Contents
- 1. Google Judges What's in It, Not Whether AI Wrote It
- 2. ChatGPT / Claude / Gemini — Splitting Work Across Three Models
- 3. Prompts That Actually Work — Persona + Sample + Constraints
- 4. The Hybrid Workflow — Four Steps Humans Add
- 5. Five "Tells" That Reveal AI — And How to Kill Them
- 6. A Six-Step Hands-On Workflow
- 7. Three Pitfalls You Must Avoid
- Summary
- FAQ
Google rolled out core updates in March and May 2026. It hasn't explained in detail what each one targeted, but its spam policies prohibit mass-producing low-value pages to manipulate rankings, however they're created. Meanwhile, the "AI drafts → expert edits → personal experience and first-party data added" hybrid approach lines up directly with the original, helpful content Google says it wants. AI writing wasn't banned — we've entered an era where what matters is what the human added.
The other major thing is that model selection now determines the outcome. A long-form piece written in Claude, a research-backed draft from ChatGPT, and a Workspace-integrated article in Gemini come out as visibly different pieces of writing. "Just use ChatGPT for everything" is a 2024 mindset.
My stance up front. The people who want AI to make life easier are the ones who quit AI fastest. The people who keep winning at both SEO and LLMO are the ones who do "AI for the draft, me for the edit, my own experience layered on top" — the unglamorous version. This article covers the three-model split, prompt structures that work, the hybrid writing workflow, how to kill the AI tells, and the common pitfalls — all as of May 2026.
In 2026, "Hybrid" Is the Right Answer for AI Writing
— The sweet spot is between "let AI do it all" and "write everything by hand"
2024: "Write it with AI" was the trend →
2026: "AI drafts it, human finishes it" is the right answer for SEO and LLMO.
Neither pure AI nor pure handcraft wins — the hybrid in between does.
1. Google Judges What's in It, Not Whether AI Wrote It
In March 2026 Google rolled out a spam update and a core update, followed by another core update in May (Google Search Status Dashboard). Google hasn't explained the goals of these core updates in detail, and it has not said they targeted AI articles. What is spelled out is in its spam policies: generating many low-value pages primarily to manipulate search rankings is prohibited as "scaled content abuse", "no matter how it's created," and using generative AI to churn out pages is listed as an example. At the same time, Google says it rewards high-quality content, however it is produced — using AI is not in itself a reason for demotion. The yardstick it points to is E-E-A-T (Experience / Expertise / Authoritativeness / Trustworthiness).
That distinction matters. Using AI is not the problem in itself. The problem is "no human experience, no expertise, no first-party data, no distinct point of view". The flip side is that articles where AI drafted, AI structured, but a human layered on experience, primary data, and a viewpoint during the edit have no reason to be penalized under Google's stated policies. In its guide to creating helpful content, Google lists self-assessment questions such as whether it's clear who wrote the page (bylines, author background) and whether it offers original information or analysis.
The same structure works for LLMO (being cited by AI search). From the point of view of AI search such as ChatGPT, Perplexity, and Google AI Overview, the generic prose AI produces sits in the same pool as the existing internet content it was trained on — there's little reason to cite it. "What only you can write" is itself the value.
2. ChatGPT / Claude / Gemini — Splitting Work Across Three Models
Three families have reached usable territory for text generation: OpenAI's ChatGPT, Anthropic's Claude, and Google's Gemini. Each vendor ships a new version every few months, so this section compares the strengths and weak spots of each product family rather than specific version names. The question isn't "which is the strongest" — it's "which task goes to which." For the models you can pick right now, see the list of major AI models.
