate. The hash cannot be turned back into an address. IP addresses are truncated and hashed rather than stored. A submission that is never confirmed is purged in full after 72 hours.

On top of deletion, the published aggregates enforce k-anonymity with k set to 8. Any group with fewer than 8 respondents is folded into an 'other' bucket rather than shown on its own, and if that 'other' bucket itself stays under 8 it is suppressed entirely, because a lone small remainder could otherwise be recovered by subtraction against the published total. For the same reason, a module's dimension means and its score distribution are withheld until the module has at least 8 counted respondents. Below that threshold the module reports only that it is not yet statistically safe to show.

Benchmarks

The peer comparison, and the modelled baseline

A report positions a respondent's total against a peer cohort. The cohort is chosen by narrowing from the most specific match to the least: first other confirmed respondents in the same vertical, then failing that the same company-size band, then failing that all confirmed respondents. A cohort is only used once it holds at least 8 members, and a cohort never mixes modules, so a governance score is only ever compared against other governance scores.

Until a real peer group of that size exists, the comparison falls back to a modelled baseline, a distribution calibrated to the premise of the source essay rather than measured from respondents. Wherever that baseline is used it is labelled as a modelled baseline in plain text, and it switches to measured respondent data as soon as the confirmed sample is large enough. The site never presents modelled numbers as measured ones. A readiness instrument that faked its own data would be a strange thing to trust.

Company classification

Industry from the domain, never the person

The industry attached to a submission is derived from the company domain in the email address, and that derivation runs on our own servers. There is no third-party enrichment, no lookup service, no data broker. The classification describes a company, never the individual who answered, and it is used only to group aggregates by vertical.

Publication and citation

Where the numbers appear, and how to cite them

As the sample grows, the statistics are published across the Thorsten Meyer AI network of sites. Because the figures move as more respondents confirm, a citation should carry the date it was read. The suggested form is:

World Model Readiness statistics, worldmodelreadiness.com, retrieved July 2026.

The raw aggregates for any module are available at /api/stats.php?module=<slug>, returned under the same anonymity rules described above: k-anonymised, means and distributions withheld below 8 respondents, quality-flagged rows excluded. Ready-to-embed charts of the same data live at /embed/ for anyone quoting the numbers on their own page.

Limitations

What these numbers are not

The answers are self-reported, and the sample is self-selected: these are people who chose to run a readiness assessment, not a random draw from the economy. The corporate-email requirement narrows the population further, toward organisations already thinking seriously enough about AI to test their footing, which skews the picture relative to companies that are not. Counts per module are shown only when it is statistically safe to show them, so an absent number means too few respondents, not zero. Read the statistics as an honest, transparent read of a particular population, useful and citable, but not a census.