A client rang me last year, genuinely upset. He'd paid a content agency £4,000 for sixty blog posts, got them back in three weeks, published the lot, and watched his organic traffic drop 22% over the following two months. When he forwarded me the posts, I could spot the problem inside thirty seconds. Every single article opened with a rhetorical question. Every section used "Furthermore" or "In addition" as a transition. The conclusions all started with "In conclusion." It was GPT-4 output, barely touched, bulk-published across a site that Google had previously trusted.
This happens constantly. And the frustrating thing is it's not the AI's fault. GPT-4, Claude, Gemini, they're genuinely impressive at generating structured prose. The failure is in how people treat the output. Raw generation is not a finished product. It never was.
What "Raw AI Content" Actually Looks Like
You've seen it. You might have published some. I have, early on at Seahawk, when we were stress-testing workflows in 2022 before we knew better.
Raw AI content has a specific texture. Sentences are remarkably uniform in length. The vocabulary is wide but oddly flat, lots of technically correct words that nobody actually says. Transitions are formulaic. The structure is almost always: intro paragraph, three to five H2s each with two to three paragraphs, a summary. Every time. It's the prose equivalent of a stock photo: technically fine, instantly recognisable as generic.
The bigger problem is what's missing. No point of view. No specific numbers tied to real situations. No friction, no contradiction, no "actually this is more complicated than it looks." Real human writing pushes back on itself. AI writing just... continues forward, confidently, forever.
Why Google Cares (And Why "AI Detection" Is Only Part of the Story)
There's a persistent myth that Google penalises content because it was written by AI. That's not quite right. Google's helpful content guidance targets content that exists primarily for search engines rather than people. Raw AI output fails that test not because a machine wrote it, but because it demonstrably doesn't help anyone. It answers surface-level questions with surface-level answers, and trained human readers can feel the emptiness even if they can't articulate it.
The second issue is E-E-A-T. Experience, Expertise, Authoritativeness, Trust. A raw AI article has none of these signals because it has no author perspective, no cited specific experience, no original data. It's citation soup. It reads like a Wikipedia summary written by someone who has never actually done the thing they're describing.
The Specific Failure Patterns I Keep Seeing
After reviewing content for hundreds of client sites, and building internal content processes at Seahawk that we've refined over two years, the failure modes cluster into a few repeating categories.
The confidence without specificity problem. AI writes declaratively. "Email marketing achieves an average ROI of 4,200%." That number comes from a Litmus report and it's real, but in AI output it just floats there, uncontextualised, with no acknowledgment of what type of campaigns, what list sizes, what industries. It sounds authoritative. It teaches nothing.
The structure-first, insight-second problem. AI plans a post before it writes it, and you can always tell. The headings are perfectly logical. The flow is textbook. But the actual insights feel reverse-engineered to fill the structure rather than the structure emerging from genuine things worth saying.
The hedging problem. "It's important to note," "It's worth mentioning," "Keep in mind that." These phrases exist to soften claims that the AI isn't confident about. They also make the prose unbearable to read. I counted eleven instances of "it's worth noting" in a 1,200-word article a client sent me last spring. Eleven.
The vocabulary tell. Certain words appear in AI content at a rate wildly above their natural frequency. "Delve." "Tapestry." "Nuanced." "Multifaceted." "Underscore." If you see three of these in a single post, someone published raw output.
What a Humanization Pipeline Is (And Isn't)
Here's the thing: "humanization" sounds like a single step. Run it through a tool, done. That's not how it works.
A humanization pipeline is a sequence of deliberate interventions that transform generated text into something that reflects genuine expertise, has a consistent voice, and contains information a real person would actually use.
It is not:
- Spinning the text with synonyms
- Running it through Undetectable.ai and calling it done
- Adding a paragraph at the start and one at the end
- Having a junior editor "just read through it quickly"
Those approaches produce content that fails slightly less obviously. The underlying emptiness is still there.
What the Pipeline Actually Contains
The pipeline I run at Seahawk has five stages. Not every post needs all five at full intensity, but every post goes through all five at some level.
- Brief enrichment before generation. The quality of your output is almost entirely determined by the quality of your input. A brief that says "write a 1,500-word post about email marketing best practices" will produce generic garbage. A brief that says "write for a founder of a 10-person SaaS who manages their own Klaviyo account, has a list of 4,000 subscribers, and is seeing open rates drop below 20%, focus on list hygiene and segmentation, not copy tips" produces something usable. We build briefs in Notion with a structured template: audience, problem, specific angle, two or three things the post must NOT say (banning the clichés explicitly), and any internal data or client examples to weave in.
- Structural surgery. The AI's first-draft structure is rarely right for the actual reader. We move sections around. We cut the ones that exist just to pad word count. I find that almost every AI draft has one genuinely good insight buried in paragraph six of section four, and that insight should usually be the lede, not a buried detail.
