The pitch finds you on LinkedIn, or in your inbox, and it is always the same pitch. An agency you have never heard of has discovered your website and would like to publish thirty, fifty, a hundred articles a month for you. You will not have to write anything. You will not even have to read anything. A pipeline will generate the posts from your domain name and whatever PDF you can spare, on a schedule, forever. There is a word for what comes out of that pipeline, and the internet settled on it with unusual speed: slop.
The offer works because it sounds like arithmetic. Every article is a lottery ticket, so a hundred tickets must beat one. And to be fair to the sellers, the trick can work for a while; a sudden shelf of new pages sometimes does bump a ranking, which is exactly what makes the invoice feel justified. The problem is what happens after the bump, because the machines you are trying to impress have spent the last two years learning to recognize the shelf.
The internet already ran this experiment
Nobody has to speculate about what a web full of automated articles looks like, because we live there. Graphite, a growth agency that studies search data, tracked article publishing across the web and found that in November 2024 the number of new AI-generated articles overtook the number of human-written ones. Roughly half of everything published since is machine output on somebody's schedule.
Then Graphite asked the more interesting question: of all that output, what do the engines actually surface? In a study of 31,493 keywords they classified every article that ranked in Google or was cited by an answer engineAnswer Engine: A tool that responds to a question with a direct answer instead of a list of links to click — like ChatGPT, Perplexity, or the AI summaries now sitting at the top of Google. Read the full entry in the glossary.Answer EngineA tool that responds to a question with a direct answer instead of a list of links to click — like ChatGPT, Perplexity, or the AI summaries now sitting at the top of Google.Read full entry. The results are lopsided. 86 percent of the articles in Google's results were written by humans. 82 percent of the articles cited by ChatGPT and Perplexity were written by humans. And when AI-generated articles ranked at all, they sank: among top results they showed up at half the rate their share of the web would predict.
So the flood happened, and the engines shrugged. Graphite's own reading of the data is the quiet punchline: the growth of AI-article publishing has plateaued, they suspect, because the people running the pipelines noticed it was not working.
Half of the internet is now automatic. The half that gets cited is not.
Google wrote the punishment down
In March 2024 Google stopped treating this as a quality problem and started treating it as spam. A new policy named scaled content abuse targets exactly the product described above: producing "content at scale to boost search ranking", and the policy takes care to add that it applies whether automation, humans or a mix of the two did the producing. Note that wording. It is not an AI policy. Paying a hundred freelancers to churn the same filler is punished identically. What the policy bans is the schedule without the substance.
The enforcement was not subtle. Within weeks, hundreds of websites were removed from the index entirely, not demoted; one early count put the number above eight hundred. When the detection company Originality.ai examined the deindexed sites, every single one showed signs of AI-generated posts, and half of them were almost nothing else, with 90 to 100 percent of their content machine-generated. Google had said it expected the update to cut unoriginal content in its results by 40 percent. After the rollout finished, it reported the real figure at 45.
Put the two sections together and you can see the actual trade on offer. The short-term bump is real, sometimes. The stake is your domain: the years of trust attached to your name are what the agency is playing with, and the agency is not the one who loses them.
What the machines actually reward
There is research on the winning side of the ledger too. In the first large academic study of generative engines, researchers at Princeton and IIT Delhi rewrote real pages nine different ways to see what makes an engine pick a page up and repeat it. The three strategies that worked are almost embarrassing in their plainness: add citationsCitation: When an answer engine names the source of what it just told someone — a business, a site, a particular set of words — instead of leaving the answer unattributed. The AI-era version of being recommended by name. Read the full entry in the glossary.CitationWhen an answer engine names the source of what it just told someone — a business, a site, a particular set of words — instead of leaving the answer unattributed. The AI-era version of being recommended by name.Read full entry to credible sources, quote people who know the field, and include statistics. Pages rewritten that way became up to 40 percent more visible in generated answers. Keyword stuffing, the old reflex, did roughly nothing.
