A homepage told me "Not in a log. Not in a profile. Not for sale." The same page loaded an ad-conversion tag and an analytics beacon that recorded my browser, my country, and every URL I visited.
The contradiction surfaced in an outside-in audit of a consent-first AI startup. The company is beside the point, so it stays unnamed. The distinction the audit forced applies to every trust claim in the industry.
Two registers
Every company that sells trust produces two bodies of text. The marketing speaks to the prospect. The operating documents, the privacy policy, the terms, the security page, speak to the lawyer and the regulator. Most people read one or the other. The audit method is to read both, build a concept map of each, and measure what each contains that the other lacks.
Both registers describe the same product, so in an honest company they mostly agree. Where they diverge, the divergence is the finding.
Consent as copy
Consent-as-copy has a recognizable sound. It speaks in absolutes: nothing collected, nothing tracked, no algorithm, nothing to target. Absolutes are effective marketing because they require no explanation. They are also almost always false, because any product that matches people to services holds data about those people somewhere.
The startup claimed "there's no targeting algorithm because there's nothing to target with." The product's core feature was matching one person's need to another person's offer, which requires needs, offers, and identities on a server. The claim described the business model, no ads, in the language of architecture.
Consent as architecture
The same company's privacy policy contained the real thing, and the real thing sounds completely different. Training use of your content requires a separate recorded opt-in, off by default. Consent records are enforced append-only at the database level, so no later code change can quietly rewrite what you agreed to. Deleting your account deletes the data, with three exceptions, listed. One sentence stands out: once your content has been included in a completed training run, withdrawing consent stops all future use but cannot reverse the past run.
That sentence is what consent-as-architecture sounds like. It's specific, it's limited, and it admits what the mechanism cannot do. No copywriter produces that sentence, because it makes the product sound weaker. Engineers produce it, because it's true.
The test
Here is the test to run on any company selling trust, ours included.
Read the marketing and the operating documents as separate texts. List every trust claim in the marketing and ask, for each: which mechanism in the operating documents delivers this? Then reverse it: list every mechanism and ask which ones the marketing mentions.
The first gap, claims with no mechanism, is the overselling. It predicts behavior under pressure, because a claim with no mechanism costs nothing to break. The second gap, mechanisms with no claim, is stranger and more hopeful: some companies build real protections and never say so, because the specific truth sounds weaker than the round absolute.
The absolute is the tell. A company that means it writes the limited sentence.