- Voice injection. This is where the post becomes something a real person wrote. We add specific examples ("a client in the legal sector running Mailchimp with a 60,000-person list"), concrete numbers that come from real situations, and opinions. Actual opinions, not "some experts believe X while others believe Y." If we think X is wrong, we say so and explain why.
- Friction and contradiction. Real expertise acknowledges where things get complicated. If we're writing about WordPress performance, we'll say: "yes, WP Rocket helps, but on a poorly-hosted site it's putting a nice curtain over a broken window." That kind of honesty is what AI strips out by default. We put it back in.
- Read-aloud pass. Someone on the team reads the piece aloud. Not skims it. Reads it aloud. You catch every sentence that sounds like it was generated rather than said. Anything you'd stumble over reading gets rewritten. This step alone catches 60% of the problems that slip through everything else.
The Tooling I Actually Use
Let me be specific, because "use tools" is useless advice.
For generation: GPT-4o for most long-form, Claude 3.5 Sonnet for anything that needs a more conversational register. I've tested Gemini 1.5 Pro and it's capable but its default verbosity is a problem, it wants to write 2,400 words when you ask for 1,500.
For detection and QA: Originality.ai is the most reliable detector I've found for publishing workflows. Winston AI is useful as a second check. Neither is a final arbiter, a well-processed piece should score high on both without any gaming, just because it's been properly edited.
For style consistency: We maintain a Notion "voice bible" for each client. Brand adjectives, prohibited phrases, tone examples, a list of topics the client has strong opinions on. The brief template pulls from this automatically for repeat clients.
For the read-aloud pass: Honestly, Natural Reader at 1.1x speed works well for flagging robotic phrasing. Some editors prefer just reading to themselves. Either works.
Why Most Agencies Skip the Pipeline
Time and margin. A proper humanization pipeline adds two to four hours to a post that an AI can generate in ninety seconds. If you've sold content at £30 per article, you cannot afford the pipeline. You publish raw output and hope nobody notices.
This is the core economic problem with the race-to-the-bottom content market. The price point makes quality impossible, so you get volume without value, and eventually the client's site gets hammered and they blame "AI content" when the real culprit was an impossible brief-to-margin ratio.
At Seahawk, we've had direct conversations with clients about this. We stopped offering content below a certain price per piece because we couldn't stand behind what the economics forced us to produce. Some clients walked. The ones who stayed have sites that are actually ranking.
Back in 2023, we onboarded an e-commerce client who'd been buying bulk AI content at scale from another agency, 200 posts a month, roughly £15 per piece. Their blog had 1,400 posts and ranked for almost nothing. We cut the output to 30 posts a month, doubled the brief quality, ran every post through the full pipeline, and within four months they were ranking page one for eighteen target terms they'd never touched before. Volume is not the answer.
When Raw AI Output Is Actually Fine
Not everything needs the full treatment. Let me be honest about this.
Internal documentation. Product changelog entries. FAQ answers for a support centre. Category page copy that's purely functional. If the goal is to inform rather than persuade, and if there's no SEO or brand voice consideration, raw-ish output with a light edit is completely defensible.
The failure comes when people apply the "good enough for internal docs" standard to public-facing editorial content. Those are different things requiring different standards. Know which one you're producing.
FAQ
Does running AI content through a humanizer tool actually work?
Short answer: no, not reliably. Tools like Undetectable.ai rephrase text in ways that reduce detector scores but don't fix the underlying problem, which is that the content lacks genuine expertise and specificity. You can fool a detector and still publish something that real readers and Google can recognise as thin. The only thing that reliably works is substantive human editing.
How long does a proper humanization pipeline take per post?
For a 1,500-word post, figure two to three hours of human time spread across brief enrichment, structural editing, voice injection, and the read-aloud pass. A 2,500-word pillar post can be four to five hours. That's the honest number. Anyone telling you they're doing it in forty-five minutes is skipping steps.
Should I disclose that content was AI-assisted?
This is more nuanced than most people admit. Google's guidance doesn't require disclosure and doesn't penalise AI use. What matters is quality and helpfulness. That said, in certain sectors, healthcare, legal, finance, transparency about how content is produced is increasingly expected, and probably the right call regardless of what Google says.
What's the single biggest mistake agencies make with AI content?
Confusing generation speed with production speed. The AI can draft in ninety seconds. That's the start of the process, not the end of it. Treating the draft as a near-finished product is where everything breaks down.
---
The pipeline isn't glamorous. It's a series of fairly tedious, careful interventions that together produce something worth publishing. There's no shortcut that skips the human judgment in the middle. I've looked for one for two years. It doesn't exist yet.