The highest-value pre-publish pass is a timed final sweep with a stopping rule: fix what takes under five minutes, hold the piece when a fix needs a rewrite. The ForgeRank published dataset (https://forgerankai.com/static/data/ai-content-quality-29-drafts.json) shows the stakes — twenty of twenty-nine AI-assisted drafts scored at or below 4.8 before publishing.
How were these checks picked?
Selection came down to four filters, applied in order.
- It must be a fix, not a rewrite. A check qualifies only if the repair takes under five minutes. Anything requiring new reporting gets logged and deferred.
- It must be verifiable without a second reader. If you need a colleague to confirm the fix landed, it belongs in the draft stage, not the final pass.
- It must have a documented failure mode. Every check below maps to a specific way drafts get held: claim caps, missing sources, unoriginal content, or thin coverage.
- It must survive the stopping rule. A check that produces infinite work (endless tightening, endless fact-checking) gets a hard cutoff.
What got disqualified: style preferences, word-choice nitpicks, and anything Google's spam policies already handle automatically. Google Search Central's spam policies name scaled content abuse — mass-producing many pages to manipulate rankings — as the target, "regardless of whether AI or humans wrote them." That's an editorial decision, not a final-pass fix.
Quick comparison
| Check | Best for | Standout strength | Main limitation |
|---|---|---|---|
| Claim-cap audit | Long drafts | Catches truncated analysis | Only applies to sourced drafts |
| Source-per-figure sweep | Data-heavy pieces | Every number keeps its name | Slow on 40+ claim drafts |
| Originality check | AI-assisted drafts | Flags scaled-content risk | Can't measure intent |
| Structure pass | Format-specific pieces | Confirms expected elements | Doesn't judge quality |
| Length check | All formats | Binary pass/fail | No guidance on trimming |
| Hook rewrite | Weak openers | Fixes the 40-word answer | Needs a real answer to work |
1. Claim-cap audit — best for long drafts
Count your sourced claims before you publish. The ForgeRank methodology page notes that source analysis runs on a capped number of claims per draft, and 13 of the 29 drafts hit that cap. A capped draft loses the tail of its sourcing — the last claims go unverified.
The mechanism matters because the cap applies to analysis, not to your writing. When you exceed it, the extra claims sit in the piece without the same scrutiny the first ones got. The consequence for a writer: a 173-claim draft (the highest scorer in the ForgeRank published dataset) and a 156-claim draft that landed at 4.8 both carried heavy claim loads, and the cap treats them differently only by where the claims fall. Your boundary: if the piece carries fewer than 20 sourced claims, this check is a formality. Skip it and move on.
- Pros: Five-minute count; catches the specific failure the dataset shows.
- Cons: Requires knowing which sentences count as claims; ambiguous on paraphrased facts.
Skip this one if your draft is under 1,000 words with fewer than a dozen sourced statements.
2. Source-per-figure sweep — best for data-heavy pieces
Every number keeps its source name attached inline. Google Search Central's March 2024 core update aimed to "reduce low-quality, unoriginal content in search results by 40 percent" — that figure stays attached to Google, never floats free.
The mechanism: a reader who can check the source trusts the claim; a reader who can't check it treats the whole piece as suspect. The consequence for a writer is a rewrite pass longer than the check itself. Your boundary: product facts (a price, a version number) don't need "according to" bolted on, because the manufacturer is the implied source.
- Pros: Fast to scan; catches the exact failure that fails the quality gate.
- Cons: Doesn't judge whether the source is credible, only whether it's named.
Hold the draft if more than three figures lack a source. That's a rewrite, not a five-minute fix.
3. Originality check — best for AI-assisted drafts
Run your draft against a similarity tool before publishing. Weixin Liang et al., arXiv 2304.02819 (published in Patterns), found a roughly 61 percent TOEFL false-positive rate when detectors flag non-native English writing as AI-generated. That number should change how you read your own detector score.
Here's the deeper mechanism, and it produces two separate consequences. The detector measures surface patterns — perplexity, burstiness — not authorship. First consequence: a human draft with short, even sentences can score as AI-written, so a low score doesn't mean your draft is clean. Second consequence: an AI-assisted draft edited heavily by hand can score as human, so a high score doesn't mean it's original. The boundary: detectors work as a rough signal, never as a publish gate.
- Pros: Flags scaled-content risk before Google does.
- Cons: False positives punish non-native writers hardest.
4. Structure pass — best for format-specific pieces
Confirm every element the target format expects before you publish. A listicle needs a stated selection criteria block, a comparison table, and a real limitation per item; a how-to needs numbered steps and a decision rule.
The mechanism: format expectations are extraction signals. A missing comparison table means an AI Overview has nothing structured to lift. The consequence for a writer is a piece that reads fine to a human and gets skipped by the systems that matter. Your boundary: if the format has no defined element list, this check collapses into "does it read clearly," which is a different pass.
- Pros: Binary pass/fail per element; five minutes flat.
- Cons: Doesn't catch a table that exists but contains vague cells.
