It is easy to find a number for how much employers want AI skills, and hard to find one that survives being asked how it was produced. Search job adverts for the word AI and a very large share of senior commercial roles appear to want it. That figure is close to meaningless, because most of what it counts is not a requirement at all. A new weekly job collects public postings, narrows them to director level and above in business to business technology sales and marketing, and then does the part that actually matters: it separates AI named as an expectation of the candidate from AI used to describe the employer, and from notices about the employer screening applications. The first wave has been published with its full methodology.

• The headline is 33.3 percent, and the interval around it is stated rather than implied. Of 700 analyzed postings open in Canada and the United States, 33.3 percent require AI skills of the candidate, with a 95 percent confidence interval of 29.9 to 36.9. The interval is a Wilson score interval rather than the usual approximation, which matters at proportions near zero or one because the common method produces bounds that are impossible. Sample size does not change how confident the figure is, it changes how wide the band is, and both are reported so a reader can judge the precision instead of taking it on trust.
• The same postings counted loosely give 74.3 percent, and the gap is the finding. A keyword search of the identical set reports 74.3 percent, because that share mentions AI somewhere in the advert. Only 40.6 percent mention it inside the requirements, and 33.3 percent ask anything of the candidate. Most of that gap turns out to be location rather than marketing: 33.7 points of it is AI language sitting in company descriptions, benefits and legal text, against 7.3 points lost to requirements that mention AI while asking nothing. A naive count therefore overstates demand by roughly a factor of two, and not for the reason most people would assume.
• Every counted posting can be traced to a specific sentence. Each classification has to carry a quotation from the advert itself, and that quotation is checked automatically against the source text before the posting is counted. A classification whose quotation does not verify is discarded rather than downgraded, and the check runs twice, in the stage that produces it and again in the stage that publishes it, so that a result recorded before a check existed can never enter a published count. This is the difference between a number that can be spot checked and one that collapses when somebody tries.
• The limitations are published with the figure, including the one that weakens it most. Although postings were collected from two boards, 87.6 percent of the analyzed set came from a single one, because postings from the other were rejected far more often by the test for a business to business technology employer. That is stated in the note rather than left for a reader to discover, along with the coverage that job boards cannot see at all, the fact that senior hiring often runs through search firms, and the plain point that an advert records what an employer says it wants, which is an imperfect proxy for what it screens on.