Ad checks that people do not learn to ignore

The failure mode for any automated ad check is not missing things. It is training the team to click past it. A tool that flags sixty trivial items per variant ends up ignored on the five that carry account risk. An ad pre-check has to earn its place on every run, which means being deliberate about what it blocks on, what it merely warns about, and what it refuses to judge at all. The design constraint is attention, not coverage.

Block on what is certainly wrong

Some constructions are facts, not opinions. A "you + personal condition" sentence is restricted framing under Meta's advertising standards regardless of the product, the market, or the account's history. An outcome guarantee with a number and a deadline — "double in 30 days" — trips promise rules on every major platform and invites the substantiation demand the FTC spells out for health and earnings claims. These checks never produce false positives about whether the pattern is present; the only honest uncertainty is how the platform will weigh it. Block here, before submission, because these are the themes that escalate into account-level review.

Warn on what is probably wrong

Patterns that policy guidance agrees on, but that sensible advertisers sometimes break on purpose, belong in the middle tier. Sensational hooks are the classic one: "shocking," "they hate this," stop-scrolling commands sit inside the sensationalized-content grey zone — the same phrase passes on a mature account and flags on a new one. Urgency language behaves the same way: a real deadline stated with a date is fine; "act now, ends tonight" on an evergreen offer is the pressure tactic regulators call deceptive (the UK's ASA has repeatedly ruled on fake countdowns). Profanity, all-caps, exclamation density: delivery-limiting signals, not violations. Warn, show the count, let a human decide — context the machine cannot see often justifies what the pattern alone condemns.

Leave alone what cannot be decided mechanically

Whether the claim is substantiated — that depends on evidence that lives outside the text. Whether the landing page matches the ad — the check never sees the page. Whether the account's history will make a reviewer suspicious — invisible from a copy paste. Whether the offer is a "prohibited category" question at all — gambling and financial-product rules are licensing questions, not phrasing questions. A tool that pretends otherwise produces confident noise, which is worse than silence: it burns the credibility you need for the checks that genuinely know things. This is also why an honest pre-check refuses to output an "approval probability." Nobody can compute it, and pretending otherwise is the check becoming the deceptive pattern it flags.

A workload that stays sane

Teams that keep an ad check alive run it on a fixed, boring cadence: one scan per variant, before the first submission, never after the campaign calendar is committed. The scan is fast precisely because it is mechanical — pattern-match the published themes against the text and it takes seconds and produces the same ranks every time. What the cadence protects is the reviewer's attention: if the block tier is already cleared by the time a human reads the variant, the human spends their effort on the judgment calls the tool deliberately refused to make. That division — machine on patterns, human on intent and evidence — is the only arrangement that has held up in practice. Teams that reverse it (humans hunting attribute framings, tools pronouncing on substantiation) produce exhausted reviewers and untrustworthy scores.

What to measure instead of compliance theatre

A single pass/fail verdict is the enemy: it hides the one finding that matters behind fifty that don't, and it invites bargaining ("it passed before"). Named, tiered findings survive disagreement — someone can argue with a rank, but nobody can argue with a quoted sentence and a named theme. If you must report a number, report the count of open HIGH findings, because that is the only figure with an obvious action attached: zero, or the variant does not submit. Rejection-velocity dashboards and "policy health scores" sound like governance and function as decoration; the fix checklist, closed line by line, is the actual control.

Tiering is the product

What teams actually want from an ad check is not more findings; it is a trustworthy ordering of them. Severity tiers — account-risk framings first, claim-sensitivity warnings second, tone noise last — are what keep the check on the desk instead of on the mute list. AdPrecheck separates these tiers rather than collapsing them into one verdict, because a single undifferentiated signal is what gets muted. And every tier stays reviewable: quoted evidence, named theme, suggested edit — nothing silently rewritten.

Decision-support only: findings quote your copy and published policy themes. You make the call before anything submits, and the platform makes the final one.