Loading video...
Video Failed to Load
Everyone argues about whether AI content works. We tested it at scale instead. We published 744 articles across 68 real websites. Half went live exactly as the AI wrote them. The other half got one 20 minute step first, and it produced 444% more traffic. In this video I... show more
32,453 views • 18 days ago •via X (Twitter)
23 Comments
Tough to pay attention to the content when the video script is very clearly written by AI. That’s the whole story.
This applies to web copy too. AI builds the structure. Human judgment makes it convert. The agencies that figure out this balance will win. The ones that ship raw AI output will lose clients fast.
Actually the obvious better step is to always use a custom humanizer skill alongside a web-focused conversion copy skill, with a model that isn't doing the watermark stuff (ie; open-source).
444% more traffic is a traffic result. Score the same 744 articles on enquiries per article and the winning half can change.
That's what we do at at scale, but with a human reviewing each of the content pieces.
744 articles across 68 sites is the rare version of this argument with an actual control group, which is why it is worth paying attention to. The number I would want alongside the traffic is what those two groups did after the click, because the gap in time on page is usually where the 444% either turns into revenue or does not.
'Readers can't tell but Google can' is the part I'd underline. Same thing in ad creative. Two versions I can't separate and the auction prices them completely differently.
That 20-minute step is doing a lot of heavy lifting. What was the step, and did you see the same lift across all site types or only certain niches?
Skill issue.
AI companies - Here, buy our AI writing tools Also AI companies - We are going to penalise you for using the same AI writing tools we sold to you MAKE THAT MAKE SENSE!
744 articles and 68 sites is a real test. Were the sites equivalent in authority and initial traffic going in? A 444% lift reads very differently depending on whether those pools were matched.
How does this work with translations? What is the data showing when an article is translated by AI? The source is written by a human, but the translated version is written by AI, because the writer does not know every language. My Spanish and Hungarian is terrible... 😆
WOW AMAZING BRO. YOU ARE REALLY GOOD AT THE SEO THING!!!
The 20 minute step is the whole game. Mine happens before the writing rather than after: the model has to fetch the ten pages currently ranking for that term, read them in full, and cover what all ten miss. Same model, and what comes out isn't the same thing at all. One site of mine runs 200 a day that way, 6x sessions since March.
The difference between 0 traffic and compounding search rankings is injecting real operator screenshots and primary data into the raw draft. Google does not penalize model generation, it penalizes the lack of information gain. A 20 minute editing pass that adds firsthand proof saves the entire domain from index rot.
the scary bit is that Google catches what readers can't — it's grading something underneath the words. genuinely curious what that 20-minute step actually is, because "add a human" is doing a lot of quiet lifting in that sentence
Loved this video, watched it on linkedin before here
Editing doesn't make a meaningful difference but makes content perform better.
The human element is still non-negotiable for SEO success
Who is everyone? These people are morons
The last part, where you talk about how to use ai and still add the human touch is so helpful. It's certainly worth the extra time it takes.
The five steps Neil mentions and some thoughts around it
444% more traffic, that step matters