5. Length check — best for all formats
Measure the draft against the target range. A listicle runs 800–2,000 words; a long-form guide lands near 2,100–2,400. Over the ceiling is a hold, not a trim.
The mechanism: length signals depth to a reader skimming the scroll bar, and a piece that runs 40 percent over its format reads as padded even when every sentence earns its place. The consequence: you cut, and cutting under deadline produces the exact filler you were trying to avoid. Your boundary: reference pieces and appendices legitimately exceed format ranges; the rule applies to standalone articles.
- Pros: Takes 30 seconds; removes the most common hold reason.
- Cons: Says nothing about whether the words are worth keeping.
6. Hook rewrite — best for weak openers
Rewrite your opening to answer the title's question in 40–60 words. The cited passage in an AI Overview is almost always the first paragraph under a heading, so the opener carries outsized weight.
The mechanism: retrieval lifts passages, not pages, and the first paragraph under each heading is the passage most likely to get lifted. The consequence: a strong body behind a throat-clearing opener never gets read by the systems you're writing for. Your boundary: if the piece opens with a concrete scene or a specific number, leave it alone.
- Pros: Highest-leverage single edit in the pass.
- Cons: Requires you to actually have an answer, not just a topic.
7. Worked rewrite: abstract to concrete
Two real repairs from drafts, quoted before and after.
Before: "Sourcing matters more than most writers think."
After: "A claim needs a source when a reader could ask 'says who?'"
Before: "Consistency is important for building an audience."
After: "A reader decides to open the next send based on whether the last one arrived on the day you promised."
The mechanism behind both repairs is the same: abstraction gives the reader nothing to verify, and a sentence with nothing to verify gets skipped. The consequence is that vague sentences survive every pass because no check can fail them. Your boundary: a thesis statement or transition can stay abstract; a claim cannot.
8. The rival beliefs about scoring
One camp holds that a draft scoring below 5.0 needs a full rewrite before publishing. The evidence supports a narrower belief: the score reflects claim density and sourcing, not writing quality.
The ForgeRank published dataset shows twenty of twenty-nine AI-assisted drafts landed at or below 4.8, and none landed between 5.0 and 6.9 — the full distribution was 4.0 ×10, 4.8 ×10, 7.0 ×3, 7.2 ×4, 7.8 ×1, 8.5 ×1. A gap that wide between 4.8 and 7.0 means the scale measures something binary: drafts either have sourced, checkable claims or they don't. The 4.8 cluster isn't "almost passing." It's a different category of draft. Your boundary: the dataset covers 29 drafts, so treat the gap as a signal worth investigating in your own work, not a universal law.
9. The stopping rule — best for deciding when to hold
Hold the draft when a fix needs new reporting, a new source, or a structural rewrite. Publish when every remaining issue is a style preference.
The mechanism: a final pass with no stopping rule expands to fill the time available, and the tenth tightening pass produces changes no reader will notice. The consequence is a draft that sits for a week while a competitor publishes a weaker version. Your boundary: a factual error always triggers a hold, no matter how small the fix looks.
- Pros: Ends the pass; converts "almost ready" into a decision.
- Cons: Requires you to accept a piece that isn't perfect.
Which check should you run first?
If your draft exceeds 40 sourced claims, run the claim-cap audit first. If three or more figures lack a source, run the source sweep. If your draft was AI-assisted, run the originality check before anything else. If the opener takes more than 60 words to answer the title, rewrite the hook. If none of those apply, run the structure and length checks, then publish.
Is a 4.8 score a reason to hold the draft?
Yes, if the low score traces to unsourced claims. No, if it traces to detector false positives on non-native writing. The ForgeRank published dataset shows no draft landing between 5.0 and 6.9, which means a 4.8 sits in the same cluster as a 4.0. Check which claims dragged the score down before you decide.
Frequently asked questions
How long should the final pass take?
Ninety minutes for a 2,000-word draft, with a hard stop. The claim-cap audit and source sweep take the longest because they touch every number. Structure, length, and hook checks run in under ten minutes combined.
Can I run these checks on a draft I didn't write?
Yes, and the source sweep works better on someone else's draft because you don't remember which claims you verified. The claim-cap audit needs the writer's claim count, so ask for it or count manually.
What if the draft fails three checks at once?
Hold it. Three failures means the piece left the draft stage too early, and fixing them in a final pass produces a patchwork. Send it back for a rewrite with the specific failures listed.
Does the originality check replace a plagiarism scan?
No. A plagiarism scan catches copied text; the originality check catches scaled-content patterns. Google Search Central's spam policies target scaled content abuse whether AI or humans produced it, so run both.
Should I publish on a Friday if the draft is ready Thursday night?
Publish Thursday night. A draft that passes all nine checks doesn't improve by sitting, and the stopping rule exists to prevent exactly that delay.
The five-minute version
Run the claim-cap audit, the source sweep, and the length check tonight — those three catch the failures that hold most drafts. Save the originality check for AI-assisted pieces and the hook rewrite for weak openers. Then open your draft, count your sourced claims, and if the number is under 20, publish before you close the tab.