The fix order matters more than the fix. Across the 29 drafts in the ForgeRank published dataset (https://forgerankai.com/static/data/ai-content-quality-29-drafts.json), phrasing never moved a draft up a band on its own, while source compliance alone was worth three tiers. Strip the em dashes first, then rebuild the substance underneath.

Who this is for

This guide assumes you already have a draft and a way to read it cold. You need a text editor with find-and-replace, a word processor that shows sentence length, and roughly 90 minutes. No writing credentials required. By the end you will have a draft that survives a phrase scan, carries at least 150 sourced claims, and reads like one specific person wrote it. The steps run in a fixed order because each one removes a cause the next one would otherwise hide.

What do you need before you start?

Step 1: Count your phrasing hits before you change a word

Run a find-and-replace count for each phrasing family you suspect. The 29 drafts recorded 594 phrasing hits across 23 distinct families, and em dashes alone accounted for 182 hits spread over 19 drafts, per the ForgeRank published dataset. The most-flagged single draft carried 73 hits.

Counting first matters because the number tells you whether you have a phrasing problem or a substance problem. Eight of the nine drafts scoring 7.0 or above still received a heavy-AI-phrasing verdict, per the same dataset. A clean phrase scan does not mean a good draft; the only draft whose phrasing came back clean scored 4.0.

Write the counts down on paper. You now have a baseline, and you can confirm the scan worked because the total matches the number of hits you can see highlighted in the text.

The trap here is fixing phrasing before you know the score. A writer who strips 73 hits from a draft with 40 claims has produced a cleaner draft that still fails.

Step 2: Cut the phrasing families, not the sentences

Delete or rewrite each hit from your count, working family by family. Em dashes get replaced with periods or commas. Ordinal openers like "The first" and "The second" get replaced with the concrete action that follows them. Hedges like "usually," "often," and "tends to" get replaced with the specific case they were hiding.

This step works because phrasing families are measurable and finite. Twenty-three families is a closed list you can clear in one pass, which is why it goes second: it is the cheapest edit in the whole sequence and it removes noise that would otherwise mask the substance problems underneath.

When you finish, re-run the count. Zero hits on your original list means the scan is clean.

The boundary: this step stops applying when a phrasing family is doing real work. A numbered list of five steps needs its numbers. Cut the tic, keep the structure.

Step 3: Replace every hedge with a named instance

Go sentence by sentence and find each claim that could apply to anything. Then name the thing. A sentence about writers hedging becomes a sentence about a draft that scored 6/10 for the same habits. A sentence about teams shipping thin content becomes a sentence about the 4.0 drafts in the dataset that all scored 4.0 across all four dimensions.

This is the dimension that moved the top of the set. The 0.8 that separates the 4.0 drafts from the 4.8 drafts is source compliance alone, worth three tiers at a 25 percent weight, per the ForgeRank methodology page. Information gain, specificity, and argument depth stayed at their floor across all twenty of those drafts.

The consequence for you: specificity is the highest-leverage edit available, and it is also the slowest. Budget most of your 90 minutes here.

The boundary: specificity stops paying when the instance is decorative. A named brand that does not change the claim is noise with a logo on it.

Step 4: Attach a source to every figure

Every number in the draft gets a source name inline. Not a footnote, not a link at the bottom. The name sits in the sentence.

Mechanism first: an unattributed figure is a claim about a third party that you cannot defend. A draft that lands at 4.8 carried 156 claims, and the highest-scoring draft in the set carried 173, per the ForgeRank published dataset. Claims without sources are not claims; they are liabilities.

For you, this means the drafting process changes. You cannot write the sentence until you have read the source. That constraint is the point.

The boundary: product facts you can verify yourself do not need a "according to" bolted on. A price you checked on the vendor's own page is its own citation.