| Model | Where it wins | Weakness | Tasks to give it |
|---|---|---|---|
| Claude (Anthropic) | Long-form voice, tone consistency, naturalness | No image generation | Essays, columns, thought leadership, ghostwriting |
| ChatGPT (OpenAI) | Tool ecosystem, Deep Research, images | Tone slants corporate, harder to steer | Research-backed drafts, image-rich pieces, SEO optimization |
| Gemini (Google) | Workspace integration, current data, reading long documents | Voice and tone reproduction one notch lower | Google Docs writing, large reference summarization, current news |
How to re-check when a new version ships: don't take the table on faith — run the same test side by side. (1) Paste 1,000 words of your own past writing, send the same "write in this style" request to all three, and see how much of your voice survives (tone consistency). (2) Ask for research involving proper nouns and numbers, and check whether you can trace the sources (research). (3) Check whether the model can be called directly from the tools you already use, such as Google Docs or Gmail (integration). If the ranking flips on any of these three, just reassign that row of the table.
How I actually use them: "Claude for the body of a blog post, ChatGPT for research and fact-checking, Gemini for internal docs and anything touching current news." Even this article was drafted in Claude, then I asked ChatGPT separately for "fact-checking, latest benchmark numbers, and SEO heading suggestions." Running all three at $60/month is dramatically better cost-for-quality than trying to do everything in one tool. The $60 = Claude Pro $20 + ChatGPT Plus $20 + Google AI Pro $19.99 (US monthly list prices as of October 4, 2026).
3. Prompts That Actually Work — Persona + Sample + Constraints
"AI sounds like AI" because most prompts end at "write an article about X." Add three elements and the AI smell drops sharply.
Template: You are [persona]. Read the sample below and write [topic] in the same voice. Constraints: [list]. Sample: [paste]. Topic: [topic]
Of the three, ② sample-paste is the most effective. Vaguely writing "avoid an AI-sounding tone" is far weaker than pasting a thousand words of your own past writing and saying "continue in this voice." AI is much better at copying a positive example than following a list of don'ts. For this article, I fed Claude three paragraphs from my earlier "How LLMs work" piece as the sample, then hand-edited.
4. The Hybrid Workflow — Four Steps Humans Add
Here are the four steps where a human adds what AI can't to an AI draft, using a furniture retailer's product category pages as the example.
Why it works: in its guide to creating helpful content, Google lists questions such as whether it's clear who wrote the page (bylines, background) and whether it offers original information or analysis. How much it helps varies by site and field, so check against your own numbers
The hinge of this workflow is the structural choice to make humans handle the three things AI can't do: ① specific field experience, ② first-party data only your company has, ③ trust through real names and credentials. None of these can be replaced no matter how much any AI model advances. The moment you decide you can automate them, you lose at SEO.
5. Five "Tells" That Reveal AI — And How to Kill Them
Five patterns that scream "this was written by AI" — to readers, editors, and Google alike. Mechanically removing them raises quality a tier.
| Tell | Examples | How to kill it |
|---|---|---|
| ① Set phrases | "delve into," "comprehensive guide," "in the ever-evolving landscape," "navigate the complexities of" | Ban in prompt, then Find & Replace |
| ② Perfect structure | Every H2 the same length, always exactly five bullets | Deliberately uneven sections, mix in paragraph-only passages |
| ③ Zero lived experience | "Generally," "in most cases," "it is said that" | Force in "in my case," "after three months of use" |
| ④ Neutral stance | "On the other hand…" both-sides-ism that could mean anything | "I pick A. Here's why" — make the position explicit |
| ⑤ Connector tic | "Furthermore," "additionally," "importantly" sprayed everywhere | Delete, shorten sentences, swap in spoken-language phrasing |
The one that personally moves the needle most is ③ forcing in lived experience. AI can only produce "general statements." Inserting two or three deliberate "in my case…" / "when I actually tried this…" passages immediately gives the piece a sense of specificity. The SEO upside is real too — these are the literal locations that match Google's Experience signal in E-E-A-T.
6. A Six-Step Hands-On Workflow
The hybrid writing flow I actually use for one blog article. Total time per article: 2–3 hours (vs. 6–8 hours pure handcraft, 30 minutes pure AI).
