How hands-on SEOs use AI to make content that holds up
The short answer
None of the 19 practitioners we read is against AI writing. All of them treat one-click articles as a dead end. The difference is a system: method, brief, inputs, human check.
- Write your method down once as a reusable instruction file, instead of typing a new prompt every time.
- Make 20 pages as if each cost $200, read them yourself, and only then scale. That's Karl Kangur's rule.
- The cheapest win is updating old articles. Charles Floate took one from 17 to 2,000+ monthly visits in under 30 minutes.
How this was researched
We focused on 19 practitioners (Kyle Roof, Michał Suski, Koray Tuğberk Gübür, Matt Diggity, Kalin Karakehayov, Craig Campbell, James Dooley, Kasra Dash, Ivana Flynn, Alan Cladx, Szymon Słowik, Robert Niechciał, Stewart Vickers, Chris Walker, Navneet Kaushal, Vaibhav Kakkar, Mads Singers, Karl Kangur, Charles Floate). From their X and LinkedIn posts about AI writing that describe a process, we took the top 10 by engagement per person, about 110 posts, and read them in full, plus two Diggity case studies, a Campbell piece on outsourced content and Kyle Roof's notes on his AI writer.
None of the 19 practitioners we read is against AI writing. They argue about how to do it.
The shared view, in one line: AI does the work, people set the direction and check the result.
The two who post most about it with an actual process are Matt Diggity (481 posts in our library) and Charles Floate (471).
Six things they agree on
1. Write the method down once
Don’t type a fresh prompt every time. Put your method in a file the AI reuses. People call these skills or agents.
- Floate (March 2026) built five, connected to Ahrefs:
- What the top results have in common.
- What competitor articles have in common.
- How to improve a page.
- Where competitors get their links.
- A self-check before anything is published.
- Kasra Dash (September 2026): the process lives in the framework; the AI model is the part you can swap.
2. Start every article from a brief
A brief is a one-page plan for an article: what it must cover and how it should be shaped.
Floate’s way to build one (March 2026):
- Paste the full text of the top 5 Google results into Claude.
- Ask for every heading, every question they answer, and the topics at least 3 of them share. Those are your “must cover”.
- Ask which topics only 1 or 2 of them cover. Those are your chance to say something new.
- Paste your own page and ask what’s missing.
Craig Campbell adds: don’t auto-publish AI text to your site. Hire an editor whose job is AI drafts.
3. Feed it your best work
Karl Kangur (March 2026): not “write me an article about X”. Instead, give it:
- your 10 best articles,
- the competitor pages you want to beat,
- a full transcript of a good copywriting course.
Then give feedback the way you would to a $200-an-hour writer.
Diggity and Kangur say the same thing from two sides. Write what AI can’t copy: your own results, your own data, your own tools. If AI can write it, AI already knows it.
4. Make 20 great pages before you make 400
Kangur had a client with 400+ pages to build. His answer: build the best 20 first. Read each one yourself and ask whether it is actually good.
- If yes, you have a system worth scaling.
- If no, you’ve saved yourself 380 bad pages.
Koray (November 2026) adds that there’s no safe daily number. 5,000 human-written pages a day can get a site in trouble; 10 AI pages a day can too. Publish a small batch, and keep going only while Google keeps indexing the new pages.
5. Update old articles first
Floate’s case (April 2026): an old review untouched for a year. Under 30 minutes of work. Monthly visits went from 17 to over 2,000.
What he had the AI change:
- prices, details and screenshots,
- dates, including the hidden date in the page code,
- a short FAQ based on what top results answer,
- topics competitors cover that he didn’t,
- a first-person opening about his own experience.
6. A person checks every page
- James Dooley (September 2026): AI collects and sorts. People check the facts, cut what’s irrelevant and add real experience.
- Stewart Vickers: three rules for his AI. No writing without research. No made-up numbers. A person reviews before anything goes out.
- Kyle Roof (2023): AI gets you about 80% there. He’s never seen output you can publish without editing.
Where they disagree
Should you try to make AI text “undetectable”?
- Floate (February 2026): AI writing has a predictable rhythm, and Google notices. He mixes formats (paragraphs, lists, tables) and switches between casual and technical, so it reads like a slightly messy person who knows the subject.
- Koray: text from one AI model carries a statistical fingerprint. If you and a competitor use the same tool, your pages sound alike.
My read: Floate gives no before-and-after data. But both point to the same risk: a whole site written in one AI voice.
Which model?
- Stewart Vickers (2026): DeepSeek for final drafts, about a quarter of Claude’s cost. One pipeline’s model bill fell from $2,200 to $340 a month.
- Karl Kangur (2026): from the same sales-call transcript, Claude’s ideas were clearly better than ChatGPT’s.
What went wrong
- Kasra Dash (May 2026) joked that he fired his whole SEO team and handed everything to Claude. Result: 14,700 fewer clicks and 923,000 fewer impressions, a slow slide.
- Karl Kangur has seen owners publish hundreds of cheap AI articles to get into AI answers. Rankings collapsed within months.
- James Dooley: the last dozen people who asked him for help had all scaled with AI faster than they could manage.
What worked (all self-reported)
| Who | What they did | Result |
|---|---|---|
| Diggity’s agency, The Search Initiative | AI wrote outlines, then a second step expanded them | Monthly revenue $166K to $491K |
| Charles Floate | Updated one old article with AI | 17 to 2,000+ monthly visits |
| Casey Botticello | 29 articles by hand, then about 700 with AI, each edited 1 to 3 hours | About $21.7K a month |
From Diggity’s write-up:
“Our client’s monthly revenue from $166k to $491k.”
One detail matters. His prompts were fixed per page type. A recipe outline had to include prep time, quantities, substitutions, steps, storage and nutrition. A fixed outline per page type beats “write an article”.
What Google says
- Google updated its AI content guidance on 1 October 2026, as reported by Lily Ray: a person should check all AI content before it’s published, including titles, descriptions and image text.
- Graphite found 86% of articles ranking on Google are written by people. They didn’t test “AI plus heavy human editing”.
Limits of this evidence
- Almost all of this is posts and self-reported cases. No controlled tests.
- Several of these people sell courses, communities or tickets.
- We read the top 10 posts per person by engagement, not everything they’ve said.
✓ Do this next
Pick your oldest article that still gets a few visits. Give an AI the page plus the top 10 results for its search term, ask what's missing, and update it this week.
My takeAI doesn't make content good or bad. It makes the gap between people with a process and people without one much wider, and much faster.
Sources
- Matt Diggity: $491K revenue a month by building an AI SEO genius (case study)diggitymarketing.com
- Lily Ray: It works until it doesn't (the risks of AI content)lilyraynyc.substack.com