Step 5: Break the sentence rhythm

Read the draft aloud and mark every run of three sentences with the same length and the same opening word. Rewrite them. Google Search Central states that its helpful content systems reward content "written by people, for people" rather than content "created primarily for search engines" — a rhythm test is the fastest way to hear which one you produced.

Uniform rhythm is the acoustic signature of generated text. A reader cannot name it, but they stop trusting the page at roughly the third identical clause.

Vary deliberately: one sentence of six words, one of twenty-eight, one that opens with a verb. Read it again. The paragraph should sound like someone talking, not someone formatting.

The boundary: technical sequences stay uniform on purpose. Installation steps and API call orders should read the same way every time.

Which belief about AI writing is actually true?

The common belief is that AI writing sounds robotic because of the words. The evidence supports a different cause: the words are a symptom, and the substance underneath is the disease. The ForgeRank published dataset settles this. Eight of the nine drafts scoring 7.0 or above still carried a heavy-AI-phrasing verdict, and the only clean-phrasing draft scored 4.0. If vocabulary were the cause, the clean draft would have scored highest. It scored lowest. The drafts that cleared 7.0 did so on source compliance and claim count, not on vocabulary.

What if it doesn't work?

The draft still reads flat after all five steps. Check claim density first. If you are under 150 sourced claims, the draft is thin and no phrasing edit will fix that. Add instances before you touch another word.

You cannot find sources for half the figures. Drop the figures. Rewrite the sentences to carry the mechanism instead of the number. An unsourced statistic is worse than no statistic.

The phrase scan keeps finding new families. You are editing in the wrong direction. Stop rewriting sentences and start deleting them. A draft with 40 strong claims reads better than one with 80 weak ones.

The rhythm pass made it worse. You are varying length without varying structure. Change what the sentence is doing, not just how long it is. A question, then a fact, then a consequence.

The draft scores well but you do not believe it. Read it as a stranger. If you cannot name the person who wrote it, the voice is still missing, and no dataset will tell you that.

How long does this take?

Ninety minutes for a 2,000-word draft, with roughly 50 of those minutes in Step 3. The phrasing passes in Steps 1 and 2 are fast because they are mechanical. Source attachment in Step 4 is slow because it depends on reading, not editing. Budget the time before you start, and do not compress Step 3 to hit the deadline.

Finish checklist

  • Phrase count is zero on your original family list
  • Every figure carries a source name inline
  • Claim count is above 150
  • No three consecutive sentences share a length and opening pattern
  • You can read the first paragraph aloud without wincing

Frequently asked questions

Does using AI to draft automatically make the content low quality? No. Google's guidance targets content created primarily for search engines, not content created with AI tools. The measured signal in the ForgeRank published dataset is claim density and source compliance, both of which a human editor controls regardless of how the first draft was produced.

Which fix should I do first if I only have 20 minutes? Step 1. Counting phrasing hits takes five minutes and tells you whether you have a vocabulary problem or a substance problem. Most drafts have both, and knowing the ratio tells you where the remaining 15 minutes should go.

Is 150 sourced claims realistic for a 1,500-word piece? That works out to roughly one claim per ten words, which is dense. The 4.8-scoring draft in the ForgeRank published dataset carried 156 claims, so the target is achievable. Cut word count before you cut claims.

Can I run these steps in a different order? You can, but Step 3 before Step 2 means you will spend time adding instances to sentences you are about to delete. The order exists to prevent wasted edits, not because the steps are dependent.

What counts as a claim? Any sentence a reader could check. A named price, a version number, a date, a measured result, a stated rule. Opinions and transitions do not count toward the 150.

You can now take a draft that reads like a template and rebuild it into something a named person would sign. The next move is small: open your current draft, run a find-and-replace count for em dashes, and write the number at the top of the page before you change anything.

Related reading: Why does AI writing sound fake? · AI content review · What 29 scored drafts revealed · How to review AI-generated articles
Working on a draft right now? You can run any piece through the same 4-dimension quality read before it ships. It is free, no signup, at forgerankai.com.