There is a shape that shows up again and again in the traffic charts of companies that decided to publish AI content at scale. Pages climb for six to twelve months. Organic traffic peaks a few months behind them. Then the line falls, and it usually keeps falling past the point where the program started.

That shape is not theoretical. Lily Ray, who has spent a decade working on Google algorithm recoveries, has been tracking more than 220 domains that were publicly named as customers in the case studies of a dozen-plus AI content platforms. The point of the exercise was simple: find out what happens after the win is announced. The answer, across industries and across vendors, is that it works until it doesn’t.

What the numbers look like after the case study is published

The headline figures from that dataset are blunt:

• 54% of the sites lost 30% or more of their peak organic traffic.
• 39% lost half or more.
• 22% lost 75% or more.

The measurements come from third-party tools, not first-party analytics: Ahrefs organic traffic estimates and page-count time series, cross-checked against Sistrix visibility data, current as of May 2026. Those tools are well validated but they are estimates, and traffic can move for reasons that have nothing to do with content strategy: redesigns, migrations, acquisitions, seasonality, competitors. Nobody is claiming a specific tool caused a specific decline. What the data shows is a correlation, repeated across hundreds of sites that share the same content patterns and the same trajectory.

One detail deserves more attention than it usually gets. Most of the drops happened after the case studies went live. And a substantial number of those brands have since quietly shrunk their content footprints in 2025 and 2026, removing, redirecting or 410’ing the very pages the success stories were built on. The case studies are often still published. The URLs they celebrate are not.

The boom-bust curve is not new, only faster

Glenn Gabe has a name for the chart: “Mount AI.” Steep ascent, sharp summit, matching descent once Google’s systems have gathered enough signals to work out what a site has been doing. Anyone who has audited through the last few years of core updates recognises the silhouette immediately.

What is different this time is the slope. Scaling to thousands of near-identical pages used to take budget, freelancers and months of coordination. Now it takes a prompt and a weekend. The bust arrives on the same schedule as always; the boom just gets there sooner.

The industry already ran this experiment

September 2023 brought the Helpful Content Update, Google’s most aggressive move in years against pages that, in its own words, feel “created for search engines instead of people.” Six months later came the March 2024 core update, the longest in Google’s history, explicitly designed to “reduce unhelpful, unoriginal content in search results by 45%.” Alongside it, Google formalised a spam policy named Scaled Content Abuse: generating many pages to manipulate rankings, regardless of who or what wrote them.

Note the wording. Google never framed the target as AI. It framed the target as scale without purpose. Human-written content produced inside a rigid SEO template got caught too, which is why so many small publishers who never touched a generator still lost most of their traffic and, in many cases, have not recovered. We wrote about that dynamic in the content framework that worked in 2019 and now works against you.

The material coming out of many current AI content platforms looks and reads a great deal like the material those two updates removed from the index.

Four template families that keep appearing in the declines

Looking at the top-traffic URLs across the declining domains, eight recurring page templates emerge. Most affected sites use three or four of them; the worst-hit use all eight, usually across hundreds or thousands of URLs. They fall into four families, and the common thread is that every one of them can be reproduced by a competitor tomorrow from the same prompt.

Head-to-head and alternatives pages

Programmatic [product-A]-vs-[product-B] pages covering every plausible matchup in a category, plus [competitor-brand]-alternatives landing pages built for every named rival. In one case in the dataset, the majority of a site’s top traffic pages were dedicated to other companies’ brand names. Some sites had drifted so far that they were publishing concept-vs-concept comparisons unrelated to what they sell.

Definition and single-question pages

The what-is-[term] glossary and the FAQ farm, where each URL answers exactly one question with the answer in the opening paragraph, bullets in the body and schema at the bottom. Both are engineered for extraction by AI engines rather than for a reader. Both create enormous low-value page counts when run at scale, and multilingual versions generated without human review compound the quality problem across the whole domain. It is worth noting that Google has since deprecated FAQ rich results, which is not an unrelated development.

