Search any mid-tier marketing question this week and read the first page properly. Not skim it. Read it.
You’ll find the same article about nine times. Same three subheadings in the same order, same worked example, same tidy closing paragraph telling you consistency is key. Different logos at the top.
Nobody got penalized for writing those. That’s the part people get wrong. Every one of them is sitting there, indexed, technically fine, and completely invisible, because being the ninth identical answer is its own punishment and no algorithm had to lift a finger.
So here’s the honest version. Google’s spam policy has never once mentioned how the words got made. It talks about content produced primarily to manipulate rankings rather than to help someone. Whether a person typed it, a model drafted it, or a template stamped it out is not the test and never was. The risk you actually carry isn’t that a crawler sniffs out the model. It’s that you published something indistinguishable from four hundred other pages and then wondered why nothing happened.
The policy is about intent and outcome, not authorship
Scaled content abuse is the rule everyone half-remembers. It covers producing large volumes of low-value pages mainly to game search, and it is explicitly method-agnostic. Human sweatshop, spun template, frontier model, doesn’t matter. Volume plus low value plus manipulative intent is the trigger. Any one of those alone is not.
That last part matters more than the scary headlines suggest. A large publisher can ship thousands of genuinely useful pages and be completely fine. A small site can publish fifty thin ones and be in violation. The ratio that gets you is value per page, not pages per month.
The March core update this year made scaled content abuse its headline enforcement target, and the sites that got flattened were the ones running bulk unedited output. The pattern in the case studies going around is consistent enough to be worth repeating: teams publishing fifty to a hundred AI-drafted pieces with real editorial work on top reported traffic up somewhere in the 30 to 80 percent range, while sites dumping a thousand or more unedited pages reported drops of 40 to 90 percent.
Same tool. Opposite outcome. The variable isn’t the model.
Why everyone using the same tool produces the same article
Here’s the mechanism nobody says out loud.
There are about four models doing most of the world’s content drafting. Most people prompt them roughly the same way, because the same prompt guides circulate on the same LinkedIn feeds. And a huge share of workflows feed the model the current top ten results as research before asking for a draft.
Think about what that last step does. You are explicitly instructing the machine to produce the average of what already ranks.
It does that well. It’s very good at it. The output lands neatly in the middle of the existing consensus, which reads as competent, passes every quality checklist your team uses, and adds precisely nothing to the page it’s competing with.
You didn’t get penalized for using AI. You asked a model to write the median article on the topic and it did exactly what you asked.
The tell is what happens next. Somebody reads the draft, thinks it seems fine, cannot articulate what’s wrong with it, and publishes. It seems fine because it is fine. Fine is the problem.
Median works for search and fails completely in AI answers
For a while you could survive this. Ranking is graded on a curve, and the seventh-best version of an article still picks up scraps of traffic.
That’s changing, and not slowly.
When someone asks ChatGPT or Perplexity the same question, there is no seventh place. Retrieval selects a passage to quote and attribute, usually from a small handful of sources. Being the ninth interchangeable page doesn’t get you a smaller slice of the citation. It gets you nothing, because there is no reason on earth to pick you over the eight pages saying the identical thing with a longer track record.
I’ve written before about what actually earns those citations, and the short version is that specificity does the heavy lifting. A passage that states a real number, a named constraint, or a genuine tradeoff is quotable. A passage saying quality content builds trust over time cannot be quoted, because it doesn’t say anything a model would need to attribute to you.
Consensus writing was mediocre strategy in a ranked world. In a retrieved one it’s a dead end.
The one input a model cannot generate for you
E-E-A-T gets recited constantly and the first E gets skipped constantly. Experience. Not expertise, which you can read your way into. Experience, meaning you actually did the thing and something happened.
A model can organize your experience. It can sharpen it, structure it, cut it down. It cannot have any. Feed it nothing first-hand and it will fill the gap with the average of everyone else’s, which is precisely how you end up as article number nine.
So the useful question isn’t whether to use the tool. It’s what you’re putting into it that nobody else could.
Four things that qualify, none of which any model can invent for you:
- Numbers from accounts you actually run. Not industry benchmarks lifted off a stats roundup. What your cost per lead did when you changed one setting, and over what period.
- The thing you tried that failed. Every competitor publishes what worked. Almost none publish the approach that ate three weeks and went nowhere, which is usually the more useful post and always the more memorable one.
- A real objection from a real conversation. The sentence a customer actually said when they hesitated, in their words rather than your category’s words.
- The constraint you work under. Budget ceilings, a platform limitation, a compliance rule that kills the obvious answer. Constraints are specific, and specificity is unforgeable.
Give a model those and it becomes genuinely useful, because now it’s arranging something only you have. Give it nothing and it hands back the internet’s average opinion with your logo on top.
What I’d change on Monday
Take the last five posts you published. For each one, find the sentence that could not appear on a competitor’s site.
If you can’t find one, that’s your answer, and it has nothing to do with how the post was drafted.
The teams that come out of this fine won’t be the ones who swore off the tools, and they won’t be the ones who scaled hardest either. They’ll be the ones who worked out that the scarce input was never the writing. It was having something to say, and that got scarcer the moment writing got free.
If you want to think that through out loud for your own content, drop me a line on email, WhatsApp or LinkedIn and we can have a quick chat. I’m contracted full time so this isn’t a pitch. My notes on content marketing cover how I’d audit a library for this if you’d rather just read.