How to Edit AI-Generated Content: A Four-Pass Protocol

TL;DR

• Marketers' top AI use case flipped from drafting to editing, 19% to 38%, per Averi's 2026 benchmarks.

• Run four ordered passes: facts (20 min), structure (15 min), voice (10 min), de-slop (10 min).

• Pass 1 verifies every number, name, claim, and link before you touch a sentence.

• Pass 4 hunts the AI tells that survive structural edits — opener clichés, hedging clouds, fake tricolons.

• A 1,500-word draft should cost about 55 minutes of editing, not 5.

The shift: editing is the job now

If you want to know how to edit AI generated content well, start by accepting what the job actually is in 2026. The drafting step is cheap. The editing step is the product.

Averi's 2026 State of AI Content Marketing benchmarks report shows the primary AI use case among marketers moved from drafting to editing, climbing from 19% to 38%. That's a doubling in the share of teams who treat the model as a first-draft engine and the human as the finisher. It tracks with what we see on the ground: the cost of a 1,200-word draft has collapsed, and the cost of a 1,200-word draft that's actually worth publishing has not.

The quality signal is moving in the right direction, slowly. The Content Marketing Institute and MarketingProfs 2026 B2B Content and Marketing Trends study reports that 58% of B2B marketers using AI for content creation say content quality has improved. Improved, not solved. The 42% who didn't see improvement are usually the teams shipping unedited drafts and hoping the model's median output beats their median writer's bad day. It rarely does.

So the job is editing. Below is the protocol we use internally and with clients on our content marketing engagements. Four passes, in order, with time budgets that hold for a 1,500-word draft.

Pass 1 — facts: verify every claim, number, name, and link

Facts first because every later pass becomes wasted motion if Pass 1 finds a fabricated statistic in paragraph two.

The rule is simple. Every number, every named source, every proper noun, every URL gets checked against a primary source before you read for flow. If the model cited a study, open the study. If it quoted a person, find the quote. If it linked a page, click the link and verify the page says what the draft claims it says.

In our client work, somewhere between 1 in 4 and 1 in 3 model-generated stats either don't exist, point to the wrong year, or misrepresent the source. That's not a model failure to be outraged about — it's a known behavior to plan around. Plan around it by treating Pass 1 as non-negotiable.

The Pass 1 checklist:

1. Every numeric claim has a source link and a date.

2. Every named source resolves to a real document or person.

3. Every URL returns a 200 and matches the claim around it.

4. Every product name, person name, and company name is spelled correctly.

5. Anything that can't be verified gets cut or hedged explicitly ("about," "in our experience," "roughly").

If a fact survives Pass 1, it's load-bearing. If it doesn't, it's gone before Pass 2 starts.

Pass 2 — structure: cut the throat-clearing, reorder for the reader

Models open with throat-clearing. They restate the prompt, set up the topic, ease in. Human editors who've read enough drafts can usually find the real first sentence somewhere around paragraph three.

Pass 2 is structural, not stylistic. You're asking three questions. Does the piece answer the reader's question in the first 150 words? Are the sections in the right order for someone scanning, not someone reading linearly? Is there a section that exists only because the model wanted symmetry?

Most AI drafts have one section that can be deleted whole. Find it. Delete it.

The reorder usually goes like this. The conclusion paragraph the model wrote becomes part of the introduction. The third H2 — almost always the most concrete one — moves up to second. The "what is X" section gets compressed into a single sentence or cut entirely if the reader already knows what X is. Snippet-eligible answers move directly under the H2 that asks the question, not buried three paragraphs in.

If you're working from a brand voice doc, this is the pass where you check structural conventions: do H2s match the house style, are TL;DRs where they should be, is the close where the reader expects it. We keep our conventions in a separate brand voice guidelines document for AI content so editors aren't pattern-matching from memory.

Pass 3 — voice: dials, lexicon, and the banned-phrase list

Voice is where most teams stop too early. They read the draft, it sounds "fine," they ship. Fine is the problem. Fine is the median model voice that 100,000 other content teams are also shipping.

Pass 3 has three sub-steps and they're worth doing in order.

First, run the banned-phrase pass. Every team needs one. Ours blocks the obvious tells: synergy, in today's landscape, empower, unprecedented, effortlessly elevate. If your team doesn't have a banned list, build one this week from the last five model drafts you read — the repeat offenders will name themselves.

Second, calibrate the dials. We think of voice as a set of sliders: technicality, swagger, empathy, urgency, authority. A founder essay for LinkedIn isn't the same dial setting as a programmatic SEO page. Pass 3 asks: is this draft at the dial setting the channel needs, or is it stuck at the model's default (medium-everything, opinion-light, authority-by-adjective)?

Third, the lexicon swap. Models reach for abstract nouns; operators use concrete ones. Solution becomes the system that routes inbox triage. Optimize becomes cut the response time from 4 hours to 11 minutes. If a sentence works with or without its adjectives, the adjectives are decoration and the sentence is undercooked.

Pass 4 — de-slop: the AI writing tells that survive passes 1–3

Pass 4 is the one most editors skip and most readers notice. Even after facts, structure, and voice, model drafts carry a residue. The ai writing tells that survive are subtle and they accumulate.