The pattern should sound familiar. It is roughly what Google's own quality raters are told to look for under the name E-E-A-TE-E-A-T: Experience, Expertise, Authoritativeness and Trustworthiness — the shorthand for whether whoever wrote a page has any standing to say what they are saying. Read the full entry in the glossary.E-E-A-TExperience, Expertise, Authoritativeness and Trustworthiness — the shorthand for whether whoever wrote a page has any standing to say what they are saying.Read full entry: experience, expertise, authoritativeness and trust. Two very different measuring sticks, one from an academic benchmark and one from a search giant's rulebook, keep landing on the same spot.
Now look at what those three winning strategies have in common. Each one is a property of an article somebody actually researched. A pipeline that knows your logo and a brand PDF cannot cite the study it never read, cannot quote the customer it never met, and cannot publish the number it never measured. This is why quality wins, and it has nothing to do with taste. To a machine deciding what to repeat to a stranger, quality is evidence density, and evidence cannot be generated from a domain name.
An answer engine is not counting your articles. It is deciding whether you are safe to repeat.
Design the article instead
Here is the alternative, and it costs an afternoon rather than a retainer. Instead of ordering articles by the dozen, design one. Start from the two things no pipeline has access to: what you actually offer, and what your customer actually gains from it. And then, this part matters, feel free to write it with AI. The machine is a fine drafting partner and a disastrous editor-in-chief. The difference between a designed article and slop was never which species typed it. It is whether anyone was responsible for it.
Start from a question a customer has actually asked you. The email you have answered three times this year is an article outline. It is also, nearly word for word, what someone will one day type into an answer engine.
Write the part only you can write. Your process, your prices, the job that went wrong and what it cost, the honest trade-offs in your own product. Answer-shaped contentAnswer-Shaped Content: Writing that leads with the answer to a real question instead of arriving at it — the single change that most often decides whether a machine can quote you. Read the full entry in the glossary.Answer-Shaped ContentWriting that leads with the answer to a real question instead of arriving at it — the single change that most often decides whether a machine can quote you.Read full entry built from real knowledge is the one thing a competitor cannot order from a menu.
Put your own numbers in it. The Princeton study says statistics earn citations, and it means measured ones: how long the job took, what it saved, what it cost. One real number outweighs a page of adjectives.
Edit it like your name is on it, because it is. Draft with a model if you like, then read every sentence and ask whether you could defend it to a customer on the phone. Delete what you cannot. Whatever survives is the article.
Where Limen stands in this
You may have noticed what Limen conspicuously does not sell. There is no button in our dashboard that generates articles from your brand PDF, and there never will be, because we would be handing you the exact liability this essay describes. Limen works on the other side of the equation: the content you already have. We scan your site the way an answer engine reads it and tell you, in plain language, what stands between your pages and AI visibilityAI Visibility: AI visibility is how often, and how accurately, AI assistants name your business when someone asks them a question you could have answered. Read the full entry in the glossary.AI VisibilityAI visibility is how often, and how accurately, AI assistants name your business when someone asks them a question you could have answered.Read full entry: the questions left unanswered, the vague claims sitting where a number should be, the structure a machine cannot parse. Each finding arrives with the fix written out underneath.
The order of operations matters more than people think. A designed article published on a site machines cannot read is a good speech into a dead microphone. Fix the way machines read your siteMachine Readability: Whether the meaning of a page survives being read by software — the plainest test of whether an AI can understand what is actually on it. Read the full entry in the glossary.Machine ReadabilityWhether the meaning of a page survives being read by software — the plainest test of whether an AI can understand what is actually on it.Read full entry first, then publish one article a month you would put your name on. That combination, a readable site and a small shelf of honest pages, is precisely what the research above says the engines are starving for.
Your first scan is free, so paste your address in and see what the machines currently make of your pages. And the next time someone offers you fifty articles a week, remember what half the internet has already learned at its own expense. The machines have all the slop they will ever need. What they are short of is you.
Sources
AI Content in Search and LLMs — Graphite
More Articles Are Now Created by AI Than Humans — Graphite
GEO: Generative Engine Optimization — Aggarwal et al., KDD 2024
New ways we're tackling spammy, low-quality content on Search — Google
Google's March 2024 Core Update Impact: Hundreds of Websites Deindexed — Search Engine Journal