No invented links or facts: how the quality gate works

The most common complaint about AI writing is not tone. It is invented detail: a link to a page that does not exist, a statistic with no source, a feature your product never had. Serpio's answer is structural rather than hopeful. Links can only point at pages in your index, every claim is checked against recorded evidence, and nothing ships without a passing report. This guide walks through the gate in the order the pipeline applies it.

Internal links come from the site index

The writer is not allowed to make up a URL. During the Research step, a link picker builds a menu of candidate pages from Serpio's private site index. A page qualifies only if it is on your own domain, has finished indexing, is not paused or removed, is not set to background context only, and has not passed its review date. Legal pages, login pages, and the blog index are excluded because they make poor editorial links.

Candidates are ranked by how closely they match the brief, using the page's meaning, its keywords, and its role: product and feature pages, conversion pages such as pricing, and support pages get a small boost because they are useful destinations. The writer receives at most twelve, and the draft's links are validated against that list before the article is saved. Uploaded documents are never link targets.

Every article carries an evidence bundle

Alongside the links, the job assembles the passages the writer may rely on: retrieved knowledge with facts permission, up to two background-only passages for context, and public sources from bounded research. Half the slots are reserved for your own knowledge so public pages cannot crowd it out. Sources on your blocked list are dropped. The bundle is recorded once and stays with the article.

You can inspect it on any article. Public sources lists external pages with a URL and excerpt. Knowledge used lists your own material, labelled Website knowledge, Knowledge document, or Saved brand profile. Private identifiers such as filenames, storage paths, and signed download links are redacted from every view and replaced with a placeholder, so evidence can be shared without leaking where a file lives.

Quality review outcomes

In the Finish step, a reviewer checks each claim in the draft against the bundle and produces a report with one of three outcomes.

Unsupported

A claim with no passage in the evidence bundle behind it.

Contradicted

A claim that disagrees with the evidence, for example a price or limit stated differently in your own docs.

Fabricated

A citation, quote, or detail that does not exist in any retrieved source.

Private

Text that reproduces a source marked background only, or a private identifier.

Outdated

A claim that relies on a source past its review date or marked stale.

Unsafe

Content that fails the safety review.

Incomplete

A required element is missing, such as a title, or the check itself could not finish and must be retried.

One category is deliberately advisory. Where a citation is placed relative to a sentence is an editorial suggestion, not evidence that the sentence is false, so placement notes never block on their own.

Editing reruns the checks

1. 1

   You edit and save

   The article becomes a new revision. Its status resets to needs review and a fresh check is queued for that revision.

2. 2

   The checks run against the new text

   The reviewer uses the same evidence bundle. If you added a claim the bundle cannot support, it is flagged.

3. 3

   You approve

   Approval, which sets the article to Ready to publish, requires a passing report for the current revision. Trying early returns the message that article checks are still needed for this revision.

4. 4

   Export or delivery checks again

   Guided exports and connected destinations assert the same condition before they send anything, so a report cannot be bypassed by publishing from a different screen.

What the gate does not do

• It does not verify claims that have no evidence. It flags them as unsupported and stops. It never fills the gap with a guess.
• It does not re-check older articles when your Knowledge Base changes. Each report belongs to the revision and evidence it was written against.
• It does not judge whether an article is interesting, well argued, or right for your audience. That is your job as the reviewer.
• It does not replace your own reading. A passing report means the text is consistent with the evidence, which is a floor, not a ceiling.

Put together, the gate makes one promise Serpio can actually keep: nothing in a published article points somewhere that does not exist or claims something the evidence does not support. For the wider argument about quality and search engines, read Will Google penalise AI-written content?.