The trick is to put your biggest time chunk — 60 minutes — into STEP 4 (human edit). The 60 minutes of editing decides the output quality far more than the 30 minutes of AI drafting. Treating Step 4 as "AI saved me time, I'll skip ahead" leaves you with the same result as unedited mass-produced articles — "just another AI-sounding article."
7. Three Pitfalls You Must Avoid
Pitfall ①: Letting AI decide the topic
"Give me 10 interesting blog topics" → pick one of the suggestions — this is the single biggest reason you don't win at SEO. AI-proposed topics are aggregations of internet content the AI already trained on; the same ideas surface for everyone else too. The topics worth writing are "experience only you have," "field intuition from your industry," "your personal opinion on current news" — and AI cannot produce these. The right division of labor is: human picks the topic, AI handles only the writing.
Pitfall ②: Leaving hallucinations in
AI will cheerfully write nonexistent statistics, quotes, URLs, and people. If AI hands you something like "a 2024 McKinsey study found productivity rose 42%," always verify against the original URL. Most of the time it's real, but pure fabrications slip in. Leaving that in and publishing it gets you called out by readers and other outlets — trust collapses fast. "See a proper noun + a number? Open the original." Make it a mechanical habit.
Pitfall ③: Failing to kill the "good-student" tone
By training, AI has a tendency to "avoid criticism, present both sides, feign neutrality, and end on a hedge." This is the single most boring writing style for a reader. Deliberately insert strong stance statements during the edit: "I think A. B is wrong. Here's why." The "potentially controversial" opinions are the ones that end up getting shared, cited, and gathering fans. Asking AI to "add a strong opinion" won't work, so a human has to add this in by hand.
Summary
AI writing has shifted from its 2024 framing as a "tool for taking it easy" to a 2026 framing as a "foundation that raises quality." Both Google and AI search aren't looking at "was this written by AI" — they're looking at "is human experience, expertise, and unique data riding on top of it." The irony is that the thing whose value rose the most in the AI era is "what only a human can write." Use AI to speed up the drafting and pour the freed time into field reporting, primary-data collection, and building strong opinions — that's the core of AI writing practice in 2026.
FAQ
Will Google penalize articles written with AI?
Not for "being AI-generated" alone (per Google's official line). But mass-producing low-value pages to manipulate rankings violates Google's spam policies, whether or not AI was used. If AI drafts and a human edits in experience and original information, using AI is not a problem in itself.
Which AI is best for writing?
Depends on the job. For long-form naturalness and tone consistency: Claude. For research and tool integration: ChatGPT. For Workspace and current information: Gemini. Rankings shift with every new version, so the reliable approach is to paste your own past writing, send the same request to all three, and compare how well each keeps your voice and gets the facts right (current model names are in the model list). Better to spend $60/month running all three than to bet on a single "best one."
How do I make AI writing not sound like AI?
Always include three things in the prompt: ① Persona (who's writing), ② Sample (write like this — paste your own past writing), ③ Constraints (don't do this, length limits). Of these, "paste 1,000 words of your past writing as a sample" is the most powerful. 10× more effective than asking "avoid an AI-sounding tone."
How much time should I spend on human editing?
Roughly 2× the AI drafting time. If AI drafts in 30 minutes, edit for 60. If AI drafts in an hour, edit for two. Cutting edit time leaves the AI smell and loses you the SEO race. "AI saves time" only applies to drafting; editing rewards every minute you put in.
Is it okay to ask AI to "suggest interesting topics"?
SEO-wise, no. AI-suggested topics are aggregations of internet content it already trained on — everyone else lands on the same ones. The topics worth writing are experience only you have, field intuition from your industry, and your personal opinion on current news. Only humans can produce these. The right split is: human owns the topic, AI owns the writing.
How deep does fact-checking need to go?
Verify every number, quote, URL, and proper noun against the original. AI casually produces nonexistent statistics like "a 2024 McKinsey study found 42% improvement." Most exist; leave in the fakes that slip through and trust evaporates fast. Make "see a proper noun + a number? open the source" a mechanical habit.
