Two structural points that matter more than the wording. First, the ledger holds no personal data: there are no candidates, no applications, no CVs and no jobseeker accounts, so most of what a privacy policy usually has to cover simply does not arise. Second, removal is self-service and verified — a site owner does not have to email anybody or wait for a human to act.

Consent decisions are recorded, so there is an auditable record of what was chosen and when, rather than only a cookie in your browser. What the analytics provider does and does not receive is stated explicitly on /privacy rather than left to the provider's own documentation.

What is sent to a model is the text of a public job posting and a company name — the same words the employer published on their own careers page. No customer data: your account, your API key, your queries and your usage are never used as model input, never used to train anything, and never shared with a model vendor. No candidate data, because none exists.

The disclosure also states where the classification is weakest, which is worth reading if you are building on category or seniority: on a clear job title they are dependable, on a vague or invented one they are a guess, and the marker says inferred either way because the row cannot tell you which case you have. The full page, including which models are used for which pass, is at /ai.

The reasoning behind the asymmetry, since it is a deliberate choice rather than a reflex: a job page ranking in a search engine sends a human here, and that human can see a price. An answer engine that ingests the same page answers the question inside its own product and the visit never happens. Handing two and a half million rows to a model vendor for free, while selling the same rows through the API, is selling something you are also giving away — to the buyer least likely to come back. What answer engines keep is deliberately generous: enough to describe the dataset accurately, cite it, and tell somebody where to get it.

/*? — the faceted query space — is refused to every crawler. The home page exposes country, city, category, remote and employment as query parameters, so the number of distinct URLs is combinatorial. One crawler enumerating them generated 1,562 requests in three minutes, 97% of them a query string never seen before, so every one missed the page cache and forced a fresh render and a fresh query; that single pattern held the server at 96% CPU for as long as it continued. Nothing of value is lost — a filtered view is duplicate content whose canonical is the bare page, and every individual job and company already has its own indexable URL in the sitemap.