Search Engine Journal describes a three-layer audit for AI readiness: retrievability, attribution and meaning, and agent transaction. Its audit of 50 large sites found respectable implementation of familiar technical foundations, much weaker semantic signals, and almost no support for agent transactions. The article deserves credit for separating being fetched from being understood and being acted upon. It also acknowledges that a low score may reflect a deliberate policy.
That qualification exposes a deeper problem with the score. Readiness is presented as a percentage of implemented protocols, yet readiness is not the same as openness. A publisher that blocks model-training crawlers may be commercially prepared precisely because it has made a conscious refusal. A retailer without agent-facing transaction endpoints may be avoiding fraud and support risks that exceed current demand. Counting missing capabilities can make restraint look like technical debt.
The framework mixes established web hygiene with unsettled bets. Semantic HTML, server-rendered content, accessibility labels, structured data, OAuth discovery, and emerging commerce protocols do not belong to one linear maturity ladder. They solve different problems, carry different threat models, and have different beneficiaries. A high composite score can conceal that a company has excellent machine access but poor authorization boundaries. A low score can hide that its core content is clear and useful to customers.
The sample also limits the conclusions. Fifty major sites across unrelated industries provide an informative snapshot, not a benchmark for general readiness. The article notes that the same weights were used despite different business models. That makes comparison convenient but strategic interpretation difficult. Wikipedia, a bank, a news publisher, and a travel marketplace should not converge on the same agent-access posture. Their liability, data freshness, transaction reversibility, and economic dependence on visits differ too much.
Even the meaning layer is more complicated than schema presence. Structured data can reduce ambiguity, but it does not make claims true, current, or authoritative. A page can perfectly label a price that is wrong. Models may ignore markup, reconcile it with outside sources, or retrieve stale copies. Likewise, robots directives and content signals communicate preferences but do not enforce them, as the article itself explains. The measurable signal is not the control.
A useful audit should therefore begin with decisions rather than points. Which agents may perform which tasks? What data can they see? How is user intent authenticated? Which actions are reversible? Who bears the cost of an error? Only then should the team choose technical mechanisms and test whether real systems honor them. Report coverage by risk and use case, not a single readiness percentage.
The addendum is that an AI-readiness score is best treated as an inventory of possible interfaces. It is not a grade. The goal is not maximum machine accessibility; it is deliberate, observable, and revocable participation. A company is ready when it knows what it wants machines to do, can verify what they actually do, and can refuse the rest without being told it failed the future.