A particular chart keeps circulating on social media. A site scales AI-generated articles, organic traffic climbs steeply for a few months, plateaus, then falls off a cliff and keeps falling. The shape resembles a mountain. The accompanying advice is usually that Google has caught on and you should buy the poster's SEO course.
The pattern is real. Most of the specific case studies attached to it are not verifiable. Those two facts sit uncomfortably together, and the gap between them is where clients lose money — first by believing the hype, then by over-correcting on the basis of a story that was never sourced.
We went through the primary sources. Here is what survives checking.
What Google's policy actually says
Start here, because most summaries paraphrase it into something Google did not say.
The relevant policy is scaled content abuse, introduced in March 2024. The current wording on Google Search Central is:
"Scaled content abuse is when many pages are generated for the primary purpose of manipulating search rankings and not helping users. This abusive practice is typically focused on creating large amounts of unoriginal content that provides little to no value to users, no matter how it's created."
That final clause is the part people skip, and it is the most important one in the document. The policy is not about AI. It is about unoriginal, low-value pages produced at volume to manipulate rankings, and it applies identically to a content farm staffed by underpaid humans.
Google's separate guidance on AI content, published February 2023, is equally explicit:
"Appropriate use of AI or automation is not against our guidelines. This means that it is not used to generate content primarily to manipulate search rankings, which is against our spam policies."
And, more bluntly:
"Using AI doesn't give content any special gains. It's just content. If it is useful, helpful, original, and satisfies aspects of E-E-A-T, it might do well in Search. If it doesn't, it might not."
So the honest reading is narrower than either camp wants. AI-generated content is not penalised for being AI-generated. Generating many low-value pages to manipulate rankings is penalised, whoever or whatever wrote them. The intent and the value are what is under assessment; the production method is explicitly not.
What actually happened in March 2024
On 5 March 2024 Google announced three new spam policies — expired domain abuse, scaled content abuse and site reputation abuse — alongside a core update it described as unusually complex, touching multiple core systems and taking up to a month to roll out.
The widely quoted "40%" needs correcting, because it is misused constantly. Google's actual words were that "the combination of this update and our previous efforts will collectively reduce low-quality, unoriginal content in search results by 40%" — a forward-looking expectation for a combined effort, not a measured outcome of that single update. Google later revised the figure upward to 45%. Anyone citing "Google's March update removed 40% of AI content" is quoting a prediction as a result, and getting its scope wrong as well.
The measured effects came from third parties. Originality.ai monitored 49,345 sites and found 837 fully removed from the index, with a broader sample of ~79,000 sites turning up 1,446 manual actions. In their deindexed sample, 100% showed signs of AI-generated content and about half were 90–100% AI-written. Worth noting that Originality.ai sells an AI detector and is therefore an interested party — but unlike most claims in this space, its sample size and methodology are published.
The dataset that actually supports the pattern
The strongest evidence for spike-then-collapse is not a viral chart. It is Lily Ray's May 2026 analysis, which examined 220+ sites that AI content platforms themselves published as success stories — that is, cases selected by the vendors as their best results.
Of those:
| Outcome | Share of sites | |---|---| | Lost 30% or more of peak organic traffic | 54% | | Lost 50% or more | 39% | | Lost 75% or more | 22% |
The typical shape: pages grow over 6–12 months, traffic peaks three to six months after the content peak, and most of the gain is erased within the following year — frequently ending below the original baseline. Most of the declines occurred after the case studies were published. A number of the sites later 410'd or redirected the same pages they had been celebrated for.
A parallel analysis by Malte Landwehr and Tomek Rudzki found that of one AI content tool's flagship clients, 57% declined and 22% showed the classic spike-then-crash; a second tool's featured clients showed 75% losing visibility.
One methodological point deserves emphasis: both analyses deliberately withhold the vendor and site names. As Ray put it, the pattern is the story, not the specific actors. That is a more rigorous choice than it might appear, and it is precisely why these datasets are more trustworthy than the named anecdotes that circulate on social media.
Where the viral version goes wrong
We checked one widely shared example in detail: a claim that the brand Momcozy went from roughly 700 pages to 62,000 pages, spiked, then collapsed.
We could not corroborate it anywhere. No SEO industry writeup, no Ahrefs or Semrush visibility chart, no news coverage, and nothing on the site of the person presenting it. The only substantive public analysis of that brand's SEO describes it in positive terms and says nothing about a page explosion or a collapse.
That does not prove the story false. It means it is an unsourced practitioner anecdote presented with the visual grammar of data, and it should not be repeated as fact — particularly when genuinely rigorous datasets covering hundreds of sites exist and reach a similar conclusion by better means.
The counter-evidence worth carrying
Intellectual honesty requires the other side. Ahrefs studied 331,000 pages and concluded that Google does not punish AI content — it punishes bad content. Their finding is consistent with Google's stated policy and with the Ray dataset: what correlates with collapse is thin, unoriginal, volume-first publishing, not the involvement of a language model.
It is also true that the spike half of the pattern is real and can be substantial. Published case studies show AI-assisted content operations taking sites from hundreds of thousands to millions of monthly organic visits inside a couple of months. The upside is not imaginary. The question is what fraction of it is still there eighteen months later, and on the vendors' own selected examples, roughly half is not.
What we tell clients
Volume is the risk factor, not the tool. If your plan involves a number of pages rather than a set of questions your customers actually ask, the plan is the problem. Publishing 60,000 pages is a strategy that fails on its own terms before Google is involved — nobody can maintain 60,000 pages.
AI in the drafting process is unremarkable. Using a model to draft, restructure, translate or edit content that a knowledgeable person then shapes and stands behind is not what any of this is about, and treating it as forbidden is over-correction.
The test is whether the page would exist without search. If the only reason a page exists is that a keyword tool returned a volume figure, it is in the category the policy targets, regardless of who wrote it.
Assume the measurement lags. The most uncomfortable finding in the Ray data is that traffic peaks three to six months after the content peak. If you are evaluating an AI content programme at month four, you are reading the part of the curve that always looks good. Judge it at eighteen months or do not claim to have judged it.
Reversibility is worth designing in. Sites that recovered generally did so by removing or redirecting large numbers of pages. If you are going to scale content, keep it structurally separable from the rest of the site so that removing it later is a decision rather than a rebuild.
Frequently asked questions
Does Google penalise AI-generated content? Not for being AI-generated. Google's published position is that appropriate use of AI is not against its guidelines, and that content is assessed on quality and originality regardless of how it was produced. What is penalised is generating many low-value pages primarily to manipulate rankings.
What is scaled content abuse? A spam policy introduced in March 2024 covering the mass generation of unoriginal, low-value pages intended to manipulate search rankings — explicitly "no matter how it's created."
Is the spike-then-collapse pattern real? The pattern is documented across 220+ sites in published analysis, with 54% losing at least 30% of peak traffic. Individual viral examples of it are often unsourced.
Can we use AI for content at all? Yes, and most organisations already do. The distinction that matters is whether a person with domain knowledge is shaping and standing behind the output, and whether the page would be worth publishing if search engines did not exist.
How long before a scaled content programme shows its real result? On the available data, traffic peaks three to six months after the content peak and the decline plays out over the following year. Assume eighteen months before you have a real number.
Verified 12 August 2026. Policy text quoted from Google Search spam policies and Google Search's guidance about AI-generated content; March 2024 announcement from Google Search Central and blog.google. Dataset from Lily Ray, "It Works Until It Doesn't"; deindexing figures from Originality.ai via Search Engine Journal; counter-evidence from Ahrefs. Third-party figures come from interested parties in several cases and are flagged as such above; treat them as directional.
