AI Content Plus Human Editing: The Formula That Ranks

The does-AI-content-rank argument is over, and the data that ended it points somewhere more useful: at the edit. AI-assisted pages rank routinely. Pure AI pages rank occasionally. And the gap between those two outcomes is human editing, which the ranking studies and the practitioner surveys now agree on with unusual consistency. This post is my reading of that evidence and the five-pass edit I run to be on the right side of it - the workhorse companion to the E-E-A-T question in my AI SEO playbook.

What the ranking data actually shows

Two large studies this past year, from vendors with opposite incentives to flatter AI, landed on the same shape. Ahrefs analyzed a million pages from the top 10 across 100,000 SERPs in June 2026: 5.3% of top-3 results were fully AI-generated - so pure AI does rank - but 82.2% of top-3 pages were less than half AI, and pages with low or moderate AI content pulled two to three times the organic impressions of heavily-AI pages. Ahrefs’ own conclusion was that Google punishes poor quality, not AI provenance, and heavy AI content is simply poorer on average.

Semrush’s study, across 42,000 ranking blog posts in November 2025, is starker at the very top: human-written content took position 1 with 80.5% probability against 10% for AI-generated, while from position 5 down the gap narrows to nearly nothing. The same study surveyed 224 SEO professionals and found 72% believe AI content ranks as well or better - a perception gap I would frame this way: AI content competes fine for the middle of page one. The top of it still belongs overwhelmingly to work a human shaped.

My read of both studies together: the ranking penalty attaches to unedited-ness, not to AI. Which happens to be exactly Google’s stated position - appropriate use of AI is fine, and what the systems reward is helpful, original content whatever produced the first draft. The policy detail lives in my breakdown of what Google actually says.

The finding that names the formula

The sharpest evidence for the edit itself comes from Orbit Media’s survey of 1,042 content marketers. Only around 14% of bloggers report strong results at all this year - and the marketers with a formal human editing process beat that benchmark by nearly two to one, the strongest predictor in the dataset. The same survey’s grim mirror image: among editing approaches, having AI do the editing scored lowest for strong results.

I want to give that last finding a sentence of respect, because it kills the tempting cheap version of this formula. Running a draft through the model again with “improve this” is not editing - the model optimizes toward its own idea of good prose, which is the sameness that made the draft mediocre. Editing that moves rankings adds the things the model structurally lacks: verified facts, lived experience, a point of view, and knowledge of your reader. A second AI pass adds none of them. It just launders the tells.

The five-pass edit I run

My edit is five passes, in a fixed order, and naming them keeps the work honest when the deadline argues for skipping one. I hold the order because each pass protects the ones after it.

Pass 1: claims. Every fact, number and named source gets verified or cut, using the fact-checking prompts to extract the claim list and my own reading to settle it. This pass is first because nothing else matters if the piece lies.

Pass 2: experience. I add what only I can: a first-person judgment, a real observation, a rule I actually follow, a place I disagree with the consensus the model reproduced. This is the E-E-A-T pass in the literal sense, and the one that separates a page from the forty others the same model drafted for other people this week.

Pass 3: structure for retrieval. Sections reworked so each opens with its answer and stands alone - the discipline from passage-level citability. It serves human skimmers and AI answer engines with the same edit.

Pass 4: voice and de-sameness. The tells go: the inflated adjectives, the “not just X but Y” reflexes, the em-dash habit, whatever your banned list holds. Sentence rhythms get varied. This is the pass a reader never notices and always feels.

Pass 5: subtraction. The model pads, and the last pass cuts: the restated headings, the summary nobody needs, the third parallel example. My drafts typically lose 15 to 25% of their words here, and the piece is better by roughly that fraction.

What this costs, and why the math still works

The honest cost accounting: a real five-pass edit on a substantial draft takes me one to two hours, which sounds like the productivity gain evaporating. It is not, and the arithmetic matters. The draft that took a day now takes an hour, the edit costs the same hour it always should have cost, and total time per publishable piece drops by more than half while quality holds. The version where the gains evaporate is the one where teams count the edit as optional overhead and then wonder why traffic sank - the pattern the Orbit numbers describe at industry scale, with 92.4% adoption and strong results at an all-time low. My accounting rule: the edit hour is part of the cost of the piece, tracked as such, which is exactly how my content ROI method books it. Pieces where the edit costs more than the draft saved are pieces I should not have pointed a model at.

Where the edit sits in the workflow

Placement matters as much as the passes. The edit runs once, after the draft is complete, against the brief - not as a running commentary during generation, which produces the endless-tinkering failure mode. Upstream of it, a real brief does half the edit’s work in advance by giving the draft a thesis, a source list and a structure to be checked against, which is why the brief workflow is the highest-leverage document in the pipeline. And at volume, the edit is the gate that decides what publishes at all: in my prompt chaining pipeline, a draft that needs more than the standard edit gets rebriefed or killed rather than heroically rescued, because rescue editing costs more than writing.

FAQ

Does Google penalize AI content? No - Google’s documentation says appropriate AI use is fine, and Ahrefs’ million-page study found fully AI pages in top positions. What the data shows instead is a quality correlation: heavily-AI pages rank worse and earn a fraction of the impressions, because most of them are unedited. The penalty is for being mediocre, which AI makes easy at scale.

How much should I edit AI content for SEO? Until all five passes hold: claims verified, something only you could say added, sections that stand alone, the tells gone, the padding cut. As a crude gauge, if the published piece is not meaningfully different from what the model handed you, you have proofread, not edited, and the Semrush position-1 numbers say proofread AI is exactly what fails at the top.

Can I use AI to help with the editing itself? As an instrument, yes: extracting the claim list, flagging banned phrases, checking sections against the brief. As the editor, no - the Orbit data has AI-edited content performing worst of all approaches. The distinction is who decides. AI can surface what to look at. A person decides what is true, what is missing and what dies.

Is fully human content worth the cost now? For the pages where position 1 is the business case - your money pages, your flagship guides - the 80.5% figure says human-led is still the winning bet, with AI as research help. For the long middle of a content program, AI draft plus real human edit is the honest optimum the data supports.

This post is part of my AI SEO playbook. The policy backdrop is Google’s AI content policy, the expertise signals live in E-E-A-T for AI content, and the verification pass runs on the fact-checking prompts.