Listicles, especially the self-serving kind

“Best [X] for [Y]” is the oldest AI content template there is, inherited from the affiliate era. The riskier variant is the self-promotional listicle: the publisher ranks its own category, and names itself number one, with no evidence it ever tested the alternatives, which is exactly what Google asks for on review content. Sites running dozens to thousands of these saw traffic collapse on effectively the same day in late January 2026. We covered that event as it happened in our look at Google’s crackdown on self-promotional listicles.

Programmatic geography and off-topic filler

City, state and country pages spun from one template, frequently for locations where the business has no presence, is a trick that has been drawing manual and algorithmic penalties for well over a decade. Off-topic volume plays are the other reliable way to get flagged: jokes, name lists, horoscopes, memes and biographical filler published on B2B and services sites that have no business ranking for any of it.

The unconfirmed late-January 2026 adjustment

A secondary pattern sits inside the data. At least 40 tracked sites began declining around 20 January 2026, with losses between 40% and 95% over the January to April window. Google confirmed no update by name. The affected sites overwhelmingly shared one trait: heavy, explicitly GEO-optimised, self-promotional content, usually concentrated in a blog or resources subfolder.

The damage was rarely contained to the offending pages. In most cases the entire subfolder went down with them, and in some cases the whole domain did. That is the signature of a site-level quality assessment, not a page-level one, and it is the most expensive kind to unwind.

How to keep AI in the workflow without buying the crash

None of this makes the tools unusable. The failure mode is almost never the model; it is the operating procedure around it, and specifically the “set it and forget it” configuration where the goal is maximum pages per month with no expert reviewing what ships.

The parts of the pipeline where these tools genuinely earn their keep are research, clustering, outlining, brief creation, synthesising internal data and speeding up the unglamorous middle of the process. The parts where they cost you are judgement, positioning and final publish decisions. We drew that line in more detail in what not to automate with AI, and the trust dimension is covered in our five-pillar framework for AI content audiences actually trust.

Whatever a model contributes, the published page still has to demonstrate real experience and expertise, add information that is not already available on the first ten results, and be honest about how it was made. Google recommends disclosure, and disclosure costs nothing.

Five questions to ask before the next page goes live

1. Does this page exist because a customer needs it, or because a search engine or a language model might quote it?
2. Could a competitor produce a near-identical version tomorrow from the same prompt?
3. Would you be comfortable if a journalist, a client or Google reviewed the complete URL list in this subfolder?
4. If the page is inherently biased in your favour, does it say so plainly?
5. Is there any first-party data, original testing or genuine perspective here that is not already published elsewhere?

If the honest answer to the first two is “the search engine” and “yes,” the page is not an asset. It is a liability with a publish date.

Frequently asked questions

Does Google penalise AI-written content?

Not for being AI-written. Google’s stated targets are unoriginal content produced at scale to manipulate rankings, and its Scaled Content Abuse policy applies the same standard regardless of authorship. A single AI-assisted page with real substance is fine. Ten thousand templated ones are the problem.

Why would AI search visibility fall along with organic traffic?

Because most assistants ground their answers in search results through retrieval. When a domain’s quality signals drop in organic search, the pool of pages that models retrieve from shrinks with it. What is risky for SEO tends to be risky for AI visibility too.

We already published thousands of these pages. What now?

Audit by template, not by page. Identify which URL patterns make up your low-value bulk, consolidate what has genuine demand behind it into fewer, deeper pages, and remove or redirect the rest. Many of the brands in this dataset are doing exactly that, which is also why some of their traffic lines have started to recover.

The bottom line

The playbooks being sold as “AI-first SEO” and “GEO content at scale” are, structurally, the playbooks that were flattened by the Helpful Content Update and the March 2024 core update. The packaging changed; the pattern did not. Across the tracked dataset, the brands still growing are the ones whose content does not match those eight templates, and the brands deleting pages are the ones that scaled straight into them. Differentiation is the whole defence, which is why non-commodity content is worth more per page than any volume target you could hit with AI content at scale.