A pre-publish gate AI-assisted content has to clear

Keep the detector out of the gate. A 2023 Stanford evaluation found that GPT detectors flagged more than half of TOEFL essays written by non-native English speakers as machine-generated, while flagging almost none of the essays written by US eighth graders. If anyone on your team writes in a second language, a detector threshold is a policy for punishing them. Google, meanwhile, has never asked how a page was produced. Its spam policy asks whether pages were made at scale mainly to rank, which is a question about intent and repetition. A gate that survives contact with an editorial team therefore checks three other things: whether the page contains anything not already on the results page it targets, whether every number in it resolves to a source, and whether a named human is willing to be wrong in public about it. Where it sits: In the publishing path, before a page can go live, never in a monthly audit afterwards. In a Git-backed stack that means a required status check on the pull request. In Sanity, Contentful or WordPress it is a webhook on the draft-to-review transition that can refuse the transition. Refusal is the property that matters. A dashboard reporting quality is not a gate, and every team that has one ships the page anyway. What each page has to clear: Claim attribution. Every statistic, date, price and legal or clinical assertion carries a source URL in a field, and the check confirms the URL returns 200 and still contains the claim. Corpus duplication. The draft is embedded and compared against your own published archive, so near-identical drafts fail instead of becoming the fourth page you own on one question. Overlap with the results page. The current top ten for the target query are pulled and the draft's substantive claims are compared against them. A page adding nothing new fails regardless of who wrote it. Named ownership. An author with a checkable footprint, not a persona, recorded in the CMS and in the page's structured data. Reviewer routing. Pages touching money, health or law go to a named reviewer and cannot skip that route. Thresholds are yours, and setting them is most of the work. I bring the defaults I use, then argue with your editor about them for a week, because a gate calibrated by an outsider gets switched off in month two. What happens when a page fails: It stays in draft with a named reason and a named owner. There is an override, because a gate with no override gets ripped out entirely, but the override is signed, logged and reported monthly with the name attached. The first page your own CMO wants pushed through is the test the gate either survives or does not, so I write the escalation path before that happens rather than during. I do not remove a check because it fired correctly. Evidence it worked: Not a detector score. The number is the share of pages published in a quarter that have received an organic click within ninety days. Teams running a real gate publish less and watch that share rise, because the filler pages never existed to dilute it. I report it monthly against a baseline captured before installation. If Search Console history is missing, the baseline cannot be honest and I will say so at the start rather than invent one later. Boundaries: I do not write your content, edit your archive or supply writers. I do not license detection software or recommend a vendor. I do not rewrite pages that fail; the gate names the reason and your team fixes it. Cleaning up an existing library is separate work with a different shape, and I will say so rather than let it eat this scope quietly. Turn this down if: Your plan is to publish at a volume no editor could read, and what you want is a check that lets it through. You want a document for a client or an auditor rather than an enforcement point in a pipeline. Or your editorial lead has not agreed that the gate can block their own work. In that last case the gate does not exist yet, and no amount of configuration will create it.