The checklist:

• Opener clichés. Any paragraph starting with In today's, Imagine, Picture this, or a rhetorical question the next sentence answers. Cut or rewrite.

• Hedging clouds. Stacked hedges (it could potentially be argued that perhaps) that make the sentence vibrate without saying anything.

• Fake tricolons. Three-item lists where the third item is filler invented to complete the pattern. Read every list of three and ask if the third is real.

• Em-dash overload. More than one em-dash per 250 words is a tell. Convert most to periods or commas.

• Smooth-talker connectives. Moreover, furthermore, additionally, in closing — none of these earn their place. Cut or replace with a hard sentence break.

• "X, not Y" antithesis. Powerful once per piece. Tiresome by the third use. Keep one, cut the rest.

• The bolded thesis on every section. If every section has the same shape, no section has weight. Vary it.

One more thing worth flagging: don't run the output through an AI "humanizer" tool as a substitute for Pass 4. Those tools optimize against detector signatures, not against reader experience. We've written separately on why AI detector false positives make the whole detection layer unreliable as a quality bar. De-slop is a reader-experience pass, not a detector-evasion pass.

Time budgets: what each pass should actually cost

For a 1,500-word draft, here's the budget we hold ourselves to. Adjust proportionally for longer or shorter pieces.

Budgets hold for a 1,500-word draft. Scale proportionally for longer pieces.

Notice Pass 1 gets the most time. That's deliberate. The cost of shipping a fabricated stat is asymmetric — one wrong number can cost more reader trust than ten clean paragraphs earn. Pass 2 is the second-most expensive because structural surgery is slow and Pass 3 and Pass 4 only work on a draft that's structurally sound.

If any pass surfaces a problem that requires a real rewrite, kick the draft back to the revise queue rather than fixing it inline. Inline fixes during a later pass corrupt the pass discipline — you start mixing voice edits with factual rewrites and you lose track of what's been checked.

Before and after: a worked example

Here's a 78-word paragraph from a real model draft we edited for a client in March 2026. The topic was inbox triage for a 12-person services firm.

Before (model draft): In today's fast-paced business environment, organizations are increasingly turning to AI-powered solutions to streamline their workflows. By embracing unprecedented agentic systems, teams can effortlessly automate repetitive tasks and achieve unprecedented productivity gains. Studies have shown that companies adopting these technologies experience significant improvements in efficiency, allowing them to focus on higher-value strategic initiatives. This unprecedented shift represents a fundamental change in how modern businesses approach their daily operations.

Pass 1 cut the unsourced "studies have shown." Pass 2 deleted the throat-clearing opener. Pass 3 swapped abstract nouns for the actual mechanism. Pass 4 removed the effortlessly, unprecedented, and unprecedented tells.

After (40 words): The client's inbox was sitting at 4.2 hours from receipt to first reply. We wired a triage agent to their CRM with a human approver above any client-impacting decision. First-reply time moved to 11 minutes inside three weeks.

Same topic. Half the words. A specific number, a named mechanism, an honest time-frame. That's the bar. The reader can act on the second version. The first version is wallpaper.

FAQ

How long should it take to edit an AI-generated blog post?

For a 1,500-word draft, budget about 55 minutes across four passes: 20 minutes on facts, 15 on structure, 10 on voice, 10 on de-slop. Longer or more technical pieces scale up proportionally. If you're spending less than 30 minutes editing a model draft, you're almost certainly shipping fabricated stats or median-voice prose.

What is the most important step in editing AI content?

Fact-checking, by a wide margin. Models generate plausible statistics that don't exist, misattribute quotes, and link to pages that don't say what the draft claims. Every number, name, and URL needs verification against a primary source before any stylistic editing happens. A well-written paragraph with a fake statistic is worse than a clunky paragraph with a real one.

What are the most common AI writing tells to look for?

The recurring tells: opener clichés like In today's, stacked hedges, three-item lists where the third item is filler, em-dash overload (more than one per 250 words), connective slop like moreover and furthermore, and the X, not Y antithesis used more than once per piece. A focused 10-minute de-slop pass catches most of them.

Should I use an AI humanizer tool instead of manual editing?

No. Humanizer tools optimize against detector signatures, not against reader experience. They produce prose that fools a classifier and bores a human. Manual editing — especially the structural and voice passes — is what makes a draft worth a reader's time. Detection-evasion is a different problem from quality, and conflating them ships bad content.

Has AI improved content quality for most marketing teams?

Partially. The CMI/MarketingProfs 2026 B2B Content and Marketing Trends study reports 58% of B2B marketers using AI for content creation say quality has improved. That leaves 42% who haven't seen gains — typically teams shipping unedited drafts. The quality lift correlates with editorial discipline, not with model choice.

What to do this week

Pick your next AI-drafted post. Run the four passes in order with a timer. Note where the budget breaks. That single audit will tell you which pass your team is skipping — and that's the one to systematize first.

If you want a second set of eyes on your editorial protocol, start a conversation here.